[Perf, Test] (MG_Util): libfork execution engine for the shader compile pool, runtime-selectable

Adds libfork v3.8.0 (3rdparty submodule, header-only, wired like the asio
precedent) as a second execution engine behind ShaderCompilePool, selected
per process by MOBILEGL_ASYNC_POOL=asio|libfork (default asio; unknown
values warn and fall back). The engine boundary is deliberately tiny: the
queue, the concurrency budget and its clamping, the suspension latch,
cancel request-vs-outcome, the stopped-is-synchronous fallback and the
drain all stay in the shared Impl - an engine only answers how a
budget-cleared job reaches a worker.

The libfork engine runs detached root tasks as CHAINS: a finished body
takes the next queued job in the same coroutine on the same worker, so
the refills a worker posts are absorbed without a scheduler round trip
(the naive dispatch-thread shape measured 4x worse than asio on short
jobs). Absorption is bounded at one job per live chain - unbounded
absorption serialized bursts posted from inside the pool, which is the
shipped shape (one compile settling fans out link jobs via SubmitAfter
and the adoption map), caught by the review and pinned by a permanent
peak-concurrency regression test (pre-fix: libfork peak 1 vs asio peak 4
on a 16-job worker-posted burst). External submissions go through a
round-robin adaptor instead of lf::lazy_pool::schedule, which both
avoids a data race on lazy_pool's unsynchronized xoshiro under
concurrent submits and beats birthday-collision placement by ~1.3x at
budget == thread count.

The measured answer to "does asio scale poorly": no - the executor was
never the bottleneck. On real pack corpora extracted from the trace
fixtures (BSL 61 shaders, Complementary 277), interleaved best-of-5 per
cell, the engines are within noise of each other at every thread count
(complementary: 4965/2506/1376/824 ms at 1/2/4/8 threads for asio;
libfork within 1%), both ~6x at 8 threads. perf counters show the
flattening past 4 threads is machine-level (instructions flat at 22.1e9
from 1 to 16 threads - no added work, no lock spinning - while cycles
and LLC misses double: memory-stall bound), and the separating control
- N fully independent single-threaded processes with no shared
scheduler at all - scales WORSE than the pool (5.27x vs 5.94x at 8).
The pool microbenchmark does favor libfork on pure dispatch (518 vs
530 ns/job at 1 worker, growing with worker count), but a real compile
body is 1-100 ms, so dispatch is under 0.1% either way. asio therefore
stays the default; this branch exists to make the comparison
reproducible (MG_Test/Util/AsyncPoolBench drives either engine over a
corpus directory) and to keep the alternative viable.

613/613 unit tests in all four combos ({asio, libfork} x {async default
on, kill switch}), integration scenarios byte-identical between engines
on both backends.

.github/workflows/test.yml

@@ -177,9 +177,13 @@ jobs:
         uses: lukka/get-cmake@v4.3.3
 
       - name: Install runtime dependencies
+        # libegl-mesa0 is the EGL vendor library itself: DriverBench brings up a
+        # real GL context, and libegl1 is only glvnd's dispatch. It normally
+        # arrives as a Recommends of libegl1, which is too quiet a dependency for
+        # the one job that needs a working driver.
         run: |
           sudo apt-get update
-          sudo apt-get install -y libvulkan1 libegl1 libgles2 libgl1-mesa-dri mesa-vulkan-drivers
+          sudo apt-get install -y libvulkan1 libegl1 libegl-mesa0 libgles2 libgl1-mesa-dri mesa-vulkan-drivers
 
       - name: Download Linux runtime
         uses: actions/download-artifact@v8

.gitmodules

@@ -31,3 +31,9 @@
 [submodule "3rdparty/apitrace"]
 	path = 3rdparty/apitrace
 	url = https://github.com/MobileGL-Dev/apitrace.git
+[submodule "3rdparty/asio"]
+	path = 3rdparty/asio
+	url = https://github.com/chriskohlhoff/asio.git
+[submodule "3rdparty/libfork"]
+	path = 3rdparty/libfork
+	url = https://github.com/ConorWilliams/libfork.git

CMakeLists.txt

@@ -4,6 +4,11 @@ project("MobileGL")
 
 option(MOBILEGL_BUILD_TEST        "Build MobileGL tests"                             ON )
 option(MOBILEGL_BUILD_BENCHMARK   "Build MobileGL benchmarks"                        ON )
+# Headless end-to-end GPU scenarios (MobileGL/MG_IntegrationTest). They need a
+# real GPU/ICD to do anything, so they are off by default for CI; every scenario
+# skips cleanly where there is none. Registered under the `integration-gpu`
+# ctest label so a run can select or exclude them.
+option(MOBILEGL_BUILD_INTEGRATION_TEST "Build MobileGL headless GPU integration tests" OFF)
 option(MOBILEGL_FORCE_RELEASE_OPT "Enable Release optimization flags in Debug build" ON )
 option(MOBILEGL_ENABLE_TRACY      "Enable tracy for profiling"                       OFF)
 option(MOBILEGL_BUILD_TRACE_REPLAY "Build desktop apitrace replay runner"            OFF)
@@ -17,7 +22,9 @@ if (ANDROID)
     set(MOBILEGL_BUILD_BENCHMARK  OFF CACHE BOOL "Build MobileGL benchmarks" FORCE)
 endif()
 
-if (NOT CMAKE_BUILD_TYPE STREQUAL "Debug" OR MOBILEGL_FORCE_RELEASE_OPT)
+option(MOBILEGL_ENABLE_LTO "Build with ThinLTO/IPO" OFF)
+
+if ((NOT CMAKE_BUILD_TYPE STREQUAL "Debug" OR MOBILEGL_FORCE_RELEASE_OPT) AND MOBILEGL_ENABLE_LTO)
     # Check if ThinLTO or LTO is suppported
     include(CheckIPOSupported)
     include(CheckCCompilerFlag)
@@ -147,6 +154,9 @@ set(SOURCE_FILES
 
     MobileGL/MG_Util/Debug/Log.cpp
 
+    MobileGL/MG_Util/Async/JobNode.cpp
+    MobileGL/MG_Util/Async/ShaderCompilePool.cpp
+
     MobileGL/MG_Util/Math/VectorTypes.cpp
     MobileGL/MG_Util/Metrics/TextureMetrics.cpp
 
@@ -180,6 +190,7 @@ set(SOURCE_FILES
 
     MobileGL/MG_Util/Classifiers/TextureEnumClassifier.cpp
 
+    MobileGL/MG_Util/ShaderTranspiler/CompileEnv.cpp
     MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpvcSession.cpp
     MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.cpp
@@ -187,6 +198,7 @@ set(SOURCE_FILES
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/FlattenInterfaceStructPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EliminateFloatEqualsZeroPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RenameSamplerFunctionParameterPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RenameBuiltinShadowingFunctionsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecomposeWorkgroupVec3Pass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecoratePositionInvariantPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/LowerDrawParametersPass.cpp
@@ -220,6 +232,7 @@ set(SOURCE_FILES
     MobileGL/MG_Impl/GLImpl/Framebuffer/Validators.cpp
     MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp
     MobileGL/MG_Impl/GLImpl/Program/GL_Program.cpp
+    MobileGL/MG_Impl/GLImpl/Program/ProgramInterface.cpp
     MobileGL/MG_Impl/GLImpl/Program/GL_ProgramPipeline.cpp
     MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.cpp
     MobileGL/MG_Impl/GLImpl/Texture/Validators.cpp
@@ -243,6 +256,7 @@ set(SOURCE_FILES
     MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp
     MobileGL/MG_Backend/DirectGLES/Utils.cpp
     MobileGL/MG_Backend/DirectGLES/Managers.cpp
+    MobileGL/MG_Backend/DirectGLES/MultiDraw.cpp
 
     MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp
     MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp
@@ -281,7 +295,11 @@ set(SOURCE_FILES
     MobileGL/MG_State/GLState/TextureState/TextureUnit.cpp
     MobileGL/MG_State/GLState/TextureState/TextureState.cpp
     MobileGL/MG_State/GLState/ProgramState/ProgramObject.cpp
+    MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.cpp
+    MobileGL/MG_State/GLState/ProgramState/ShaderCompileTask.cpp
     MobileGL/MG_State/GLState/ProgramState/ShaderObject.cpp
+    MobileGL/MG_State/GLState/ProgramState/ShaderPreprocessCache.cpp
+    MobileGL/MG_State/GLState/ProgramState/ShaderCompileAdoptionMap.cpp
     MobileGL/MG_State/GLState/ProgramState/ProgramState.cpp
     MobileGL/MG_State/GLState/RenderState/RenderState.cpp
     MobileGL/MG_State/GLState/FramebufferState/FramebufferObject.cpp
@@ -315,6 +333,11 @@ if (WIN32)
     )
 endif()
 
+# The shader-compile pool runs standalone Asio on real threads. This host's glibc (>= 2.34)
+# merged pthread into libc, so it links without asking, but the NDK and musl are not
+# guaranteed to be as forgiving - ask for it explicitly rather than rely on the accident.
+find_package(Threads REQUIRED)
+
 set(MOBILEGL_LINK_LIBRARIES
         glslang::glslang
         spirv-cross-c
@@ -324,12 +347,17 @@ set(MOBILEGL_LINK_LIBRARIES
         GPUOpen::VulkanMemoryAllocator
         Vulkan::UtilityHeaders
         spirv-reflect-static
+        Threads::Threads
 )
 
 set(MOBILEGL_COMPILE_DEF
         -DVMA_STATIC_VULKAN_FUNCTIONS=0
         -DVMA_DYNAMIC_VULKAN_FUNCTIONS=1
         -DVMA_VULKAN_VERSION=1001000
+        # Header-only Asio, no Boost, no deprecated interfaces. Set on the definition list
+        # rather than per-target so the shared library and the _s static target agree.
+        -DASIO_STANDALONE
+        -DASIO_NO_DEPRECATED
 )
 
 message(STATUS "MOBILEGL_COMPILE_DEF=${MOBILEGL_COMPILE_DEF}")
@@ -341,6 +369,17 @@ set(MOBILEGL_INCLUDE_DIR
         ${spirv-tools_SOURCE_DIR}/include
         ${spirv-tools_BINARY_DIR}
         ${SPIRV-Headers_SOURCE_DIR}/include
+        # Header-only submodule: no add_subdirectory, no link target. Only
+        # MG_Util/Async/ShaderCompilePool.cpp includes it, and it stays behind that file's
+        # pimpl so no consumer target needs this path.
+        ${CMAKE_SOURCE_DIR}/3rdparty/asio/asio/include
+        # The second shader-compile execution engine (MOBILEGL_ASYNC_POOL=libfork), on the
+        # same terms as Asio above: header-only, no add_subdirectory (its CMakeLists only
+        # declares an INTERFACE target plus install/test scaffolding we do not want), no link
+        # target, and reachable from exactly one translation unit. libfork's own
+        # target_compile_features asks for cxx_std_23, which this project already sets
+        # globally, so its C++20 coroutines need no per-source standard override.
+        ${CMAKE_SOURCE_DIR}/3rdparty/libfork/include
 )
 
 add_library(${CMAKE_PROJECT_NAME} SHARED
@@ -538,6 +577,12 @@ if (NOT ANDROID)
         add_subdirectory(MobileGL/MG_Test)
     endif()
 
+    # After MG_Test so googletest is already available when the unit tests are
+    # built; the module fetches its own copy when they are not.
+    if (MOBILEGL_BUILD_INTEGRATION_TEST)
+        add_subdirectory(MobileGL/MG_IntegrationTest)
+    endif()
+
     if (MOBILEGL_BUILD_BENCHMARK)
         add_subdirectory(MobileGL/MG_Benchmark)
     endif()

MobileGL/Config.h

@@ -29,6 +29,33 @@ namespace MobileGL::MG_Config {
         ForceOff,
     };
 
+    // Preferred DirectVulkan dispatch tier for the glMultiDraw* families. A preference,
+    // never a demand: the renderer clamps it to what the device supports at device
+    // creation, falling down the chain ext -> indirect -> unroll with one log line.
+    enum class MultiDrawMode : Uint8 {
+        Auto = 0, // unset: best supported tier
+        Ext,      // VK_EXT_multi_draw: one vkCmdDrawMultiEXT / vkCmdDrawMultiIndexedEXT
+        Indirect, // multiDrawIndirect feature: one vkCmdDraw*Indirect over a transient command array
+        Unroll,   // one vkCmdDraw* per sub-draw
+    };
+
+    // Preferred DirectGLES emulation tier for glMultiDrawElements(BaseVertex). GLES has no
+    // such entry point in core, so every tier below is an emulation; they differ only in
+    // which driver capability they lean on and how many driver calls a batch costs. Like
+    // the Magma knob this is a preference, clamped at resolution time to what the ES
+    // driver actually supports, with one log line when it falls back.
+    enum class GLESMultiDrawMode : Uint8 {
+        Auto = 0,      // unset: best supported tier
+        Ext,           // one glMultiDrawElementsBaseVertexEXT
+        MultiIndirect, // one glMultiDrawElementsIndirectEXT over a scratch command buffer
+        Indirect,      // one glDrawElementsIndirect per sub-draw over that same buffer
+        BaseVertex,    // one glDrawElementsBaseVertex per sub-draw
+        DrawElements,  // baseVertex folded into a scratch index buffer on the CPU, then plain
+                       // glDrawElements per sub-draw (for drivers with no base-vertex draw at all)
+        Compute,       // a compute shader flattens every sub-draw into one rebased index buffer,
+                       // drawn by a single glDrawElements
+    };
+
     // Feature toggles parsed once from environment variables in MG_ConfigLoader::Init()
     // (ConfigLoader.cpp), before the accepted-env map is destroyed. All Bool fields share
     // one truthy rule: the variable is set, non-empty, not "0", and not "false"
@@ -39,6 +66,12 @@ namespace MobileGL::MG_Config {
     //   - DISPLAY: X11 session variable, not MobileGL configuration.
     //   - MOBILEGL_LOG_FILE_PATH: log-file init runs before MG_ConfigLoader::Init
     //     (see MG_Util/Debug/Log.cpp).
+    //   - MOBILEGL_ASYNC_POOL: a ShaderCompilePool is constructed by binaries that never call
+    //     MobileGL::Initialize() and so never run MG_ConfigLoader::Init - MG_Test's
+    //     JobNodeTest builds pools directly, and it is the suite that runs the whole async
+    //     matrix against both execution engines. Mirroring it here would resolve to the
+    //     default in exactly the tests that exist to tell the engines apart (see
+    //     MG_Util/Async/ShaderCompilePool.cpp, DetectAsyncPoolEngine).
     struct FeaturesTable {
         // MOBILEGL_DISABLE_TIMERQUERY: do not advertise or use GPU timer queries.
         Bool DisableTimerQuery = false;
@@ -91,6 +124,24 @@ namespace MobileGL::MG_Config {
         // feature off. It is enabled by default to match GL's defined out-of-range fetch
         // behavior; this escape hatch exists to measure or dodge its GPU cost on a device.
         Bool DisableRobustBufferAccess = false;
+        // MOBILEGL_MAGMA_MULTIDRAW_MODE: preferred DirectVulkan multi-draw dispatch tier
+        // ("ext" | "indirect" | "unroll", see MultiDrawMode). Clamped to device support;
+        // unset picks the best supported tier.
+        MultiDrawMode MagmaMultiDrawMode = MultiDrawMode::Auto;
+        // MOBILEGL_ESPRYT_MULTIDRAW_MODE: preferred DirectGLES glMultiDrawElements emulation
+        // tier ("ext" | "multiindirect" | "indirect" | "basevertex" | "drawelements" |
+        // "compute", see GLESMultiDrawMode). Clamped to driver support; unset picks the best
+        // supported tier, which never includes "compute" - see the note on its resolution.
+        GLESMultiDrawMode EsprytMultiDrawMode = GLESMultiDrawMode::Auto;
+        // MOBILEGL_ASYNC_SHADER_COMPILE: overrides asynchronous shader compilation. Unset
+        // keeps the built-in default (MG_Util::Async::kAsyncShaderCompileDefault); falsy
+        // forces every glCompileShader/glLinkProgram to run synchronously on the calling
+        // thread AND withdraws GL_KHR_parallel_shader_compile, so the single switch reverts
+        // both the threading and the application-visible behaviour change.
+        QuirkOverride AsyncShaderCompile = QuirkOverride::Auto;
+        // MOBILEGL_ASYNC_SHADER_COMPILE_THREADS: shader-compile worker count. 0 (unset) means
+        // auto, which is min(4, big cores); an explicit value is honoured as given.
+        Uint32 AsyncShaderCompileThreads = 0;
     };
     extern FeaturesTable Features;
 } // namespace MobileGL::MG_Config

MobileGL/ConfigLoader.cpp

@@ -97,6 +97,47 @@ namespace MobileGL::MG_ConfigLoader {
                                          : MG_Config::QuirkOverride::ForceOff;
     }
 
+    // Multi-draw mode is a named-value preference: unset keeps Auto (best supported tier),
+    // a recognized name selects that tier as the ceiling, anything else warns and keeps Auto.
+    inline MG_Config::MultiDrawMode QueryEnvMultiDrawMode(const String& key) {
+        auto it = acceptedEnvVariablesMap->find(key);
+        if (it == acceptedEnvVariablesMap->end()) {
+            return MG_Config::MultiDrawMode::Auto;
+        }
+        String lowered = it->second;
+        std::transform(lowered.begin(), lowered.end(), lowered.begin(),
+                       [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
+        if (lowered == "ext") return MG_Config::MultiDrawMode::Ext;
+        if (lowered == "indirect") return MG_Config::MultiDrawMode::Indirect;
+        if (lowered == "unroll") return MG_Config::MultiDrawMode::Unroll;
+        if (lowered.empty() || lowered == "auto") return MG_Config::MultiDrawMode::Auto;
+        MGLOG_W("Config: Ignoring invalid env variable %s='%s'; expected ext|indirect|unroll|auto, using auto",
+                key.c_str(), it->second.c_str());
+        return MG_Config::MultiDrawMode::Auto;
+    }
+
+    // Same contract as QueryEnvMultiDrawMode, over the DirectGLES tier names.
+    inline MG_Config::GLESMultiDrawMode QueryEnvGLESMultiDrawMode(const String& key) {
+        auto it = acceptedEnvVariablesMap->find(key);
+        if (it == acceptedEnvVariablesMap->end()) {
+            return MG_Config::GLESMultiDrawMode::Auto;
+        }
+        String lowered = it->second;
+        std::transform(lowered.begin(), lowered.end(), lowered.begin(),
+                       [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
+        if (lowered == "ext") return MG_Config::GLESMultiDrawMode::Ext;
+        if (lowered == "multiindirect") return MG_Config::GLESMultiDrawMode::MultiIndirect;
+        if (lowered == "indirect") return MG_Config::GLESMultiDrawMode::Indirect;
+        if (lowered == "basevertex") return MG_Config::GLESMultiDrawMode::BaseVertex;
+        if (lowered == "drawelements") return MG_Config::GLESMultiDrawMode::DrawElements;
+        if (lowered == "compute") return MG_Config::GLESMultiDrawMode::Compute;
+        if (lowered.empty() || lowered == "auto") return MG_Config::GLESMultiDrawMode::Auto;
+        MGLOG_W("Config: Ignoring invalid env variable %s='%s'; expected "
+                "ext|multiindirect|indirect|basevertex|drawelements|compute|auto, using auto",
+                key.c_str(), it->second.c_str());
+        return MG_Config::GLESMultiDrawMode::Auto;
+    }
+
     inline Uint32 QueryEnvUint32(const String& key, Uint32 defaultValue, Uint32 minValue, Uint32 maxValue) {
         auto it = acceptedEnvVariablesMap->find(key);
         if (it == acceptedEnvVariablesMap->end()) {
@@ -138,6 +179,10 @@ namespace MobileGL::MG_ConfigLoader {
         features.MagmaDisableBlendedDepthWriteQuirk =
             QueryEnvQuirkOverride("MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE");
         features.DisableRobustBufferAccess = QueryEnvFlag("MOBILEGL_DISABLE_ROBUST_BUFFER_ACCESS");
+        features.MagmaMultiDrawMode = QueryEnvMultiDrawMode("MOBILEGL_MAGMA_MULTIDRAW_MODE");
+        features.EsprytMultiDrawMode = QueryEnvGLESMultiDrawMode("MOBILEGL_ESPRYT_MULTIDRAW_MODE");
+        features.AsyncShaderCompile = QueryEnvQuirkOverride("MOBILEGL_ASYNC_SHADER_COMPILE");
+        features.AsyncShaderCompileThreads = QueryEnvUint32("MOBILEGL_ASYNC_SHADER_COMPILE_THREADS", 0, 0, 64);
     }
 
     inline void InitBackendType() {

MobileGL/Init.cpp

@@ -15,6 +15,8 @@
 #include <MG_Impl/GLImpl/Texture/ProxyTexture.h>
 #include <MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h>
 #include <MG_Impl/GLImpl/Sync/GL_Sync.h>
+#include <MG_Util/Async/ShaderCompilePool.h>
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
 
 #include <atomic>
 #include <mutex>
@@ -37,7 +39,12 @@ namespace MobileGL {
             if (logLifecycle) {
                 MGLOG_I("MobileGL closing...");
             }
-            glslang::FinalizeProcess();
+            // First, before anything else is torn down. In-flight compile/link jobs own
+            // their own inputs and are safe against everything below EXCEPT glslang's
+            // process globals and the TShader/TProgram objects hanging off pGLContext,
+            // both of which this function is about to destroy. This is the one
+            // cancellation path in the whole design that waits.
+            MG_Util::Async::ShaderCompilePool::Get().StopAndDrain();
             // GL syncs die with their contexts, and every context is gone by the
             // time full teardown runs: drain the live-sync registry while the
             // backend function table can still release the backend handles (and
@@ -49,6 +56,16 @@ namespace MobileGL {
             MG_State::pEGLContext.reset();
             MG_Impl::GLImpl::TextureImpl::pProxyTextureManager.reset();
             MG_Impl::GLImpl::FramebufferImpl::pDefaultFramebufferInfo.reset();
+            // Must run AFTER pGLContext.reset(). FinalizeProcess -> ShFinalize deletes
+            // glslang's process-wide pool allocator and every cached built-in symbol table,
+            // while the TShader/TProgram objects owned by the shader and program objects
+            // still reference levels adopted from those tables. Finalizing first left live
+            // glslang objects pointing at freed memory for the rest of the teardown.
+            glslang::FinalizeProcess();
+            // Immediately after, and never apart from it: FinalizeProcess just deleted the
+            // built-in symbol tables the prewarm latch stands for, so leaving it set would
+            // make the next Initialize() skip a prewarm it genuinely needs.
+            MG_Util::ShaderTranspiler::ShaderCompiler::ResetPrewarmLatch();
             MG_Backend::gBackendFunctionsTable = {};
             g_isInitialized = false;
             if (logLifecycle) {
@@ -76,6 +93,19 @@ namespace MobileGL {
         MG_Impl::Init();
         MGLOG_D("MG_Impl initialized");
         glslang::InitializeProcess();
+        // On the GL thread, before any worker can exist. glslang builds its built-in symbol
+        // tables lazily under a process-wide lock held for the whole build, so without this
+        // the first concurrent compiles of a shaderpack all serialize behind the very first
+        // parse and asynchronous compilation looks like it is doing nothing.
+        //
+        // Gated on the flag, because the problem it solves only exists when there are
+        // workers: with compilation synchronous, nothing ever contends for that lock and the
+        // three throwaway parses buy nothing - they just add to every eglInitialize. Read the
+        // flag here rather than inside PrewarmBuiltins so ShaderCompiler keeps no dependency
+        // on the async subsystem (ProgramUtilTest compiles that file without it).
+        if (MG_Util::Async::AsyncShaderCompileEnabled()) {
+            MG_Util::ShaderTranspiler::ShaderCompiler::PrewarmBuiltins();
+        }
         MGLOG_D("glslang initialized");
         g_isInitialized = true;
         MGLOG_I("MobileGL initialized");

MobileGL/MG_Backend/BackendObject.h

@@ -181,15 +181,18 @@ namespace MobileGL {
             void (*GetIntegeri_v)(GLenum target, GLuint index, GLint* data);
             void (*GetInteger64i_v)(GLenum target, GLuint index, GLint64* data);
             void (*GetProgramiv)(GLuint program, GLenum pname, GLint* params);
-            void (*GetProgramInterfaceiv)(GLuint program, GLenum programInterface, GLenum pname, GLint* params);
+            // The GL program interface (glGetProgramInterfaceiv / glGetProgramResource*) is NOT
-            GLuint (*GetProgramResourceIndex)(GLuint program, GLenum programInterface, const GLchar* name);
+            // a backend query: it describes the program the application wrote, in the
-            void (*GetProgramResourceName)(GLuint program, GLenum programInterface, GLuint index, GLsizei bufSize,
+            // application's namespace, which neither backend program is in. It is answered
-                                           GLsizei* length, GLchar* name);
+            // entirely by MG_Impl/GLImpl/Program/ProgramInterface from the frontend reflection.
-            void (*GetProgramResourceiv)(GLuint program, GLenum programInterface, GLuint index, GLsizei propCount,
+            // Takes the block's GL NAME, not glShaderStorageBlockBinding's index. The index
-                                         const GLenum* props, GLsizei bufSize, GLsizei* length, GLint* params);
+            // the application passes is the frontend interface-query enumeration's, and no
-            GLint (*GetProgramResourceLocation)(GLuint program, GLenum programInterface, const GLchar* name);
+            // backend shares that index space: DirectVulkan enumerates SPIR-V descriptor
-            GLint (*GetProgramResourceLocationIndex)(GLuint program, GLenum programInterface, const GLchar* name);
+            // bindings and DirectGLES asks a real driver about SPIRV-Cross-generated ESSL.
-            void (*ShaderStorageBlockBinding)(GLuint program, GLuint storageBlockIndex, GLuint storageBlockBinding);
+            // The name is the one coordinate all three agree on, so the frontend resolves the
+            // index against its own enumeration and each backend maps the name to its own.
+            void (*ShaderStorageBlockBinding)(GLuint program, const GLchar* storageBlockName,
+                                              GLuint storageBlockBinding);
             // GL fence sync objects. All entries are optional (may be null); the
             // frontend then falls back to always-signaled sync semantics.
             // FenceSync may itself return null when the backend cannot create a

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -18,6 +18,7 @@
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToStr/TextureEnumConverter.h>
 #include <MG_Util/Texture/TextureFormatProcessor.h>
+#include <MG_Util/Async/ShaderCompilePool.h>
 #include <Config.h>
 #include <algorithm>
 #include <cmath>
@@ -209,7 +210,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 reasons.push_back("GL_DEPTH_COMPONENT32 native probe failed on OpenGL ES");
             }
             if (options & PixelFormatNormalizeOptionBit::NoThreeChannelRenderTarget) {
-                reasons.push_back("no three-channel multisample storage format on OpenGL ES");
+                reasons.push_back("no colour-renderable three-channel format on OpenGL ES");
             }
             if (options & PixelFormatNormalizeOptionBit::NoSnorm16RenderTarget) {
                 reasons.push_back("EXT_render_snorm not supported");
@@ -552,26 +553,60 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
                 for (SizeT targetIndex = 0; targetIndex < kFormatCapabilityTextureTargetCount; ++targetIndex) {
                     const auto target = static_cast<TextureTarget>(targetIndex);
-                    // A multisample texture can only ever be rendered into, so its storage format
+                    // Colour-attachable targets need a colour-renderable fallback; the ordinary
-                    // has to stay colour-renderable; the ordinary fallback for a three-channel
+                    // fallback for a three-channel format is another three-channel one, which ES
-                    // format is a three-channel one, which ES accepts as a texture but rejects as
+                    // accepts as a texture but never as an attachment. Recompute the fallback per
-                    // multisample storage. Recompute the fallback per target so those formats get
+                    // target so those formats get widened where the target demands it.
-                    // widened here and nowhere else.
+                    const Flags<PixelFormatNormalizeOptionBit> renderTargetOptions =
-                    Flags<PixelFormatNormalizeOptionBit> targetOptions;
+                        TextureImpl::GetRenderTargetNormalizeOptions(capabilities, targetIndex);
-                    if (IsGLESProbeMultisampleTarget(target)) {
+                    // Multisample storage has no three-channel form on ES at all, so its widening
-                        targetOptions |= PixelFormatNormalizeOptionBit::NoThreeChannelRenderTarget;
+                    // is unconditional and skips the native probe (which cannot succeed). Every
-                        if (!capabilities.SupportsRenderSnorm || !capabilities.SupportsNorm16Texture) {
+                    // other target keeps the widening on the DRIVER branch, behind the native
-                            targetOptions |= PixelFormatNormalizeOptionBit::NoSnorm16RenderTarget;
+                    // probe: `shouldProbeFallback = !nativeCreated || !nativeRenderable` below is
-                        }
+                    // what makes the substitution conditional on the driver actually refusing, so
-                    }
+                    // a driver that does render to a three-channel image keeps allocating it byte
+                    // for byte. That is a per-format runtime answer, NOT a desktop-vs-device
+                    // split: llvmpipe renders to GL_RGB16F but refuses GL_RGB8_SNORM, GL_SRGB8,
+                    // GL_RGB32F and the RGB integer formats, so the CI driver widens those eight
+                    // too. Re-run the retrace fixtures and the glcts suites on any change here.
+                    const Bool widenUnconditionally = IsGLESProbeMultisampleTarget(target);
                     GLESProbeFormatInfo fallbackInfo = outerFallbackInfo;
                     Bool hasForcedFallback = outerHasForcedFallback;
-                    if (targetOptions) {
+                    if (renderTargetOptions) {
-                        hasForcedFallback = BuildFallbackProbeFormatInfo(
+                        // Folded into the forced options only when a forced fallback already
-                            requestedInternalFormat, forcedOptions | targetOptions, true, fallbackInfo);
+                        // applies, so the render-target bits never *create* one: ANGLE's forced
-                        if (!hasForcedFallback) {
+                        // GL_RGB8_SNORM -> GL_RGB16F is still three-channel and still needs
-                            BuildFallbackProbeFormatInfo(requestedInternalFormat, driverOptions | targetOptions, false,
+                        // widening, but a non-ANGLE driver must not lose its native probe.
+                        const Flags<PixelFormatNormalizeOptionBit> forcedProbeOptions =
+                            (outerHasForcedFallback || widenUnconditionally) ? forcedOptions | renderTargetOptions
+                                                                             : forcedOptions;
+                        hasForcedFallback =
+                            BuildFallbackProbeFormatInfo(requestedInternalFormat, forcedProbeOptions, true,
                                                          fallbackInfo);
+                        if (!hasForcedFallback) {
+                            BuildFallbackProbeFormatInfo(requestedInternalFormat,
+                                                         driverOptions | renderTargetOptions, false, fallbackInfo);
+                        }
+                        // HONEST STATUS OF THE FORCED PATH. A forced fallback is only ever built
+                        // for ANGLE (GetForcedPixelFormatNormalizeOptions returns nothing for any
+                        // other renderer), and it SKIPS the native probe entirely - the widened
+                        // format is asserted rather than measured on this device. That assertion
+                        // is validated on exactly one configuration, the android-angle retrace
+                        // golden; it is NOT covered by the headless llvmpipe suites, which take
+                        // the driver branch below and prove nothing about ANGLE's answers. So log
+                        // the choice at INFO rather than the usual MGLOG_D caveat: on any other
+                        // ANGLE device the device report is the only evidence there is of which
+                        // storage format the image really got. Once per format on the ordinary 2D
+                        // target - repeating it for all ten targets would bury the report.
+                        if (hasForcedFallback && target == TextureTarget::Texture2D &&
+                            (MG_Util::TextureFormatProcessor::GetApplicablePixelFormatNormalizeOptions(
+                                 requestedInternalFormat, renderTargetOptions) &
+                             PixelFormatNormalizeOptionBit::NoThreeChannelRenderTarget)) {
+                            MGLOG_I("Three-channel widening (FORCED path, no native probe): %s stored as %s. "
+                                    "Reason: %s. Device-validated on the android-angle golden only.",
+                                    MG_Util::ConvertTextureInternalFormatToString(logicalFormat).c_str(),
+                                    ConvertFallbackInternalFormatToString(fallbackInfo.InternalFormat).c_str(),
+                                    fallbackInfo.Reason.c_str());
                         }
                     }
 
@@ -617,8 +652,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 }
 
                 const SizeT renderbufferTargetIndex = GetRenderbufferFormatCapabilityTargetIndex();
-                Bool shouldProbeFallbackRenderbuffer = outerHasForcedFallback;
+                // A renderbuffer exists only to be attached, so it needs the same three-channel
-                if (!outerHasForcedFallback) {
+                // widening the colour-attachable texture targets get - and on the same terms: the
+                // native storage is probed first, so a driver that renders to it keeps it.
+                const Flags<PixelFormatNormalizeOptionBit> renderbufferOptions =
+                    TextureImpl::GetRenderTargetNormalizeOptions(capabilities, renderbufferTargetIndex);
+                GLESProbeFormatInfo renderbufferFallbackInfo = outerFallbackInfo;
+                Bool renderbufferHasForcedFallback = outerHasForcedFallback;
+                if (renderbufferOptions) {
+                    const Flags<PixelFormatNormalizeOptionBit> forcedProbeOptions =
+                        outerHasForcedFallback ? forcedOptions | renderbufferOptions : forcedOptions;
+                    renderbufferHasForcedFallback = BuildFallbackProbeFormatInfo(
+                        requestedInternalFormat, forcedProbeOptions, true, renderbufferFallbackInfo);
+                    if (!renderbufferHasForcedFallback) {
+                        BuildFallbackProbeFormatInfo(requestedInternalFormat, driverOptions | renderbufferOptions,
+                                                     false, renderbufferFallbackInfo);
+                    }
+                }
+
+                Bool shouldProbeFallbackRenderbuffer = renderbufferHasForcedFallback;
+                if (!renderbufferHasForcedFallback) {
                     const Bool nativeRenderbufferComplete =
                         ProbeRenderbuffer(gl, nativeInfo.InternalFormat, logicalFormat, false, 1);
                     if (nativeRenderbufferComplete) {
@@ -632,16 +685,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
                         shouldProbeFallbackRenderbuffer = true;
                     }
                 }
-                if (shouldProbeFallbackRenderbuffer && outerFallbackInfo.InternalFormat != GL_UNKNOWN_MGL &&
+                if (shouldProbeFallbackRenderbuffer && renderbufferFallbackInfo.InternalFormat != GL_UNKNOWN_MGL &&
-                    ProbeRenderbuffer(gl, outerFallbackInfo.InternalFormat, logicalFormat, false, 1)) {
+                    ProbeRenderbuffer(gl, renderbufferFallbackInfo.InternalFormat, logicalFormat, false, 1)) {
                     if (AddCaveatFormatCaps(cache, renderbufferTargetIndex, formatIndex,
                                             GetRenderbufferFeatureCaps(logicalFormat))) {
-                        LogGLESFormatCaveat(logicalFormat, renderbufferTargetIndex, outerFallbackInfo);
+                        LogGLESFormatCaveat(logicalFormat, renderbufferTargetIndex, renderbufferFallbackInfo);
                     }
                     const Int maxSamples =
-                        GetGLESFormatMaxSamples(capabilities, logicalFormat, outerFallbackInfo.ImageFormat);
+                        GetGLESFormatMaxSamples(capabilities, logicalFormat, renderbufferFallbackInfo.ImageFormat);
-                    cache.SampleCounts[renderbufferTargetIndex][formatIndex] =
+                    cache.SampleCounts[renderbufferTargetIndex][formatIndex] = ProbeRenderbufferSampleCounts(
-                        ProbeRenderbufferSampleCounts(gl, outerFallbackInfo.InternalFormat, logicalFormat, maxSamples);
+                        gl, renderbufferFallbackInfo.InternalFormat, logicalFormat, maxSamples);
                 }
             }
         }
@@ -891,6 +944,22 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // GL_KHR_parallel_shader_compile is MobileGL's own capability, not the host ES
+        // driver's: the compiler threads are MobileGL's, and glCompileShader/glLinkProgram
+        // are serviced entirely inside the frontend. Whether the device driver advertises
+        // the string is irrelevant here (the POST reports it separately, for the day the
+        // driver-side link is what gets parallelised).
+        //
+        // Gated on the async flag deliberately, and this is the whole reason the gate
+        // exists. Advertising the string is the one part of asynchronous compilation that a
+        // recorded trace can never cover: Iris and Sodium change their SUBMISSION SCHEDULE
+        // the moment they see it - they enqueue whole pipeline batches and poll
+        // GL_COMPLETION_STATUS_KHR instead of compiling one program at a time - so
+        // MOBILEGL_ASYNC_SHADER_COMPILE=0 has to withdraw the application-visible behaviour
+        // change as well as the threading, or the kill switch would only be half a switch.
+        if (MG_Util::Async::AsyncShaderCompileEnabled()) {
+            extensions.push_back(E_GL_KHR_parallel_shader_compile);
+        }
         // Only advertised when the device driver actually has usable timer queries
         // (GL_EXT_disjoint_timer_query plus its entry points) and the
         // MOBILEGL_DISABLE_TIMERQUERY escape hatch is off.
@@ -951,12 +1020,6 @@ namespace MobileGL::MG_Backend::DirectGLES {
             funcsTable.GL.GetIntegeri_v = GetIntegeri_v;
             funcsTable.GL.GetInteger64i_v = GetInteger64i_v;
             funcsTable.GL.GetProgramiv = GetProgramiv;
-            funcsTable.GL.GetProgramInterfaceiv = GetProgramInterfaceiv;
-            funcsTable.GL.GetProgramResourceIndex = GetProgramResourceIndex;
-            funcsTable.GL.GetProgramResourceName = GetProgramResourceName;
-            funcsTable.GL.GetProgramResourceiv = GetProgramResourceiv;
-            funcsTable.GL.GetProgramResourceLocation = GetProgramResourceLocation;
-            funcsTable.GL.GetProgramResourceLocationIndex = GetProgramResourceLocationIndex;
             funcsTable.GL.ShaderStorageBlockBinding = ShaderStorageBlockBinding;
             funcsTable.GL.Clear = Clear;
             funcsTable.GL.ClearBufferfi = ClearBufferfi;

MobileGL/MG_Backend/DirectGLES/DirectGLES.h

@@ -92,15 +92,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
     void GetIntegeri_v(GLenum target, GLuint index, GLint* data);
     void GetInteger64i_v(GLenum target, GLuint index, GLint64* data);
     void GetProgramiv(GLuint program, GLenum pname, GLint* params);
-    void GetProgramInterfaceiv(GLuint program, GLenum programInterface, GLenum pname, GLint* params);
+    void ShaderStorageBlockBinding(GLuint program, const GLchar* storageBlockName, GLuint storageBlockBinding);
-    GLuint GetProgramResourceIndex(GLuint program, GLenum programInterface, const GLchar* name);
-    void GetProgramResourceName(GLuint program, GLenum programInterface, GLuint index, GLsizei bufSize, GLsizei* length,
-                                GLchar* name);
-    void GetProgramResourceiv(GLuint program, GLenum programInterface, GLuint index, GLsizei propCount,
-                              const GLenum* props, GLsizei bufSize, GLsizei* length, GLint* params);
-    GLint GetProgramResourceLocation(GLuint program, GLenum programInterface, const GLchar* name);
-    GLint GetProgramResourceLocationIndex(GLuint program, GLenum programInterface, const GLchar* name);
-    void ShaderStorageBlockBinding(GLuint program, GLuint storageBlockIndex, GLuint storageBlockBinding);
     Bool InitWindowSurface(NativeWindowType window);
     Bool InitPbufferSurface(EGLint width, EGLint height);
     Bool MakeCurrent();
@@ -182,6 +174,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
     namespace XfbImpl {
         Bool AreTransformFeedbacksSupported();
+        // True while a capture span is open on the current transform feedback object
+        // (frontend Begin seen and not paused), whether or not the deferred driver-side
+        // Begin has been issued yet. Draw paths that would restructure the primitive
+        // stream, or that need to dispatch compute mid-draw, decline while it is set.
+        Bool IsCaptureSpanOpen();
         void BeginTransformFeedback(GLenum primitiveMode);
         void EndTransformFeedback();
         void PauseTransformFeedback();
@@ -191,6 +188,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
         void OnBackendContextDestroyed();
     } // namespace XfbImpl
 
+    namespace RenderStateImpl {
+        // Pushes the frontend's render-state block to the ES driver, diffed against what was
+        // last pushed.
+        //
+        // `forColorClear` names the CALLER, and the only thing it changes is the colour write
+        // mask handed to the driver. A draw into a colour attachment the backend widened from
+        // three channels to four gets that buffer's alpha channel masked OFF, so nothing can
+        // move the stored alpha away from the 1.0 the application's three-channel format
+        // implies (see FramebufferImpl::g_alphaWidenedDrawBufferMask). A CLEAR is how that 1.0
+        // gets there in the first place, so it must be allowed to write alpha - hence the flag
+        // rather than an unconditional doctoring. It is part of the sync memo, so a clear
+        // followed by a draw re-pushes the mask instead of early-outing on an unchanged
+        // frontend version.
+        //
+        // The application's own colour mask is never modified: glGet(GL_COLOR_WRITEMASK)
+        // answers from the frontend state, which this function only reads.
+        void SyncRenderState(Bool forColorClear = false);
+        void InvalidateSyncedRenderState();
+    } // namespace RenderStateImpl
+
     extern MG_External::EGLFunctionsTable g_EGLFuncs;
     extern MG_External::GLESFunctionsTable g_GLESFuncs;
     extern MG_External::GLESCapabilities g_GLESCapabilities;

MobileGL/MG_Backend/DirectGLES/Managers.h

@@ -21,45 +21,126 @@ namespace MobileGL::MG_Backend::DirectGLES {
     String EmulateBaseInstanceInVertexShader(String source, GLenum shaderType);
     String PromoteDrawParameterGlobalsToUniforms(String source, GLenum shaderType);
 
+    // True once the process has entered exit(): past that point the EGL library and
+    // the driver may already be unloaded, so a backend twin's destructor must not
+    // call into g_GLESFuncs (the observed crash is a jump through an unmapped driver
+    // pointer from __run_exit_handlers) nor touch statics in other TUs (cross-TU
+    // destruction order is unspecified). Deliberate leak: the process is exiting and
+    // the driver reclaims GPU objects. The flag is set by a std::atexit handler that
+    // EnsureProcessTeardownSentinel() registers lazily on first registry use - by
+    // then every static everywhere has finished constructing, so this handler is
+    // guaranteed to run BEFORE any static destructor (atexit is LIFO). A destructor
+    // hook on the registry itself was tried first and is WRONG: tests and cache
+    // resets destroy temporary registry instances mid-run, which would latch the
+    // flag while the process is very much alive.
+    Bool InProcessTeardown();
+    void EnsureProcessTeardownSentinel();
+
+    // Which optional pieces of state a draw needs synchronized before it is issued.
+    // Index/indirect buffer syncs and the instancing-related work are skipped for
+    // draws that provably cannot read them.
+    enum class DrawSyncBit : Uint32 {
+        None = 0,
+        IndexBuffer = 1 << 0,
+        IndirectBuffer = 1 << 1,
+        Instancing = 1 << 2
+    };
+    // Deliberately the shared Flags<> rather than hand-written operators for this enum:
+    // a namespace-local operator| here would hide MobileGL::operator|(Bit, Bit) from
+    // every other scoped-enum flag set used inside this namespace.
+    using DrawSyncFlags = Flags<DrawSyncBit>;
+
+    // The GL-defined indirect command layouts, byte-identical to what the driver reads
+    // out of a GL_DRAW_INDIRECT_BUFFER. Also the staging layout the multi-draw emulation
+    // synthesizes commands into.
+    struct DrawElementsIndirectCommand {
+        Uint32 count = 0;
+        Uint32 instanceCount = 0;
+        Uint32 firstIndex = 0;
+        Int32 baseVertex = 0;
+        Uint32 baseInstance = 0;
+    };
+
+    struct DrawArraysIndirectCommand {
+        Uint32 count = 0;
+        Uint32 instanceCount = 0;
+        Uint32 first = 0;
+        Uint32 baseInstance = 0;
+    };
+
+    // Brings the whole draw-relevant frontend state onto the native ES context and binds
+    // the program; every GL draw entry point calls it exactly once before issuing draws.
+    void PrepareForDraw(DrawSyncFlags syncBits);
+    // GLES core supports only GL_PRIMITIVE_RESTART_FIXED_INDEX. Throws when the app enabled
+    // the arbitrary GL_PRIMITIVE_RESTART with a non-fixed index for this index type.
+    void CheckPrimitiveRestartSupported(GLenum indexType);
+    // Feed the current program's gl_BaseInstance / gl_DrawID emulation uniforms. Both are
+    // no-ops when the program does not read the corresponding builtin.
+    void SetCurrentBaseInstance(Uint32 baseInstance);
+    void SetCurrentDrawID(Uint32 drawId);
+    // True when the current program actually reads gl_DrawID, i.e. when a batched
+    // (single driver call) multi-draw tier would have to feed it one value for the whole
+    // batch and would therefore be wrong.
+    Bool CurrentProgramReadsDrawID();
+
     template <typename StateObject, typename BackendObject>
     class StateBackendObjectRegistry {
     public:
+
         using StatePtr = SharedPtr<StateObject>;
         using StateWeakPtr = std::weak_ptr<StateObject>;
         using BackendPtr = SharedPtr<BackendObject>;
-        using BackendMap = UnorderedMap<StateObject*, BackendPtr>;
+
-        using StateRefMap = UnorderedMap<StateObject*, StateWeakPtr>;
+        // The backend twin and the weak reference that decides whether the raw key still
+        // names the state object the twin was built for. Both live in one entry: a
+        // separate liveness map answered nothing the backend probe had not already found
+        // and cost a second hash lookup on every Find, which the draw path runs ~10 times.
+        struct Entry {
+            BackendPtr backend;
+            StateWeakPtr stateRef;
+        };
+        using BackendMap = UnorderedMap<StateObject*, Entry>;
         using iterator = typename BackendMap::iterator;
         using const_iterator = typename BackendMap::const_iterator;
 
         BackendPtr& GetOrCreate(const StatePtr& stateObj) {
             MOBILEGL_ASSERT(stateObj != nullptr, "State object must not be null");
 
-            auto* key = stateObj.get();
+            // Twin creation is the moment a driver-owned id starts needing a guarded
-            auto trackedStateIt = m_stateRefs.find(key);
+            // destructor; cold path, so the once-guard costs nothing per draw.
-            if (trackedStateIt != m_stateRefs.end() && trackedStateIt->second.expired()) {
+            EnsureProcessTeardownSentinel();
-                EraseByKey(key);
+            auto& entry = m_entries[stateObj.get()];
+            if (entry.stateRef.expired()) {
+                // The previous owner of this address is gone and the allocator handed it
+                // to a new object: its twin describes ids the new state object never made.
+                entry.backend.reset();
             }
-            m_stateRefs[key] = stateObj;
+            entry.stateRef = stateObj;
-            return m_backendObjects[key];
+            return entry.backend;
         }
 
-        iterator find(StateObject* stateObj) {
+        // Null when no live state object owns this key. The result points into the map, so
-            if (!IsAlive(stateObj)) {
+        // it stays valid only until the next GetOrCreate/Find/CollectGarbage on this registry.
-                EraseByKey(stateObj);
+        BackendPtr* Find(StateObject* stateObj) {
-                return m_backendObjects.end();
+            const auto entryIt = m_entries.find(stateObj);
+            if (entryIt == m_entries.end()) {
+                return nullptr;
             }
-            return m_backendObjects.find(stateObj);
+            if (entryIt->second.stateRef.expired()) {
+                m_entries.erase(entryIt);
+                return nullptr;
+            }
+            return &entryIt->second.backend;
         }
 
-        const_iterator find(StateObject* stateObj) const {
+        const BackendPtr* Find(StateObject* stateObj) const {
-            return const_cast<StateBackendObjectRegistry*>(this)->find(stateObj);
+            return const_cast<StateBackendObjectRegistry*>(this)->Find(stateObj);
         }
 
-        iterator begin() { return m_backendObjects.begin(); }
+        iterator begin() { return m_entries.begin(); }
-        const_iterator begin() const { return m_backendObjects.begin(); }
+        const_iterator begin() const { return m_entries.begin(); }
-        iterator end() { return m_backendObjects.end(); }
+        iterator end() { return m_entries.end(); }
-        const_iterator end() const { return m_backendObjects.end(); }
+        const_iterator end() const { return m_entries.end(); }
 
         void CollectGarbageIfNeeded() {
             ++m_gcTick;
@@ -73,19 +154,6 @@ namespace MobileGL::MG_Backend::DirectGLES {
         void CollectGarbageNow() { CollectGarbage(); }
 
     private:
-        bool IsAlive(StateObject* stateObj) const {
-            const auto trackedStateIt = m_stateRefs.find(stateObj);
-            if (trackedStateIt == m_stateRefs.end()) {
-                return false;
-            }
-            return !trackedStateIt->second.expired();
-        }
-
-        void EraseByKey(StateObject* stateObj) {
-            m_stateRefs.erase(stateObj);
-            m_backendObjects.erase(stateObj);
-        }
-
         void CollectGarbage() {
             if (m_isCollecting) {
                 return;
@@ -94,16 +162,15 @@ namespace MobileGL::MG_Backend::DirectGLES {
             m_isCollecting = true;
 
             Vector<StateObject*> staleKeys;
-            staleKeys.reserve(m_stateRefs.size());
+            staleKeys.reserve(m_entries.size());
-            for (const auto& [stateKey, stateWeakRef] : m_stateRefs) {
+            for (const auto& [stateKey, entry] : m_entries) {
-                if (stateWeakRef.expired()) {
+                if (entry.stateRef.expired()) {
                     staleKeys.push_back(stateKey);
                 }
             }
 
             for (auto* stateKey : staleKeys) {
-                m_stateRefs.erase(stateKey);
+                m_entries.erase(stateKey);
-                m_backendObjects.erase(stateKey);
             }
 
             m_isCollecting = false;
@@ -111,8 +178,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
     private:
         static constexpr Uint32 kGCInterval = 1024;
-        StateRefMap m_stateRefs;
+        BackendMap m_entries;
-        BackendMap m_backendObjects;
         Uint32 m_gcTick = 0;
         Bool m_isCollecting = false;
     };
@@ -120,6 +186,43 @@ namespace MobileGL::MG_Backend::DirectGLES {
     namespace BufferImpl {
         const GLenum TempBufferTarget = GL_ARRAY_BUFFER;
 
+        // --- Buffer-mutation epoch -------------------------------------------------
+        // Manager-wide monotonic counter: it moves whenever ANY buffer resource may
+        // have gone from draw-clean to dirty. Draw-path memos read it once per pass
+        // (CurrentBufferMutationEpoch, acquire), re-run their IsBufferDrawClean
+        // probes only when it moved, and stamp the PRE-pass value after a pass in
+        // which every probe came up clean - so a concurrent bump lands strictly
+        // after the stamped value and forces a re-probe on the next pass no matter
+        // how the probe interleaved with the mutation. Conservative-correct: a bump
+        // never skips work, it only re-runs the probes once.
+        //
+        // Every clean->dirty transition path bumps it (BumpBufferMutationEpoch,
+        // release, AFTER the mutation lands so an acquire reader that still sees
+        // the old epoch cannot have missed the mutation):
+        //   * the frontend BufferBackendOps table - Respecify, SubData,
+        //     FlushMappedRange, AcquirePersistentMap, ReadbackFromGpu, OnDestroy -
+        //     which every frontend change-serial bump and every pending-range
+        //     queueing reaches while ops are registered (upload, orphan/respecify,
+        //     map flush/unmap writeback, persistent-map adoption, delete/pooling);
+        //   * backend-initiated shadow writebacks that bump the frontend change
+        //     serial without an op: transform-feedback capture readback
+        //     (XfbImpl::ReadbackCapturedRanges and the scatter path) and every
+        //     pack-PBO WritebackFromBackend site (glReadPixels/glGetTexImage);
+        //   * RegisterBufferBackendOps/UnregisterBufferBackendOps - while ops are
+        //     unregistered, frontend writes advance serials silently, so both edges
+        //     of that window re-open every memo;
+        //   * OnBackendContextDestroyed - the buffer context generation moved, so
+        //     every previously clean resource is invalid.
+        // NOT bumped (cleanliness provably unchanged): MarkGpuWritten (the backend
+        // copy is authoritative; IsBufferDrawClean does not consult it),
+        // NotifyContentWrite on a GPU-resident buffer (persistent-mapped resources
+        // are clean by construction), and EnsureBufferResource itself (it only
+        // repairs toward clean). A non-persistent map (draws on it are GL errors
+        // the frontend rejects) sets IsMapped without an op; persistent maps reach
+        // AcquirePersistentMap or (FLUSH_EXPLICIT) publish only via FlushMappedRange.
+        Uint64 CurrentBufferMutationEpoch();
+        void BumpBufferMutationEpoch();
+
         // The DirectGLES storage behind one frontend buffer. Owned (refcounted) by
         // the frontend BufferObject; immediate BufferBackendOps keep it current, so
         // draw-time "sync" reduces to ensuring the storage exists.
@@ -145,6 +248,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Bool pendingRespecify = false;
             VecRange1D pendingRanges;
             std::mutex pendingMutex;
+            // Buffer-mutation epoch (see CurrentBufferMutationEpoch) at which this
+            // resource last probed IsBufferDrawClean == true, 0 = never (epochs start
+            // at 1). Written only on the draw thread; per-draw resource consumers
+            // (the UBO binding walk) skip the probe while their pre-pass epoch read
+            // matches, exactly like the per-VAO memo stamps.
+            Uint64 drawCleanEpoch = 0;
             // Zero-copy coherent persistent map (EXT_buffer_storage): the GL store is
             // immutable, persistently+coherently mapped, and persistentPtr is what the app
             // (and the frontend PipeResource) write into directly. While set, draw-time
@@ -170,6 +279,17 @@ namespace MobileGL::MG_Backend::DirectGLES {
         GLESBufferResource* EnsureBufferResource(const SharedPtr<MG_State::GLState::BufferObject>& bufferObject);
         // Existing resource or nullptr; performs no GL calls.
         GLESBufferResource* GetBufferResource(MG_State::GLState::BufferObject* bufferObject);
+        // True when EnsureBufferResource(frontend) would provably fall straight through
+        // every branch and do no work — i.e. `resource` is still the frontend's own
+        // resource, its id belongs to the live ES context, and either it is the
+        // zero-copy coherent persistent store (draw-time sync is a no-op by design) or
+        // the storage is initialized at the right size with no pending ops and a synced
+        // change serial while the buffer is not mapped (an active map may owe a
+        // per-draw persistent-range push, so it always takes the full path).
+        // `frontend` must be non-null and alive; the caller guarantees that by holding
+        // (or shadowing something that holds) a SharedPtr to it. Enables the per-VAO
+        // resolved-buffers memo to skip EnsureBufferResource on clean static buffers.
+        Bool IsBufferDrawClean(const MG_State::GLState::BufferObject* frontend, const GLESBufferResource* resource);
 
         // Deletes GL buffers whose owning frontend objects died (possibly on a
         // thread without a current ES context). Called from draw-time sync.
@@ -255,11 +375,70 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Uint GetBackendVertexArrayId() const { return m_backendVAOId; }
             void Bind() const;
 
+            // Draw-path memo of SyncNeccessaryBuffers' attribute walk for this VAO: the
+            // distinct enabled-attribute buffers (deduped) and the index buffer, resolved
+            // to their backend resources once. Valid while the VAO's config version is
+            // unchanged — every attach/enable/disable/format mutation bumps it (the same
+            // invariant SyncToBackend's gate already leans on), and the VAO's attribute
+            // SharedPtrs pin each memoed frontend buffer for exactly that long, so the raw
+            // pointers cannot dangle on a hit. Per-buffer cleanliness is NOT memoed here:
+            // each hit re-checks IsBufferDrawClean (resource identity, context generation,
+            // pending ops, change serial) and falls back to EnsureBufferResource for just
+            // the dirty entries via their attribute index. The IBO entry is keyed on the
+            // slot's bound-object identity instead (its slot version is a wrapping Uint16
+            // and is not covered by the config version).
+            struct ResolvedDrawBuffers {
+                struct Entry {
+                    MG_State::GLState::BufferObject* frontend = nullptr;
+                    BufferImpl::GLESBufferResource* resource = nullptr;
+                    Uint8 attribIndex = 0;
+                };
+                Bool valid = false;
+                Uint32 configVersion = 0;
+                Uint count = 0;
+                Array<Entry, MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS> entries;
+                MG_State::GLState::BufferObject* iboFrontend = nullptr;
+                BufferImpl::GLESBufferResource* iboResource = nullptr;
+                // Buffer-mutation epoch (BufferImpl::CurrentBufferMutationEpoch) at which
+                // the LAST probe pass found every entry / the IBO clean; 0 = not stamped
+                // (epochs start at 1). While a stamp matches the pre-pass epoch read, the
+                // probes are skipped outright: any path that can dirty ANY buffer bumps
+                // the epoch (the exhaustive site list lives at the epoch declaration).
+                // The IBO stamp is only trusted together with the bound-object identity
+                // compare - the VAO's index slot can rebind with no epoch or config move.
+                Uint64 vboCleanEpoch = 0;
+                Uint64 iboCleanEpoch = 0;
+            };
+            ResolvedDrawBuffers& GetResolvedDrawBuffersMemo() { return m_resolvedDrawBuffers; }
+
+            // Memo for SyncCurrentVertexAttributeValues: which of a program's ACTIVE
+            // attribute locations lack an enabled array in this VAO (those read the
+            // context's current generic value instead of a buffer). Keyed on the VAO
+            // config version (enable/disable bumps it) and the program's active-location
+            // mask. Hosted per twin — the former function-static single entry missed on
+            // every draw once the app cycled VAOs, re-reading the cold attribute slots.
+            struct PendingAttribValueMask {
+                Bool valid = false;
+                Uint32 configVersion = 0;
+                Uint32 activeMask = 0;
+                Uint32 pendingMask = 0;
+            };
+            PendingAttribValueMask& GetPendingAttribValueMaskMemo() { return m_pendingAttribValueMask; }
+
         private:
+            ResolvedDrawBuffers m_resolvedDrawBuffers;
+            PendingAttribValueMask m_pendingAttribValueMask;
             Uint m_backendVAOId = 0;
             Array<Uint, MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS> m_clientAttributeBufferIds;
             Bool m_isInitialized = false;
             Uint16 m_syncedIndexBufferVersion = 0;
+            // Aggregate gate over the per-attribute walk below: the frontend bumps its config
+            // version on every per-attribute version bump (the three Bump*Version functions are
+            // its only writers), so an unchanged config version proves every per-attribute
+            // compare in SyncToBackend would come up clean. The index-buffer slot has its own
+            // version and is NOT covered. The Bool (not a sentinel value) marks "never synced".
+            Bool m_hasSyncedConfigVersion = false;
+            Uint32 m_syncedConfigVersion = 0;
             Array<MG_State::GLState::VertexAttributeVersion, MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS>
                 m_syncedAttributeVersions;
         };
@@ -336,6 +515,25 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return target == TextureTarget::Texture3D || target == TextureTarget::TextureCubeMap;
         }
 
+        // Components per texel the frontend format's client data carries, for the three-channel
+        // formats that can be widened to a four-channel colour-renderable target; 0 for everything
+        // else. See PrepareChannelWidenedUpload.
+        Uint GetWidenableClientComponentCount(TextureInternalFormat format);
+
+        // True when a widenable format's components are integer rather than normalized, which is
+        // what decides the synthetic alpha's value: GL_RGB8I and GL_RGB8_SNORM are both uploaded
+        // as GL_BYTE, but their 1.0 is 1 and 0x7F respectively.
+        Bool IsIntegerWidenableFormat(TextureInternalFormat format);
+
+        // Repacks three-component client data as four components with an alpha of 1.0 in
+        // `uploadType`, for a format the backend widened to keep a colour attachment renderable.
+        // Returns `data` untouched when no widening applies. Pure CPU and context-free so a unit
+        // test can exercise the exact packing the driver is handed; `widenedData` is the caller's
+        // scratch buffer and has to outlive the returned pointer.
+        const void* PrepareChannelWidenedUpload(Uint componentCount, const IntVec3& texelSize, const void* data,
+                                                SizeT byteSize, GLenum uploadType, Vector<Uint8>& widenedData,
+                                                Bool integerData = false);
+
         struct StateTextureBasicInfo { // Used for tracking texture state changes
             TextureInternalFormat internalFormat = TextureInternalFormat::Unknown;
             SizeT width = 0;
@@ -373,6 +571,38 @@ namespace MobileGL::MG_Backend::DirectGLES {
             void Bind(GLenum target, Uint unit = TempTextureUnit);
             Uint GetBackendTextureId() const;
 
+            // Aggregate first-level clean gate for the per-draw trio
+            // SyncTextureParamsToBackend + SyncBuiltinSamplerToBackend +
+            // SyncMipmapsToBackend: EXACTLY the conjunction of their own early-outs
+            // (params version == synced params version; builtin-sampler version ==
+            // synced sampler version; and SyncMipmapsToBackend's cheap gate - stamped
+            // trio + content version + Mipmap storage). True means each of the three
+            // would provably return without work, so the caller may skip the calls;
+            // false only falls through to the three calls, whose own gates re-decide
+            // individually - this gate must never be MORE permissive than they are.
+            // `contextId`/`samplingGeneration` are the frontend context's current
+            // values, hoisted by the caller so a per-draw list walk reads them once
+            // instead of per texture. `t` must be the live frontend texture.
+            Bool IsDrawSyncClean(const MG_State::GLState::ITextureObject* t, Uint64 contextId,
+                                 Uint64 samplingGeneration) const {
+                if (!m_isInitialized || m_syncedShapeContextId == 0 || m_syncedShapeContextId != contextId ||
+                    m_syncedShapeGeneration != samplingGeneration) {
+                    return false;
+                }
+                const Uint16 paramsVersion = t->GetTextureParamsVersion();
+                if (m_syncedShapeParamsVersion != paramsVersion || m_syncedTextureParamsVersion != paramsVersion) {
+                    return false;
+                }
+                if (m_syncedContentVersion == 0 || m_syncedContentVersion != t->GetContentVersion()) {
+                    return false;
+                }
+                const auto& samplerObject = t->GetSamplerObject();
+                if (!samplerObject || m_syncedSamplerVersion != samplerObject->GetVersion()) {
+                    return false;
+                }
+                return t->GetStorageType() == TextureStorageType::Mipmap;
+            }
+
         private:
             void RecreateBackendTexture();
 
@@ -388,6 +618,21 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // clean probe compares this before rebuilding shape info and scanning
             // per-level dirty flags; 0 never matches a real version (they start at 1).
             Uint64 m_syncedContentVersion = 0;
+            // First-level clean gate for SyncMipmapsToBackend, checked before even the
+            // IsComplete()/shape-probe walk. Valid only as a trio with the content and
+            // texture-params versions: the context's sampling-resolution generation moves on
+            // EVERY texture-shape mutation (BumpShapeVersion is the only writer of shape and
+            // unconditionally bumps it), the content version on every CPU pixel mutation, and
+            // the params version covers SetSamples/SetFixedSampleLocations, which bump neither
+            // of the other two but feed the shape probe. The context id pins the generation to
+            // the context that produced it - generations restart at 0 with a new context, and a
+            // texture is owned by exactly one context (share groups are not implemented), so a
+            // mutation can never happen under a context this key does not name. 0 = never
+            // stamped (real context ids start at 1). Backend-side invalidation rides on
+            // m_isInitialized: RequireImageBindableStorage and RecreateBackendTexture clear it.
+            Uint64 m_syncedShapeContextId = 0;
+            Uint64 m_syncedShapeGeneration = 0;
+            Uint16 m_syncedShapeParamsVersion = 0;
             SamplerParameters m_cacheSamplerParameters;
             UintVec2 m_cacheLodRange = {0, 1000};
             FloatVec4 m_cacheBorderColor = {0.0f, 0.0f, 0.0f, 0.0f};
@@ -404,6 +649,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
         SharedPtr<BackendTextureObject>& SyncTextureObjectToBackend(
             const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
             Bool imageBindableStorageRequired = false);
+        // Brings every texture the next draw reads - the touched units' bindings and the draw
+        // FBO's texture attachments - onto the backend, through the two borrowed-pair memos
+        // documented at their definitions. Declared here so tests can drive those memos directly.
+        void SyncNeccessaryTextures();
         extern Array<Array<BackendTextureObject*, (SizeT)TextureTarget::TextureTargetCount>,
                      MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS>
             g_boundTexturesCache;
@@ -484,11 +733,81 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // has to apply the clamp itself.
         Bool IsFixedPointFallbackReadAttachment();
 
-        extern Array<Uint16, SizeT(FramebufferTarget::FramebufferTargetCount)> g_fboBindVersions;
+        // True when the read buffer names a three-channel attachment the backend actually stores
+        // in a four-channel format (the colour-renderable widening). A format without alpha reads
+        // back as 1.0, so the readback path has to overwrite the alpha the draw left behind -
+        // unconditionally, since this is the format's own semantics rather than the
+        // GL_CLAMP_READ_COLOR rule the clamp above implements.
+        Bool IsAlphaWidenedFallbackReadAttachment();
+
+        // True when this attachment's storage carries an alpha channel its frontend format does
+        // not (the three-channel colour-renderable widening).
+        Bool IsAlphaWidenedColorAttachment(const MG_State::GLState::FramebufferAttachmentObject& attachmentObject);
+
+        // Bit i set = DRAW BUFFER i of `fbo` resolves to a colour attachment the backend widened
+        // from three channels to four. Indexed by draw-buffer slot, not by attachment point,
+        // because that is what glColorMaski / glClearBufferfv address.
+        Uint32 ComputeAlphaWidenedDrawBufferMask(const MG_State::GLState::FramebufferObject& fbo);
+
+        // The same mask for whatever is currently bound to GL_DRAW_FRAMEBUFFER, recomputed by
+        // SyncCurrentFBO (BackendFramebufferObject::SyncToBackend for the DRAW target, and reset
+        // to 0 on the default framebuffer). Read by the draw/clear state sync, so it is only
+        // trustworthy after SyncCurrentFBO has run in the same entry point.
+        //
+        // WHY IT EXISTS (the dst-alpha discipline). A widened attachment has a real alpha channel
+        // the application's format does not, and GL says a missing channel reads as 1.0. Readback
+        // can paper over that (ForceWideReadAlphaToOne), but GL_DST_ALPHA /
+        // GL_ONE_MINUS_DST_ALPHA blending and glBlitFramebuffer read the STORED alpha inside the
+        // driver where no interception is possible. So the stored alpha is kept at 1.0 instead:
+        // a clear touching a widened buffer writes alpha 1.0, and every draw into it has its
+        // alpha write mask forced off, so nothing can ever move it again. The application's own
+        // colour mask is untouched - glGet(GL_COLOR_WRITEMASK) still reports what it set.
+        extern Uint32 g_alphaWidenedDrawBufferMask;
+
+        // Bit i set = DRAW BUFFER i of the framebuffer bound as DRAW resolves to a colour
+        // attachment with an INTEGER format. Recomputed beside the mask above and for its sake:
+        // glClearBufferfv on an integer colour buffer is GL_INVALID_OPERATION, so the
+        // per-draw-buffer clear route the widening needs has to stand down when one is present.
+        // (glClear on an integer colour buffer is left undefined by ES in the first place, and
+        // an application that wants a defined answer has to call glClearBufferuiv/iv - which does
+        // carry the widened alpha substitution.)
+        extern Uint32 g_integerColorDrawBufferMask;
+
+        // The colour a clear has to hand the driver for one draw buffer: the application's value,
+        // except that a widened attachment's alpha is replaced by the 1.0 its three-channel
+        // format implies. `one` is 1.0 encoded in the clear call's own component type - the
+        // integer clears carry the integer 1, the float clear carries 1.0f.
+        //
+        // Returns `value` itself when nothing is substituted, so the ordinary path allocates and
+        // copies nothing; `scratch` is the caller's buffer and has to outlive the returned
+        // pointer. Free of GL state on purpose, so the substitution can be unit-tested exactly as
+        // the driver sees it.
+        template <typename T>
+        const T* SubstituteWidenedClearAlpha(const T* value, Bool widened, T one, T (&scratch)[4]) {
+            if (!widened || value == nullptr) {
+                return value;
+            }
+            scratch[0] = value[0];
+            scratch[1] = value[1];
+            scratch[2] = value[2];
+            scratch[3] = one;
+            return scratch;
+        }
+
+        // What SyncCurrentFBO last pushed for each target, as a (binding, object, revision)
+        // triple; it re-syncs unless all three still match. Stamped by SyncCurrentFBO and
+        // ForceBindCurrentFBO, cleared by InvalidateFramebufferBindingCache. The three are
+        // only meaningful together - see SyncCurrentFBO.
+        //
+        // The binding slot's own version, which changes whenever a different object is bound
+        // to this target. Distinguishes a rebind from an in-place edit, and keeps the raw
+        // pointer below from matching an address the allocator recycled for a new FBO.
+        extern Array<Uint16, SizeT(FramebufferTarget::FramebufferTargetCount)> g_fboSyncedSlotVersions;
         // Tracks the bound FBO's object version (bumped on any attachment/drawbuffer change)
         // per target: re-attaching textures or changing draw buffers on an already-bound FBO
         // must re-sync it even when the binding-slot version has not moved.
         extern Array<Uint16, SizeT(FramebufferTarget::FramebufferTargetCount)> g_fboSyncedObjectVersions;
+        // Which object was synced. Raw and never dereferenced: only compared for identity.
         extern Array<MG_State::GLState::FramebufferObject*, SizeT(FramebufferTarget::FramebufferTargetCount)>
             g_fboSyncedObjects;
 
@@ -584,6 +903,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
         void InvalidatePackStateCache();
     } // namespace PixelStoreImpl
 
+    namespace SamplerImpl {
+        class BackendSamplerObject; // for PrgramImpl's sampler-pass memo rows below
+    }
+
     // Image uniforms take their unit from the layout(binding=N) qualifier baked into
     // the transpiled ESSL; unlike samplers they must not (and in ES cannot) be
     // assigned through glUniform1i.
@@ -630,6 +953,43 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 Float lastAssignedLodBias = 0.0f;
             };
 
+            // Memo of the whole per-draw sampler-uniform pass (glUniform1i unit
+            // assignments, lod-bias uniform, raw-depth-fetch substitution and the
+            // per-unit sampler-object binds) in BindCurrentProgramWithResources.
+            // The pass is a pure function of the keys below, and its only driver-side
+            // effect is the sampler binding of each sampled unit, so replaying it as
+            // "do nothing" additionally requires those bindings to still be on the
+            // driver - the per-entry row compare against g_boundSamplersCache (the
+            // shadow every sampler bind in this backend already routes through).
+            //
+            // Invalidation enumeration:
+            //  * sampler-uniform unit assignment (glUniform1i) and uniform-block
+            //    binding edits -> frontend backendStateVersion;
+            //  * any texture/sampler bind moving on any unit (incl. the high-water
+            //    mark moving) -> unitBindingsEpoch;
+            //  * any sampler parameter (incl. lod bias, compare mode) or texture
+            //    shape/format change -> samplingGeneration;
+            //  * another frontend context -> contextId (never-reused id);
+            //  * ES context recreation -> textureContextGeneration;
+            //  * relink / backend program rebuild -> SyncToBackend resets `valid`
+            //    (it rebuilds m_samplerUniformBindings, whose lastAssignedUnit /
+            //    lastAssignedLodBias dedup state this memo leans on);
+            //  * any other writer moving a sampled unit's sampler binding
+            //    (BindCurrentUnitSamplers on a unit-sampler change, scratch binds)
+            //    -> the row snapshot compare.
+            struct SamplerPassMemo {
+                static constexpr SizeT kMaxEntries = 16;
+                Bool valid = false;
+                Uint8 count = 0;
+                Uint64 contextId = 0;
+                Uint64 unitBindingsEpoch = 0;
+                Uint64 samplingGeneration = 0;
+                Uint32 backendStateVersion = 0;
+                Uint textureContextGeneration = 0;
+                Array<Uint8, kMaxEntries> units{};
+                Array<SamplerImpl::BackendSamplerObject*, kMaxEntries> rows{};
+            };
+
             BackendProgramObjectImpl();
             ~BackendProgramObjectImpl();
             void SyncToBackend(const SharedPtr<MG_State::GLState::ProgramObject>& stateProgramObject);
@@ -637,6 +997,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
             void SetBaseInstance(Uint32 baseInstance) const;
             void SetBaseInstanceWordIndex(Int32 wordIndex) const;
             void SetDrawID(Uint32 drawId) const;
+            // True when the transpiled program kept a gl_DrawID uniform, i.e. SetDrawID
+            // actually reaches a shader read rather than being discarded.
+            Bool ReadsDrawID() const { return m_drawIdUniformLocation >= 0; }
             Int GetIndirectParamsBinding() const { return m_indirectParamsBinding; }
             Uint GetBackendProgramId() const { return m_backendProgramId; }
             // False when the last SyncToBackend could not produce a usable program (a
@@ -658,6 +1021,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // reflected size when the transpiled block pads differently).
             Int GetGlobalUboBackendBlockSize() const { return m_globalUboBackendBlockSize; }
             BufferImpl::UboRingAllocation& GetGlobalUboRingAllocation() { return m_globalUboRingAllocation; }
+            SamplerPassMemo& GetSamplerPassMemo() { return m_samplerPassMemo; }
             // Frontend link version this backend program (and its resource caches) was
             // built from; a mismatch means every link-derived cache here is stale.
             Uint32 GetSyncedLinkVersion() const { return m_syncedLinkVersion; }
@@ -686,6 +1050,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Uint32 m_lastUploadedGlobalUboVersion = ~0u;
             BufferImpl::UboRingAllocation m_globalUboRingAllocation;
             Uint32 m_syncedLinkVersion = ~0u;
+            SamplerPassMemo m_samplerPassMemo;
         };
 
         extern Uint32 g_snormFallbackClampOutputMask;
@@ -700,6 +1065,22 @@ namespace MobileGL::MG_Backend::DirectGLES {
         extern Uint g_lastUsedBackendProgramId;
         extern StateBackendObjectRegistry<MG_State::GLState::ProgramObject, BackendProgramObjectImpl>
             g_backendProgramObjects;
+
+        // Points one shader storage block of an ALREADY-LINKED backend program at
+        // `binding`. `blockName` is the frontend interface-query spelling; the real
+        // driver's own index for it is looked up here, because the transpiled ESSL's
+        // block order is not the frontend's. Returns false when the block does not exist
+        // on the backend program (eliminated as unused, or the driver lacks the entry
+        // points), which is not an error - GL_BUFFER_BINDING is served from the frontend
+        // record either way.
+        Bool ApplyShaderStorageBlockBinding(Uint backendProgramId, const String& blockName, Uint binding);
+        // Replays every glShaderStorageBlockBinding recorded on the program onto a backend
+        // program that was just built. The frontend record is authoritative (only the
+        // shader's DECLARED binding survives in the SPIR-V), so without this replay any
+        // rebuild would silently revert rebound blocks. Mirrors DirectVulkan's
+        // reseed-on-rebuild in BuildProgramResourceCache.
+        void ReseedShaderStorageBlockBindings(Uint backendProgramId,
+                                              const MG_State::GLState::ProgramObject& stateProgramObject);
     } // namespace PrgramImpl
 
     namespace SamplerImpl {

MobileGL/MG_Backend/DirectGLES/MultiDraw.cpp

@@ -0,0 +1,894 @@
+// MobileGL - MobileGL/MG_Backend/DirectGLES/MultiDraw.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "MultiDraw.h"
+#include "Managers.h"
+#include <MG_State/GLState/Core.h>
+#include <cstring>
+#include <limits>
+
+namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
+    using MG_Config::GLESMultiDrawMode;
+
+    namespace {
+        // ---------------------------------------------------------------------------
+        // Batch shape
+        // ---------------------------------------------------------------------------
+
+        SizeT IndexTypeSize(GLenum type) {
+            switch (type) {
+            case GL_UNSIGNED_BYTE: return 1;
+            case GL_UNSIGNED_SHORT: return 2;
+            case GL_UNSIGNED_INT: return 4;
+            default: return 0;
+            }
+        }
+
+        // The all-ones value of an index type, which is what GL restarts on once
+        // primitive restart is in play. CheckPrimitiveRestartSupported has already
+        // rejected the arbitrary-index form of GL_PRIMITIVE_RESTART, so an enabled
+        // restart always restarts here and nowhere else.
+        Uint32 RestartSentinelFor(GLenum type) {
+            switch (type) {
+            case GL_UNSIGNED_BYTE: return 0xFFu;
+            case GL_UNSIGNED_SHORT: return 0xFFFFu;
+            default: return 0xFFFFFFFFu;
+            }
+        }
+
+        Bool RestartActive() {
+            return MG_State::pGLContext->IsCapabilityEnabled(CapabilityInput::PrimitiveRestart) ||
+                   MG_State::pGLContext->IsCapabilityEnabled(CapabilityInput::PrimitiveRestartFixedIndex);
+        }
+
+        // Vertices per primitive for the modes whose sub-draws may be concatenated into a
+        // single draw without changing the primitive stream. Zero for strip/loop/fan modes
+        // (concatenation would weld one sub-draw's last primitive to the next sub-draw's
+        // first) and for GL_PATCHES, whose primitive size is dynamic tessellation state.
+        Uint32 ConcatenablePrimitiveSize(GLenum mode) {
+            switch (mode) {
+            case GL_POINTS: return 1;
+            case GL_LINES: return 2;
+            case GL_TRIANGLES: return 3;
+            case GL_LINES_ADJACENCY: return 4;
+            case GL_TRIANGLES_ADJACENCY: return 6;
+            default: return 0;
+            }
+        }
+
+        // Beyond this an emulated batch would ask for a scratch allocation measured in
+        // hundreds of megabytes (and the scratch ring never shrinks again); decline and let
+        // a per-sub-draw tier handle it instead of trying and failing inside the driver.
+        constexpr SizeT kMaxFlattenedIndices = SizeT{1} << 24;
+
+        // The flattening dispatch is one invocation per output index. ES 3.1 only
+        // guarantees 65535 work groups per dimension, and exceeding it makes
+        // glDispatchCompute an INVALID_VALUE no-op - which would leave the draw reading an
+        // uninitialised index buffer rather than failing visibly. Cap the tier there
+        // instead of querying: 4.19M indices is far past any real multi-draw batch, and
+        // beyond it the per-sub-draw tiers are the better answer anyway.
+        constexpr SizeT kComputeWorkGroupSize = 64;
+        constexpr SizeT kMaxComputeWorkGroups = 65535;
+        constexpr SizeT kMaxComputeFlattenedIndices = kMaxComputeWorkGroups * kComputeWorkGroupSize;
+
+        Uint BoundDrawIndirectBufferId() {
+            const auto& indirect =
+                MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::DrawIndirect).GetBoundObject();
+            if (!indirect) return 0;
+            const auto* resource = BufferImpl::EnsureBufferResource(indirect);
+            return resource ? resource->id : 0;
+        }
+
+        const SharedPtr<MG_State::GLState::BufferObject>& BoundIndexBuffer() {
+            static const SharedPtr<MG_State::GLState::BufferObject> none;
+            const auto& vao = MG_State::pGLContext->GetBoundVertexArray();
+            if (!vao) return none;
+            return vao->GetIndexBufferBindingSlot().GetBoundObject();
+        }
+
+        // The GL name PrepareForDraw left on GL_ELEMENT_ARRAY_BUFFER, i.e. what a tier
+        // that swaps in a scratch index buffer has to put back. Restoring the exact name
+        // matters beyond tidiness: the VAO twin memoises that it already synced this
+        // index binding and will not re-issue it on the next draw.
+        Uint BoundIndexBufferId() {
+            const auto& ibo = BoundIndexBuffer();
+            if (!ibo) return 0;
+            const auto* resource = BufferImpl::EnsureBufferResource(ibo);
+            return resource ? resource->id : 0;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Scratch GL objects
+        //
+        // All of them belong to the ES context and are abandoned (not deleted) when it
+        // dies, exactly like XfbImpl's scatter buffer: the names are the dead context's
+        // to reclaim, and deleting them would target whatever the successor context
+        // handed out for the same name.
+        // ---------------------------------------------------------------------------
+
+        struct ScratchBuffer {
+            Uint id = 0;
+            SizeT capacity = 0;
+            SizeT cursor = 0; // ring buffers only: next free byte
+        };
+
+        ScratchBuffer g_indirectCommands; // synthesized DrawElementsIndirectCommand array
+        ScratchBuffer g_rebasedIndices;   // CPU-rebased index stream
+        ScratchBuffer g_drawInfo;         // compute tier: per-sub-draw descriptors
+        ScratchBuffer g_flattenedIndices; // compute tier: flattened index stream
+
+        Uint g_computeProgram = 0;
+        Bool g_computeProgramFailed = false;
+        GLint g_uElementSize = -1;
+        GLint g_uDrawCount = -1;
+        GLint g_uTotalIndices = -1;
+
+        // Reused staging, so a steady stream of batches allocates nothing.
+        Vector<DrawElementsIndirectCommand> g_commandStaging;
+        Vector<Uint32> g_indexStaging;
+        Vector<Uint32> g_drawInfoStaging;
+        Vector<GLint> g_zeroBaseVertices;
+
+        // Everything below stages through GL_ARRAY_BUFFER, the manager-wide staging target
+        // (BufferImpl::TempBufferTarget); binding it disturbs no VAO state.
+        Bool EnsureScratchName(ScratchBuffer& buffer) {
+            if (buffer.id != 0) return true;
+            GLuint id = 0;
+            g_GLESFuncs.glGenBuffers(1, &id);
+            if (id == 0) return false;
+            buffer.id = id;
+            buffer.capacity = 0;
+            buffer.cursor = 0;
+            return true;
+        }
+
+        // Whole-buffer upload, for the two buffers that are read from offset 0 because they
+        // are bound as storage blocks. Respecifies rather than sub-updates: glBufferData
+        // orphans the previous store, so the upload never waits on a dispatch still reading
+        // the old contents out of the same name.
+        Bool UploadScratch(ScratchBuffer& buffer, SizeT bytes, const void* data) {
+            if (bytes == 0) return true;
+            if (!EnsureScratchName(buffer)) return false;
+            BufferImpl::BindBufferId(BufferImpl::TempBufferTarget, buffer.id);
+            // Grow in powers of two so a batch that creeps up in size stops respecifying.
+            SizeT capacity = buffer.capacity == 0 ? bytes : buffer.capacity;
+            while (capacity < bytes) capacity *= 2;
+            g_GLESFuncs.glBufferData(BufferImpl::TempBufferTarget, static_cast<GLsizeiptr>(capacity), nullptr,
+                                     GL_STREAM_DRAW);
+            buffer.capacity = capacity;
+            buffer.cursor = 0;
+            if (data) {
+                g_GLESFuncs.glBufferSubData(BufferImpl::TempBufferTarget, 0, static_cast<GLsizeiptr>(bytes), data);
+            }
+            return true;
+        }
+
+        // Ring upload, for the buffers whose consumers can address a byte offset (indirect
+        // commands and rewritten index streams). Respecifying per batch is what an
+        // orphan-every-time scheme costs, and on a desktop-class driver that allocation
+        // dominated the tiers that use these buffers - a multi-draw of 32 sub-draws stages
+        // 640 bytes and paid for a fresh store to hold them. Bump-allocating instead means
+        // one respecify per wrap; every byte between two wraps is written exactly once, so
+        // nothing in flight is overwritten, and the wrap itself orphans.
+        constexpr SizeT kRingAlignment = 16; // >= 4, so both command and uint32-index offsets stay legal
+        constexpr SizeT kMinRingBytes = 1u << 16;
+
+        Bool UploadScratchRing(ScratchBuffer& buffer, SizeT bytes, const void* data, SizeT& outOffset) {
+            outOffset = 0;
+            if (bytes == 0) return true;
+            if (!EnsureScratchName(buffer)) return false;
+            BufferImpl::BindBufferId(BufferImpl::TempBufferTarget, buffer.id);
+
+            const SizeT aligned = (bytes + kRingAlignment - 1) & ~(kRingAlignment - 1);
+            if (buffer.capacity < aligned) {
+                SizeT capacity = buffer.capacity == 0 ? kMinRingBytes : buffer.capacity;
+                while (capacity < aligned) capacity *= 2;
+                g_GLESFuncs.glBufferData(BufferImpl::TempBufferTarget, static_cast<GLsizeiptr>(capacity), nullptr,
+                                         GL_STREAM_DRAW);
+                buffer.capacity = capacity;
+                buffer.cursor = 0;
+            } else if (buffer.cursor + aligned > buffer.capacity) {
+                g_GLESFuncs.glBufferData(BufferImpl::TempBufferTarget, static_cast<GLsizeiptr>(buffer.capacity),
+                                         nullptr, GL_STREAM_DRAW);
+                buffer.cursor = 0;
+            }
+
+            outOffset = buffer.cursor;
+            if (data) {
+                g_GLESFuncs.glBufferSubData(BufferImpl::TempBufferTarget, static_cast<GLintptr>(outOffset),
+                                            static_cast<GLsizeiptr>(bytes), data);
+            }
+            buffer.cursor += aligned;
+            return true;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Tier resolution
+        // ---------------------------------------------------------------------------
+
+        // Best-first, and measured rather than assumed. MobileGlues orders its own Auto
+        // multiindirect -> indirect -> basevertex; on both ES drivers available here that
+        // is backwards, because staging a command buffer per batch costs more than the
+        // driver entries it saves. mc_sodium_multidraw (132 batches x 32 sub-draws),
+        // ns/op, median of three:
+        //
+        //                    NVIDIA ES 3.2   Mesa llvmpipe ES 3.2
+        //   ext                  n/a               19300
+        //   basevertex           2500              25200
+        //   multiindirect        5700              27600
+        //   drawelements         5600              28700
+        //   indirect             5800              31000
+        //
+        // Ring-allocating the command staging (instead of respecifying per batch) was
+        // tried first and moved the indirect tiers by less than noise, so the cost is the
+        // indirect draw path itself, not the upload. Only "ext" - a real multi-draw entry
+        // point rather than an indirect one - actually beats replaying the sub-draws.
+        //
+        // The compute tier is deliberately absent from the ladder: it rewrites the
+        // primitive stream rather than replaying it, and it measured slowest of all here,
+        // so it stays opt-in behind the env knob (the same call MobileGlues makes - its
+        // Auto never selects Compute either).
+        constexpr GLESMultiDrawMode kAutoLadder[] = {
+            GLESMultiDrawMode::Ext,      GLESMultiDrawMode::BaseVertex,   GLESMultiDrawMode::MultiIndirect,
+            GLESMultiDrawMode::Indirect, GLESMultiDrawMode::DrawElements,
+        };
+
+        Bool SupportsTier(GLESMultiDrawMode tier) {
+            return IsTierSupported(g_GLESCapabilities, g_GLESFuncs, tier);
+        }
+
+        GLESMultiDrawMode g_resolvedTier = GLESMultiDrawMode::Auto;
+        Bool g_tierResolved = false;
+        String g_tierResolution;
+
+        void ResolveTierOnce() {
+            if (g_tierResolved) return;
+            g_tierResolved = true;
+            g_resolvedTier =
+                ResolveTier(g_GLESCapabilities, g_GLESFuncs, MG_Config::Features.EsprytMultiDrawMode,
+                            &g_tierResolution);
+            MGLOG_I("DirectGLES multi-draw: %s", g_tierResolution.c_str());
+        }
+
+        // Which tiers have already announced themselves, one bit per GLESMultiDrawMode.
+        // The resolution line above says which tier was CHOSEN; this says which one a
+        // batch actually went through, and the two differ whenever a batch's shape
+        // demotes it. Worth a line each: a multi-draw path that resolves to a tier and
+        // then quietly runs a different one is exactly how "the batch drew nothing"
+        // hides.
+        Uint32 g_announcedTiers = 0;
+
+        void NoteTierExecuted(GLESMultiDrawMode tier) {
+            const Uint32 bit = 1u << static_cast<Uint32>(tier);
+            if (g_announcedTiers & bit) return;
+            g_announcedTiers |= bit;
+            MGLOG_I("DirectGLES multi-draw: first batch executed via tier \"%s\"", TierName(tier));
+        }
+
+        // The tier this particular batch can actually take. A tier is demoted here when
+        // the batch's own shape - not the driver - rules it out; the compute tier keeps
+        // its remaining feasibility checks inside its implementation, where the data it
+        // has to walk is already in hand.
+        GLESMultiDrawMode ResolveTierForBatch(Bool programReadsDrawID, Bool hasIndexBuffer) {
+            ResolveTierOnce();
+            GLESMultiDrawMode tier = g_resolvedTier;
+
+            // Batched tiers issue one driver entry for the whole batch, so the emulated
+            // gl_DrawID uniform can only hold one value across every sub-draw. A program
+            // that reads gl_DrawID gets an unrolled tier, which feeds each sub-draw its
+            // own index (the spec's value); nothing else observes the difference.
+            const Bool batched = tier == GLESMultiDrawMode::Ext || tier == GLESMultiDrawMode::MultiIndirect ||
+                                 tier == GLESMultiDrawMode::Compute;
+            if (batched && programReadsDrawID) {
+                tier = SupportsTier(GLESMultiDrawMode::BaseVertex) ? GLESMultiDrawMode::BaseVertex
+                                                                   : GLESMultiDrawMode::DrawElements;
+            }
+
+            // The indirect tiers describe each sub-draw as an element offset into the
+            // bound element array buffer. A client-memory index array has no such buffer,
+            // and indirect draws are not defined without one.
+            if (!hasIndexBuffer &&
+                (tier == GLESMultiDrawMode::MultiIndirect || tier == GLESMultiDrawMode::Indirect)) {
+                tier = SupportsTier(GLESMultiDrawMode::BaseVertex) ? GLESMultiDrawMode::BaseVertex
+                                                                   : GLESMultiDrawMode::DrawElements;
+            }
+            return tier;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Index rewriting, shared by the two tiers that fold base vertices into indices
+        // ---------------------------------------------------------------------------
+
+        // Both of those tiers emit GL_UNSIGNED_INT regardless of the source type. Keeping
+        // the source width would be wrong, not merely tight: GL adds baseVertex to the
+        // index at full precision, so a GL_UNSIGNED_SHORT index plus a base vertex past
+        // 65535 addresses a vertex the source type cannot spell. Widening also gives the
+        // rewritten stream a restart sentinel (0xFFFFFFFF) that survives the rebase.
+        void RebaseIndices(const Uint8* source, SizeT sourceIndexCount, SizeT indexSize, Int32 baseVertex,
+                           Bool restartActive, Uint32 restartSentinel, Uint32* out) {
+            const Uint32 baseVertexBits = static_cast<Uint32>(baseVertex);
+            for (SizeT i = 0; i < sourceIndexCount; ++i) {
+                Uint32 value = 0;
+                switch (indexSize) {
+                case 1: value = source[i]; break;
+                case 2: {
+                    Uint16 narrow = 0;
+                    std::memcpy(&narrow, source + i * 2, sizeof(narrow));
+                    value = narrow;
+                    break;
+                }
+                default: std::memcpy(&value, source + i * 4, sizeof(value)); break;
+                }
+                // Unsigned wraparound is the defined behaviour for a negative base vertex.
+                out[i] = (restartActive && value == restartSentinel) ? 0xFFFFFFFFu : value + baseVertexBits;
+            }
+        }
+
+        // CPU-readable bytes of one sub-draw's indices, from the frontend shadow of the
+        // bound index buffer or straight from the client array. Null when the sub-draw
+        // would read outside the buffer.
+        const Uint8* ResolveSubDrawIndices(const SharedPtr<MG_State::GLState::BufferObject>& indexBuffer,
+                                           const Uint8* indexBufferBytes, SizeT indexBufferSize, const void* indices,
+                                           SizeT indexCount, SizeT indexSize) {
+            if (!indexBuffer) {
+                return static_cast<const Uint8*>(indices);
+            }
+            if (!indexBufferBytes) return nullptr;
+            const SizeT byteOffset = reinterpret_cast<SizeT>(indices);
+            const SizeT byteEnd = byteOffset + indexCount * indexSize;
+            if (byteEnd > indexBufferSize || byteEnd < byteOffset) return nullptr;
+            return indexBufferBytes + byteOffset;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Tier: Ext - one glMultiDrawElementsBaseVertexEXT
+        // ---------------------------------------------------------------------------
+
+        Bool RunExt(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices, GLsizei drawcount,
+                    const GLint* basevertex) {
+            if (!SupportsTier(GLESMultiDrawMode::Ext)) return false;
+            const GLint* baseVertices = basevertex;
+            if (!baseVertices) {
+                // glMultiDrawElements: every base vertex is 0, but the entry point still
+                // wants an array. One permanently-zero vector serves every such batch.
+                if (g_zeroBaseVertices.size() < static_cast<SizeT>(drawcount)) {
+                    g_zeroBaseVertices.resize(static_cast<SizeT>(drawcount), 0);
+                }
+                baseVertices = g_zeroBaseVertices.data();
+            }
+            g_GLESFuncs.glMultiDrawElementsBaseVertexEXT(mode, count, type, indices, drawcount, baseVertices);
+            NoteTierExecuted(GLESMultiDrawMode::Ext);
+            return true;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Tiers: MultiIndirect / Indirect - synthesized indirect commands
+        // ---------------------------------------------------------------------------
+
+        Bool RunIndirect(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                         GLsizei drawcount, const GLint* basevertex, Bool batched, Bool feedDrawID) {
+            if (!SupportsTier(batched ? GLESMultiDrawMode::MultiIndirect : GLESMultiDrawMode::Indirect)) return false;
+            const SizeT indexSize = IndexTypeSize(type);
+            if (indexSize == 0) return false;
+            // Indirect commands address indices as an element offset into the bound element
+            // array buffer, and an indirect draw is not defined without one.
+            const auto& indexBuffer = BoundIndexBuffer();
+            if (!indexBuffer) return false;
+
+            g_commandStaging.resize(static_cast<SizeT>(drawcount));
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                const SizeT byteOffset = reinterpret_cast<SizeT>(indices[i]);
+                // firstIndex counts elements, so an offset that is not a whole number of
+                // them cannot be expressed as a command at all.
+                if (byteOffset % indexSize != 0) return false;
+                auto& command = g_commandStaging[static_cast<SizeT>(i)];
+                command.count = count[i] > 0 ? static_cast<Uint32>(count[i]) : 0u;
+                command.instanceCount = 1;
+                command.firstIndex = static_cast<Uint32>(byteOffset / indexSize);
+                command.baseVertex = basevertex ? basevertex[i] : 0;
+                command.baseInstance = 0;
+            }
+
+            const SizeT commandBytes = g_commandStaging.size() * sizeof(DrawElementsIndirectCommand);
+            SizeT commandBase = 0;
+            if (!UploadScratchRing(g_indirectCommands, commandBytes, g_commandStaging.data(), commandBase)) {
+                return false;
+            }
+
+            // Every synthesized command carries baseInstance 0. Say so through the direct
+            // path, which also clears the indirect-params word index a preceding real
+            // indirect draw may have left pointing into its own command buffer.
+            SetCurrentBaseInstance(0);
+
+            const Uint previousIndirectBinding = BoundDrawIndirectBufferId();
+            BufferImpl::BindBufferId(GL_DRAW_INDIRECT_BUFFER, g_indirectCommands.id);
+            if (batched) {
+                g_GLESFuncs.glMultiDrawElementsIndirectEXT(mode, type, reinterpret_cast<const void*>(commandBase),
+                                                           drawcount, 0);
+            } else {
+                for (GLsizei i = 0; i < drawcount; ++i) {
+                    if (feedDrawID) SetCurrentDrawID(static_cast<Uint32>(i));
+                    const SizeT commandOffset = commandBase + static_cast<SizeT>(i) * sizeof(DrawElementsIndirectCommand);
+                    g_GLESFuncs.glDrawElementsIndirect(mode, type, reinterpret_cast<const void*>(commandOffset));
+                }
+                if (feedDrawID) SetCurrentDrawID(0);
+            }
+            BufferImpl::BindBufferId(GL_DRAW_INDIRECT_BUFFER, previousIndirectBinding);
+            NoteTierExecuted(batched ? GLESMultiDrawMode::MultiIndirect : GLESMultiDrawMode::Indirect);
+            return true;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Tier: BaseVertex - the per-sub-draw replay
+        // ---------------------------------------------------------------------------
+
+        Bool RunBaseVertexLoop(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                               GLsizei drawcount, const GLint* basevertex, Bool feedDrawID) {
+            if (!SupportsTier(GLESMultiDrawMode::BaseVertex)) return false;
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] <= 0) continue;
+                if (feedDrawID) SetCurrentDrawID(static_cast<Uint32>(i));
+                g_GLESFuncs.glDrawElementsBaseVertex(mode, count[i], type, indices[i],
+                                                     basevertex ? basevertex[i] : 0);
+            }
+            if (feedDrawID) SetCurrentDrawID(0);
+            NoteTierExecuted(GLESMultiDrawMode::BaseVertex);
+            return true;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Tier: DrawElements - base vertices folded into a scratch index stream
+        // ---------------------------------------------------------------------------
+
+        Bool RunRebasedDrawElements(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                                    GLsizei drawcount, const GLint* basevertex, Bool feedDrawID) {
+            const SizeT indexSize = IndexTypeSize(type);
+            if (indexSize == 0) return false;
+
+            SizeT total = 0;
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] > 0) total += static_cast<SizeT>(count[i]);
+            }
+            if (total == 0) return true;
+            if (total > kMaxFlattenedIndices) return false;
+
+            const auto& indexBuffer = BoundIndexBuffer();
+            const Uint8* indexBufferBytes = nullptr;
+            SizeT indexBufferSize = 0;
+            if (indexBuffer) {
+                // The shadow is the source of truth for CPU reads, but a persistent map or
+                // a shader write may have moved past it since the last sync.
+                indexBuffer->SyncPersistentMappedRange();
+                indexBuffer->SyncGpuWrites();
+                indexBufferBytes = indexBuffer->MappedData();
+                indexBufferSize = indexBuffer->GetSize();
+            }
+
+            const Bool restartActive = RestartActive();
+            const Uint32 restartSentinel = RestartSentinelFor(type);
+            g_indexStaging.resize(total);
+            SizeT cursor = 0;
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] <= 0) continue;
+                const SizeT subDrawCount = static_cast<SizeT>(count[i]);
+                const Uint8* source = ResolveSubDrawIndices(indexBuffer, indexBufferBytes, indexBufferSize, indices[i],
+                                                            subDrawCount, indexSize);
+                if (!source) {
+                    MGLOG_E("DirectGLES multi-draw (drawelements tier): sub-draw %d reads outside the bound index "
+                            "buffer; skipping the batch",
+                            i);
+                    return false;
+                }
+                RebaseIndices(source, subDrawCount, indexSize, basevertex ? basevertex[i] : 0, restartActive,
+                              restartSentinel, g_indexStaging.data() + cursor);
+                cursor += subDrawCount;
+            }
+
+            SizeT indexBase = 0;
+            if (!UploadScratchRing(g_rebasedIndices, total * sizeof(Uint32), g_indexStaging.data(), indexBase)) {
+                return false;
+            }
+
+            const Uint previousIndexBinding = BoundIndexBufferId();
+            BufferImpl::BindBufferId(GL_ELEMENT_ARRAY_BUFFER, g_rebasedIndices.id);
+            cursor = 0;
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] <= 0) continue;
+                if (feedDrawID) SetCurrentDrawID(static_cast<Uint32>(i));
+                g_GLESFuncs.glDrawElements(mode, count[i], GL_UNSIGNED_INT,
+                                           reinterpret_cast<const void*>(indexBase + cursor * sizeof(Uint32)));
+                cursor += static_cast<SizeT>(count[i]);
+            }
+            if (feedDrawID) SetCurrentDrawID(0);
+            BufferImpl::BindBufferId(GL_ELEMENT_ARRAY_BUFFER, previousIndexBinding);
+            NoteTierExecuted(GLESMultiDrawMode::DrawElements);
+            return true;
+        }
+
+        // ---------------------------------------------------------------------------
+        // Tier: Compute - the whole batch flattened into one rebased index stream
+        // ---------------------------------------------------------------------------
+
+        // One index per invocation. The sub-draw an output slot belongs to is found by
+        // binary search over the inclusive prefix sums of the sub-draw counts, which is
+        // why the descriptors are sorted by construction. Sub-draws with a zero count
+        // repeat the previous prefix sum and are therefore skipped by the search.
+        //
+        // Three storage blocks, not the five the shape suggests: ES 3.1 only guarantees
+        // four per compute stage, so the per-sub-draw descriptors share one buffer.
+        constexpr const char* kFlattenComputeSource = R"(#version 310 es
+layout(local_size_x = 64) in;
+
+uniform uint uElementSize;
+uniform uint uDrawCount;
+uniform uint uTotalIndices;
+
+layout(std430, binding = 0) readonly buffer SourceIndices { uint sourceWords[]; };
+layout(std430, binding = 1) readonly buffer DrawInfo { uint drawInfo[]; };
+layout(std430, binding = 2) writeonly buffer FlatIndices { uint flatIndices[]; };
+
+uint ReadSourceIndex(uint element) {
+    if (uElementSize == 4u) {
+        return sourceWords[element];
+    }
+    if (uElementSize == 2u) {
+        uint word = sourceWords[element >> 1u];
+        return (word >> ((element & 1u) * 16u)) & 0xFFFFu;
+    }
+    uint word = sourceWords[element >> 2u];
+    return (word >> ((element & 3u) * 8u)) & 0xFFu;
+}
+
+void main() {
+    uint outIndex = gl_GlobalInvocationID.x;
+    if (outIndex >= uTotalIndices) {
+        return;
+    }
+
+    uint low = 0u;
+    uint high = uDrawCount - 1u;
+    while (low < high) {
+        uint mid = low + (high - low) / 2u;
+        if (drawInfo[mid * 3u + 2u] > outIndex) {
+            high = mid;
+        } else {
+            low = mid + 1u;
+        }
+    }
+
+    uint localIndex = outIndex - (low == 0u ? 0u : drawInfo[(low - 1u) * 3u + 2u]);
+    // Unsigned wraparound is the defined behaviour for a negative base vertex. No
+    // restart sentinel handling: the tier declines outright while restart is enabled.
+    flatIndices[outIndex] = ReadSourceIndex(localIndex + drawInfo[low * 3u]) + drawInfo[low * 3u + 1u];
+}
+)";
+
+        struct FlattenedStream {
+            Uint bufferId = 0;
+            SizeT indexCount = 0;
+        };
+
+        Bool EnsureComputeProgram() {
+            if (g_computeProgram != 0) return true;
+            if (g_computeProgramFailed) return false;
+            g_computeProgramFailed = true; // cleared again only on a complete success
+
+            const GLuint shader = g_GLESFuncs.glCreateShader(GL_COMPUTE_SHADER);
+            if (shader == 0) {
+                MGLOG_E("DirectGLES multi-draw (compute tier): glCreateShader(GL_COMPUTE_SHADER) failed");
+                return false;
+            }
+            const char* source = kFlattenComputeSource;
+            g_GLESFuncs.glShaderSource(shader, 1, &source, nullptr);
+            g_GLESFuncs.glCompileShader(shader);
+            GLint status = GL_FALSE;
+            g_GLESFuncs.glGetShaderiv(shader, GL_COMPILE_STATUS, &status);
+            if (status != GL_TRUE) {
+                char log[1024] = {};
+                g_GLESFuncs.glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                MGLOG_E("DirectGLES multi-draw (compute tier): index-flattening shader failed to compile: %s", log);
+                g_GLESFuncs.glDeleteShader(shader);
+                return false;
+            }
+
+            const GLuint program = g_GLESFuncs.glCreateProgram();
+            if (program == 0) {
+                MGLOG_E("DirectGLES multi-draw (compute tier): glCreateProgram failed");
+                g_GLESFuncs.glDeleteShader(shader);
+                return false;
+            }
+            g_GLESFuncs.glAttachShader(program, shader);
+            g_GLESFuncs.glLinkProgram(program);
+            g_GLESFuncs.glDeleteShader(shader);
+            g_GLESFuncs.glGetProgramiv(program, GL_LINK_STATUS, &status);
+            if (status != GL_TRUE) {
+                char log[1024] = {};
+                g_GLESFuncs.glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                MGLOG_E("DirectGLES multi-draw (compute tier): index-flattening program failed to link: %s", log);
+                g_GLESFuncs.glDeleteProgram(program);
+                return false;
+            }
+
+            g_computeProgram = program;
+            g_uElementSize = g_GLESFuncs.glGetUniformLocation(program, "uElementSize");
+            g_uDrawCount = g_GLESFuncs.glGetUniformLocation(program, "uDrawCount");
+            g_uTotalIndices = g_GLESFuncs.glGetUniformLocation(program, "uTotalIndices");
+            g_computeProgramFailed = false;
+            MGLOG_I("DirectGLES multi-draw: index-flattening compute program ready (id %u)", program);
+            return true;
+        }
+
+        // Builds the flattened stream, or leaves `out` empty when this batch's shape rules
+        // the tier out. Runs BEFORE PrepareForDraw - see the call site - so it may leave
+        // the compute program current and the first storage points unbound; the
+        // preparation that follows re-establishes both.
+        void FlattenWithCompute(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                                GLsizei drawcount, const GLint* basevertex, FlattenedStream& out) {
+            if (!SupportsTier(GLESMultiDrawMode::Compute)) return;
+            const SizeT indexSize = IndexTypeSize(type);
+            if (indexSize == 0) return;
+
+            // Merging sub-draws into a single draw only reproduces the original primitive
+            // stream for list-shaped modes: a strip, loop or fan would gain primitives
+            // spanning the seam between two sub-draws.
+            const Uint32 primitiveSize = ConcatenablePrimitiveSize(mode);
+            if (primitiveSize == 0) return;
+
+            // Primitive restart defeats the whole-multiple-of-a-primitive argument below,
+            // even for a list mode. A restart ends the current primitive, so a sub-draw of
+            // six GL_TRIANGLES indices with a restart after the third emits ONE triangle
+            // and drops the two leftover vertices - and once concatenated those leftovers
+            // find a third vertex in the next sub-draw and become a triangle that GL never
+            // draws. Splicing separator sentinels into the flattened stream could fix it,
+            // at the cost of a per-sub-draw offset the prefix-sum layout does not carry;
+            // declining is the honest trade for a tier that is already opt-in.
+            if (RestartActive()) return;
+
+            // The shader reads the source indices as a storage buffer, so there has to be
+            // a real buffer to read - a client-memory index array has none.
+            const auto& indexBuffer = BoundIndexBuffer();
+            if (!indexBuffer) return;
+
+            // A dispatch inside an open capture span is not legal, and the span would also
+            // observe one merged draw rather than the batch it asked for.
+            if (XfbImpl::IsCaptureSpanOpen()) return;
+
+            auto* sourceResource = BufferImpl::EnsureBufferResource(indexBuffer);
+            if (!sourceResource || sourceResource->id == 0) return;
+            const SizeT sourceSize = indexBuffer->GetSize();
+            // std430 addresses the source as uint[]; a tail shorter than a word is not
+            // reachable, so a narrow index type needs a word-multiple buffer.
+            if (indexSize < 4 && (sourceSize % 4) != 0) return;
+
+            g_drawInfoStaging.resize(3 * static_cast<SizeT>(drawcount));
+            SizeT total = 0;
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                const SizeT subDrawCount = count[i] > 0 ? static_cast<SizeT>(count[i]) : 0;
+                // GL drops a trailing partial primitive per sub-draw; concatenation would
+                // instead splice it onto the next sub-draw's first vertices.
+                if (subDrawCount % primitiveSize != 0) return;
+                const SizeT byteOffset = reinterpret_cast<SizeT>(indices[i]);
+                if (byteOffset % indexSize != 0) return;
+                if (subDrawCount != 0) {
+                    const SizeT byteEnd = byteOffset + subDrawCount * indexSize;
+                    if (byteEnd > sourceSize || byteEnd < byteOffset) return;
+                }
+                total += subDrawCount;
+                if (total > kMaxComputeFlattenedIndices) return;
+                const SizeT slot = 3 * static_cast<SizeT>(i);
+                g_drawInfoStaging[slot] = static_cast<Uint32>(byteOffset / indexSize);
+                g_drawInfoStaging[slot + 1] = static_cast<Uint32>(basevertex ? basevertex[i] : 0);
+                g_drawInfoStaging[slot + 2] = static_cast<Uint32>(total);
+            }
+            if (total == 0) return; // nothing to draw; the ordinary tiers no-op just as well
+
+            if (!EnsureComputeProgram()) return;
+            if (!UploadScratch(g_drawInfo, g_drawInfoStaging.size() * sizeof(Uint32), g_drawInfoStaging.data())) {
+                return;
+            }
+            if (!UploadScratch(g_flattenedIndices, total * sizeof(Uint32), nullptr)) return;
+
+            BufferImpl::BindBufferBaseCached(GL_SHADER_STORAGE_BUFFER, 0, sourceResource->id);
+            BufferImpl::BindBufferBaseCached(GL_SHADER_STORAGE_BUFFER, 1, g_drawInfo.id);
+            BufferImpl::BindBufferBaseCached(GL_SHADER_STORAGE_BUFFER, 2, g_flattenedIndices.id);
+
+            g_GLESFuncs.glUseProgram(g_computeProgram);
+            PrgramImpl::g_lastUsedBackendProgramId = g_computeProgram;
+            if (g_uElementSize >= 0) g_GLESFuncs.glUniform1ui(g_uElementSize, static_cast<GLuint>(indexSize));
+            if (g_uDrawCount >= 0) g_GLESFuncs.glUniform1ui(g_uDrawCount, static_cast<GLuint>(drawcount));
+            if (g_uTotalIndices >= 0) g_GLESFuncs.glUniform1ui(g_uTotalIndices, static_cast<GLuint>(total));
+
+            g_GLESFuncs.glDispatchCompute(
+                static_cast<GLuint>((total + kComputeWorkGroupSize - 1) / kComputeWorkGroupSize), 1, 1);
+            g_GLESFuncs.glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT | GL_ELEMENT_ARRAY_BARRIER_BIT);
+
+            // Hand the storage points back to their GL default. PrepareForDraw re-syncs
+            // only the points the app has actually touched, so leaving a scratch buffer on
+            // an untouched point would keep it visible to the next shader that declares one.
+            for (Uint point = 0; point < 3; ++point) {
+                BufferImpl::BindBufferBaseCached(GL_SHADER_STORAGE_BUFFER, point, 0);
+            }
+
+            NoteTierExecuted(GLESMultiDrawMode::Compute);
+            out.bufferId = g_flattenedIndices.id;
+            out.indexCount = total;
+        }
+    } // namespace
+
+    // -------------------------------------------------------------------------------
+    // Public surface
+    // -------------------------------------------------------------------------------
+
+    Bool IsTierSupported(const MG_External::GLESCapabilities& caps, const MG_External::GLESFunctionsTable& funcs,
+                         GLESMultiDrawMode tier) {
+        const Bool esAtLeast31 =
+            caps.GLESVersion.Major > 3 || (caps.GLESVersion.Major == 3 && caps.GLESVersion.Minor >= 1);
+        switch (tier) {
+        case GLESMultiDrawMode::Ext:
+            return caps.SupportsMultiDrawElementsBaseVertex;
+        case GLESMultiDrawMode::MultiIndirect:
+            return caps.SupportsMultiDrawIndirect && esAtLeast31 && funcs.glDrawElementsIndirect != nullptr;
+        case GLESMultiDrawMode::Indirect:
+            return esAtLeast31 && funcs.glDrawElementsIndirect != nullptr;
+        case GLESMultiDrawMode::BaseVertex:
+            return caps.SupportsDrawElementsBaseVertex;
+        case GLESMultiDrawMode::DrawElements:
+            // Plain glDrawElements over a rewritten index stream: ES 2 core, so this is
+            // the floor every other tier can fall back to.
+            return true;
+        case GLESMultiDrawMode::Compute:
+            // Three storage blocks, which is inside the four ES 3.1 guarantees per stage.
+            return caps.SupportsComputeShader && caps.MaxComputeShaderStorageBlocks >= 3 &&
+                   funcs.glBindBufferBase != nullptr;
+        case GLESMultiDrawMode::Auto:
+            break;
+        }
+        return false;
+    }
+
+    GLESMultiDrawMode ResolveTier(const MG_External::GLESCapabilities& caps,
+                                  const MG_External::GLESFunctionsTable& funcs, GLESMultiDrawMode requested,
+                                  String* explanation) {
+        const auto bestAuto = [&]() {
+            for (const GLESMultiDrawMode tier : kAutoLadder) {
+                if (IsTierSupported(caps, funcs, tier)) return tier;
+            }
+            return GLESMultiDrawMode::DrawElements;
+        };
+
+        GLESMultiDrawMode resolved = GLESMultiDrawMode::DrawElements;
+        String line;
+        if (requested == GLESMultiDrawMode::Auto) {
+            resolved = bestAuto();
+            line = String("auto -> ") + TierName(resolved);
+        } else if (IsTierSupported(caps, funcs, requested)) {
+            resolved = requested;
+            line = String("MOBILEGL_ESPRYT_MULTIDRAW_MODE=") + TierName(requested) + " -> " + TierName(resolved);
+        } else {
+            resolved = bestAuto();
+            line = String("MOBILEGL_ESPRYT_MULTIDRAW_MODE=") + TierName(requested) +
+                   " requested but unsupported by this driver -> " + TierName(resolved);
+        }
+
+        if (explanation) {
+            String supported;
+            for (const GLESMultiDrawMode tier : kAutoLadder) {
+                if (!IsTierSupported(caps, funcs, tier)) continue;
+                if (!supported.empty()) supported += ", ";
+                supported += TierName(tier);
+            }
+            if (IsTierSupported(caps, funcs, GLESMultiDrawMode::Compute)) {
+                supported += supported.empty() ? "compute (opt-in)" : ", compute (opt-in)";
+            }
+            *explanation = line + " (driver supports: " + supported + ")";
+        }
+        return resolved;
+    }
+
+    const char* TierName(GLESMultiDrawMode tier) {
+        switch (tier) {
+        case GLESMultiDrawMode::Auto: return "auto";
+        case GLESMultiDrawMode::Ext: return "ext";
+        case GLESMultiDrawMode::MultiIndirect: return "multiindirect";
+        case GLESMultiDrawMode::Indirect: return "indirect";
+        case GLESMultiDrawMode::BaseVertex: return "basevertex";
+        case GLESMultiDrawMode::DrawElements: return "drawelements";
+        case GLESMultiDrawMode::Compute: return "compute";
+        }
+        return "unknown";
+    }
+
+    GLESMultiDrawMode ResolvedTier() {
+        ResolveTierOnce();
+        return g_resolvedTier;
+    }
+
+    String DescribeTierResolution() {
+        ResolveTierOnce();
+        return g_tierResolution;
+    }
+
+    void OnBackendContextDestroyed() {
+        g_indirectCommands = {};
+        g_rebasedIndices = {};
+        g_drawInfo = {};
+        g_flattenedIndices = {};
+        g_computeProgram = 0;
+        g_computeProgramFailed = false;
+        g_uElementSize = -1;
+        g_uDrawCount = -1;
+        g_uTotalIndices = -1;
+    }
+
+    void DrawElementsBatch(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                           GLsizei drawcount, const GLint* basevertex) {
+        if (drawcount <= 0 || !count || !indices) return;
+        // State-independent and possibly throwing, so it runs before any GL work.
+        CheckPrimitiveRestartSupported(type);
+
+        const Bool hasIndexBuffer = BoundIndexBuffer() != nullptr;
+
+        // The compute tier dispatches BEFORE the draw state is established: doing it
+        // afterwards would mean unpicking the program, SSBO and index bindings
+        // PrepareForDraw just made, and a dispatch inside an open transform feedback
+        // span is not legal at all. On success it hands back a flattened index stream.
+        FlattenedStream flattened;
+        if (ResolvedTier() == GLESMultiDrawMode::Compute && !CurrentProgramReadsDrawID()) {
+            FlattenWithCompute(mode, count, type, indices, drawcount, basevertex, flattened);
+        }
+
+        PrepareForDraw(DrawSyncBit::IndexBuffer);
+
+        if (flattened.indexCount != 0) {
+            const Uint previousIndexBinding = BoundIndexBufferId();
+            BufferImpl::BindBufferId(GL_ELEMENT_ARRAY_BUFFER, flattened.bufferId);
+            g_GLESFuncs.glDrawElements(mode, static_cast<GLsizei>(flattened.indexCount), GL_UNSIGNED_INT, nullptr);
+            BufferImpl::BindBufferId(GL_ELEMENT_ARRAY_BUFFER, previousIndexBinding);
+            return;
+        }
+
+        const Bool feedDrawID = CurrentProgramReadsDrawID();
+        const GLESMultiDrawMode tier = ResolveTierForBatch(feedDrawID, hasIndexBuffer);
+
+        Bool drawn = false;
+        switch (tier) {
+        case GLESMultiDrawMode::Ext:
+            drawn = RunExt(mode, count, type, indices, drawcount, basevertex);
+            break;
+        case GLESMultiDrawMode::MultiIndirect:
+            drawn = RunIndirect(mode, count, type, indices, drawcount, basevertex, /*batched=*/true, feedDrawID);
+            break;
+        case GLESMultiDrawMode::Indirect:
+            drawn = RunIndirect(mode, count, type, indices, drawcount, basevertex, /*batched=*/false, feedDrawID);
+            break;
+        case GLESMultiDrawMode::BaseVertex:
+            drawn = RunBaseVertexLoop(mode, count, type, indices, drawcount, basevertex, feedDrawID);
+            break;
+        case GLESMultiDrawMode::DrawElements:
+            drawn = RunRebasedDrawElements(mode, count, type, indices, drawcount, basevertex, feedDrawID);
+            break;
+        case GLESMultiDrawMode::Compute:
+            // Its pre-pass ran above; reaching here means it declined this batch's shape.
+            break;
+        case GLESMultiDrawMode::Auto:
+            break; // resolution never yields Auto
+        }
+
+        // Every tier above may decline a batch whose shape it cannot express. The two
+        // below are the floor: a base-vertex replay where the driver has one, and the
+        // rewritten index stream where it does not. Both are safe for any batch these
+        // entry points can receive.
+        if (!drawn) drawn = RunBaseVertexLoop(mode, count, type, indices, drawcount, basevertex, feedDrawID);
+        if (!drawn) drawn = RunRebasedDrawElements(mode, count, type, indices, drawcount, basevertex, feedDrawID);
+        if (!drawn) {
+            MGLOG_E("DirectGLES multi-draw: no usable tier for a %d sub-draw batch (mode 0x%x, type 0x%x); "
+                    "the batch was dropped",
+                    drawcount, mode, type);
+        }
+    }
+} // namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl

MobileGL/MG_Backend/DirectGLES/MultiDraw.h

@@ -0,0 +1,64 @@
+// MobileGL - MobileGL/MG_Backend/DirectGLES/MultiDraw.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+#include <Config.h>
+#include "DirectGLES.h"
+
+// Emulation of the desktop glMultiDrawElements / glMultiDrawElementsBaseVertex entry
+// points on OpenGL ES, which has neither in core.
+//
+// Every strategy below is an emulation; they differ only in which driver capability
+// they lean on and in how many driver entries a batch of N sub-draws costs. The design
+// follows MobileGlues (MobileGL-Dev/MobileGlues, gl/multidraw.cpp) tier for tier, plus
+// the native GL_EXT_multi_draw_arrays interaction that MobileGL already had:
+//
+//   Ext           one glMultiDrawElementsBaseVertexEXT           1 driver entry
+//   MultiIndirect one glMultiDrawElementsIndirectEXT             1 driver entry + 1 upload
+//   Indirect      N x glDrawElementsIndirect                     N + 1 upload
+//   BaseVertex    N x glDrawElementsBaseVertex                   N
+//   DrawElements  N x glDrawElements over CPU-rebased indices    N + 1 upload
+//   Compute       1 x glDrawElements over a GPU-flattened,       1 dispatch + 1 entry
+//                 rebased index stream
+//
+// Which one runs is resolved once per ES context from the driver's capabilities,
+// capped by MOBILEGL_ESPRYT_MULTIDRAW_MODE, and can additionally be demoted per batch
+// when the batch's own shape rules a tier out (see ResolveTierForBatch in the .cpp).
+namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
+    // The tier this ES context resolved to, computed on first use and stable after.
+    MG_Config::GLESMultiDrawMode ResolvedTier();
+    // "multiindirect", "compute", ... - stable identifiers, also used by the POST row.
+    const char* TierName(MG_Config::GLESMultiDrawMode tier);
+    // One line naming the resolved tier, the tiers the driver can support, and the env
+    // clamp if one applied. For DriverPost and the startup log.
+    String DescribeTierResolution();
+
+    // The resolution itself, as a pure function of a capability set: the backend feeds
+    // it the live ES context's capabilities, DriverPost feeds it the ones it probed
+    // standalone, and both therefore report the same tier. `explanation`, when non-null,
+    // receives the "requested -> resolved (driver supports: ...)" line.
+    MG_Config::GLESMultiDrawMode ResolveTier(const MG_External::GLESCapabilities& caps,
+                                             const MG_External::GLESFunctionsTable& funcs,
+                                             MG_Config::GLESMultiDrawMode requested, String* explanation);
+    // Whether one tier is runnable on the given capability set, for per-row POST output.
+    Bool IsTierSupported(const MG_External::GLESCapabilities& caps, const MG_External::GLESFunctionsTable& funcs,
+                         MG_Config::GLESMultiDrawMode tier);
+
+    // Runs `drawcount` indexed sub-draws as one glMultiDrawElements(BaseVertex) call
+    // would. `basevertex` is null for the plain glMultiDrawElements entry point (every
+    // base vertex is 0). Owns the whole draw, preparation included: callers must not
+    // have run PrepareForDraw, because the compute tier has to dispatch before the
+    // draw state is established.
+    void DrawElementsBatch(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                           GLsizei drawcount, const GLint* basevertex);
+
+    // The ES context is gone: every scratch buffer and the compute program belonged to
+    // it, so drop the names without deleting them (the dead context reclaims them).
+    void OnBackendContextDestroyed();
+} // namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl

MobileGL/MG_Backend/DirectGLES/Utils.cpp

@@ -61,29 +61,6 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 requestedInternalFormat, GetDriverPixelFormatNormalizeOptions() | extraOptions);
         }
 
-        // Multisample textures can only ever be rendered into, never uploaded to, so a fallback
-        // format for them has to stay colour-renderable - a three-channel float fallback is a legal
-        // ES texture format but not a legal multisample storage format. Widening to four channels
-        // is safe here precisely because there is no transfer path that would have to expand
-        // three-channel client data, and the alpha the draw writes for a three-channel source is
-        // already the 1.0 the frontend format implies.
-        Bool TargetRequiresRenderableFormat(SizeT targetIndex) {
-            return targetIndex == static_cast<SizeT>(TextureTarget::Texture2DMultisample) ||
-                   targetIndex == static_cast<SizeT>(TextureTarget::Texture2DMultisampleArray);
-        }
-
-        Flags<PixelFormatNormalizeOptionBit> GetRenderTargetNormalizeOptions(SizeT targetIndex) {
-            Flags<PixelFormatNormalizeOptionBit> options;
-            if (!TargetRequiresRenderableFormat(targetIndex)) {
-                return options;
-            }
-            options |= PixelFormatNormalizeOptionBit::NoThreeChannelRenderTarget;
-            if (!g_GLESCapabilities.SupportsRenderSnorm || !g_GLESCapabilities.SupportsNorm16Texture) {
-                options |= PixelFormatNormalizeOptionBit::NoSnorm16RenderTarget;
-            }
-            return options;
-        }
-
         Bool HasCachedFormatCapability(TextureInternalFormat internalFormat,
                                        SizeT targetIndex,
                                        Bool caveat,
@@ -141,14 +118,61 @@ namespace MobileGL::MG_Backend::DirectGLES {
             const GLenum requestedInternalFormat = MG_Util::ConvertTextureInternalFormatToGLEnum(internalFormat);
             Flags<PixelFormatNormalizeOptionBit> options;
             if (!pActiveBackendObject || ShouldUseCaveatFormat(internalFormat, targetIndex)) {
-                options = GetRuntimeFallbackNormalizeOptions(requestedInternalFormat,
+                options = GetRuntimeFallbackNormalizeOptions(
-                                                             GetRenderTargetNormalizeOptions(targetIndex));
+                    requestedInternalFormat,
+                    TextureImpl::GetRenderTargetNormalizeOptions(g_GLESCapabilities, targetIndex));
             }
             NormalizePixelFormat(requestedInternalFormat, options, outInternalFormat, outFormat, outType);
         }
     } // namespace
 
     namespace TextureImpl {
+        // Every image that can back a colour attachment needs a colour-renderable storage format,
+        // and ES has no renderable three-channel format at all: a three-channel float fallback is
+        // a legal ES texture but neither legal multisample storage nor a legal attachment, so
+        // GL_RGB8_SNORM / GL_RGB16F / ... have to be widened to four channels for any of them.
+        // This used to cover the multisample pair alone, on the grounds that only those can never
+        // be uploaded to; the transfer paths now expand three-channel client data themselves
+        // (Managers.cpp PrepareFallbackUpload) and hide the added alpha again on sample and
+        // readback, so the same substitution is available everywhere.
+        //
+        // The widening only ever *happens* where the driver refuses the native form (see
+        // PopulateFormatCapabilitiesImpl: outside multisample storage it rides the driver branch,
+        // behind the native probe), so a driver that does render to a three-channel image keeps
+        // allocating it byte for byte.
+        //
+        // Do NOT read that as "nothing changes off-device". Measured on Mesa 26.1.6 llvmpipe
+        // (the headless CI driver), an ES 3.2 GL_TEXTURE_2D colour attachment is COMPLETE for
+        // GL_RGB8 and GL_RGB16F but INCOMPLETE_ATTACHMENT for GL_RGB8_SNORM, GL_SRGB8 and every
+        // RGB integer format, and UNSUPPORTED for GL_RGB32F. Those eight formats therefore DO
+        // take the widened path on llvmpipe, which is where the retrace fixtures and the glcts
+        // green suites run - the substitution is driver-conditional, not desktop-exempt.
+        //
+        // A buffer texture is the one image that can never be an attachment; its storage is the
+        // buffer object's, and widening it would misdescribe the application's data.
+        Bool TargetRequiresRenderableFormat(SizeT targetIndex) {
+            if (targetIndex >= kFormatCapabilityTargetCount) {
+                return false;
+            }
+            if (targetIndex == kFormatCapabilityRenderbufferTargetIndex) {
+                return true;
+            }
+            return static_cast<TextureTarget>(targetIndex) != TextureTarget::TextureBuffer;
+        }
+
+        Flags<PixelFormatNormalizeOptionBit> GetRenderTargetNormalizeOptions(
+            const MG_External::GLESCapabilities& capabilities, SizeT targetIndex) {
+            Flags<PixelFormatNormalizeOptionBit> options;
+            if (!TargetRequiresRenderableFormat(targetIndex)) {
+                return options;
+            }
+            options |= PixelFormatNormalizeOptionBit::NoThreeChannelRenderTarget;
+            if (!capabilities.SupportsRenderSnorm || !capabilities.SupportsNorm16Texture) {
+                options |= PixelFormatNormalizeOptionBit::NoSnorm16RenderTarget;
+            }
+            return options;
+        }
+
         void GenerateTextureFormatInfo(TextureInternalFormat internalFormat, GLenum* outInternalFormat,
                                        GLenum* outFormat, GLenum* outType, TextureTarget target) {
 #ifdef TRACY_ENABLE
@@ -178,20 +202,29 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return ShouldUseCaveatFormat(internalFormat, GetRenderbufferFormatCapabilityTargetIndex());
         }
 
+        namespace {
+            Bool BackendFormatAddsAlpha(TextureInternalFormat internalFormat, SizeT targetIndex) {
+                if (!TargetRequiresRenderableFormat(targetIndex)) {
+                    return false;
+                }
+                if (pActiveBackendObject && !ShouldUseCaveatFormat(internalFormat, targetIndex)) {
+                    return false;
+                }
+                const GLenum requestedInternalFormat = MG_Util::ConvertTextureInternalFormatToGLEnum(internalFormat);
+                const Flags<PixelFormatNormalizeOptionBit> options = GetRuntimeFallbackNormalizeOptions(
+                    requestedInternalFormat, GetRenderTargetNormalizeOptions(g_GLESCapabilities, targetIndex));
+                return static_cast<Bool>(options & PixelFormatNormalizeOptionBit::NoThreeChannelRenderTarget);
+            }
+        } // namespace
+
         Bool BackendTextureFormatAddsAlpha(TextureInternalFormat internalFormat, TextureTarget target) {
             const SizeT targetIndex =
                 target == TextureTarget::Unknown ? kFormatCapabilityTargetCount : GetFormatCapabilityTargetIndex(target);
-            if (!TargetRequiresRenderableFormat(targetIndex)) {
+            return BackendFormatAddsAlpha(internalFormat, targetIndex);
-                return false;
+        }
-            }
+
-            if (pActiveBackendObject && !ShouldUseCaveatFormat(internalFormat, targetIndex)) {
+        Bool BackendRenderbufferFormatAddsAlpha(TextureInternalFormat internalFormat) {
-                return false;
+            return BackendFormatAddsAlpha(internalFormat, GetRenderbufferFormatCapabilityTargetIndex());
-            }
-            const GLenum requestedInternalFormat = MG_Util::ConvertTextureInternalFormatToGLEnum(internalFormat);
-            const Flags<PixelFormatNormalizeOptionBit> options =
-                GetRuntimeFallbackNormalizeOptions(requestedInternalFormat,
-                                                   GetRenderTargetNormalizeOptions(targetIndex));
-            return static_cast<Bool>(options & PixelFormatNormalizeOptionBit::NoThreeChannelRenderTarget);
         }
     } // namespace TextureImpl
     namespace PrgramImpl {
@@ -1107,6 +1140,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 }
             }
         }
+            if (pixelPackBufferObject) {
+                // WritebackFromBackend bumps change serials with no backend op; re-open
+                // the buffer draw-clean memos (once for the whole row loop).
+                BufferImpl::BumpBufferMutationEpoch();
+            }
             return true;
         }
     } // namespace ReadbackImpl

MobileGL/MG_Backend/DirectGLES/Utils.h

@@ -9,6 +9,8 @@
 #pragma once
 #include <Includes.h>
 #include <MG_State/GLState/Core.h>
+#include <MG_Util/BackendLoaders/OpenGL/Loader.h>
+#include <MG_Util/Texture/TextureFormatProcessor.h>
 
 namespace MobileGL::MG_Backend::DirectGLES {
     namespace DebugImpl {
@@ -34,6 +36,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
     } // namespace VertexArrayImpl
 
     namespace TextureImpl {
+        // Whether images on this format-capability target can back a colour attachment, and so
+        // need a colour-renderable storage format even when the frontend asked for a
+        // three-channel one ES never renders to. Shared by the capability probe (which passes the
+        // capabilities it has just queried, before the globals are published) and by the
+        // allocation path (which reads the active backend's), so the format the cache was probed
+        // with is always the format the image is created with.
+        Bool TargetRequiresRenderableFormat(SizeT targetIndex);
+        Flags<PixelFormatNormalizeOptionBit> GetRenderTargetNormalizeOptions(
+            const MG_External::GLESCapabilities& capabilities, SizeT targetIndex);
+
         void GenerateTextureFormatInfo(TextureInternalFormat internalFormat, GLenum* outInternalFormat,
                                        GLenum* outFormat, GLenum* outType,
                                        TextureTarget target = TextureTarget::Unknown);
@@ -41,10 +53,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                             GLenum* outFormat, GLenum* outType);
         Bool ShouldUseCaveatTextureFormat(TextureInternalFormat internalFormat, TextureTarget target);
 
-        // True when the format the texture is actually created with has an alpha channel the
+        // True when the format the image is actually created with has an alpha channel the
-        // frontend format does not (the three-channel multisample widening). GL reads such a
+        // frontend format does not (the three-channel colour-renderable widening). GL reads such
-        // channel back as 1.0, so any swizzle source of ALPHA has to be answered with ONE.
+        // a channel back as 1.0, so any swizzle source of ALPHA has to be answered with ONE and
+        // any readback of the image has to overwrite the alpha the draw happened to leave there.
         Bool BackendTextureFormatAddsAlpha(TextureInternalFormat internalFormat, TextureTarget target);
+        Bool BackendRenderbufferFormatAddsAlpha(TextureInternalFormat internalFormat);
         Bool ShouldUseCaveatRenderbufferFormat(TextureInternalFormat internalFormat);
     } // namespace TextureImpl
 

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -16,6 +16,7 @@
 #include "MG_Util/Converters/MGToStr/TextureEnumConverter.h"
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
 #include "MG_Util/Texture/TextureFormatProcessor.h"
+#include "MG_Util/Async/ShaderCompilePool.h"
 
 #include <Config.h>
 #include <cmath>
@@ -523,6 +524,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (shaderSubgroupSupported && !MG_Config::Features.DisableSubgroup) {
             extensions.push_back(E_GL_KHR_shader_subgroup);
         }
+        // GL_KHR_parallel_shader_compile is MobileGL's own capability, not the Vulkan
+        // device's: the compiler threads belong to MobileGL's shader pool and
+        // glCompileShader/glLinkProgram are serviced entirely inside the frontend, so there
+        // is no device feature to condition this on.
+        //
+        // Gated on the async flag deliberately, and this is the whole reason the gate
+        // exists. Advertising the string is the one part of asynchronous compilation that a
+        // recorded trace can never cover: Iris and Sodium change their SUBMISSION SCHEDULE
+        // the moment they see it - they enqueue whole pipeline batches and poll
+        // GL_COMPLETION_STATUS_KHR instead of compiling one program at a time - so
+        // MOBILEGL_ASYNC_SHADER_COMPILE=0 has to withdraw the application-visible behaviour
+        // change as well as the threading, or the kill switch would only be half a switch.
+        if (MG_Util::Async::AsyncShaderCompileEnabled()) {
+            extensions.push_back(E_GL_KHR_parallel_shader_compile);
+        }
         // GL_ARB_timer_query gates MC's F3 GPU% (LWJGL checks the extension string);
         // only advertised when the device actually supports timestamp queries and the
         // MOBILEGL_DISABLE_TIMERQUERY escape hatch is off.
@@ -601,12 +617,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             funcsTable.GL.GetIntegeri_v = GetIntegeri_v;
             funcsTable.GL.GetInteger64i_v = GetInteger64i_v;
             funcsTable.GL.GetProgramiv = GetProgramiv;
-            funcsTable.GL.GetProgramInterfaceiv = GetProgramInterfaceiv;
-            funcsTable.GL.GetProgramResourceIndex = GetProgramResourceIndex;
-            funcsTable.GL.GetProgramResourceName = GetProgramResourceName;
-            funcsTable.GL.GetProgramResourceiv = GetProgramResourceiv;
-            funcsTable.GL.GetProgramResourceLocation = GetProgramResourceLocation;
-            funcsTable.GL.GetProgramResourceLocationIndex = GetProgramResourceLocationIndex;
             funcsTable.GL.ShaderStorageBlockBinding = ShaderStorageBlockBinding;
             funcsTable.GL.FenceSync = FenceSync;
             funcsTable.GL.ClientWaitSync = ClientWaitSync;

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp

@@ -16,6 +16,7 @@
 #include "MG_Util/Metrics/TextureMetrics.h"
 #include "MG_Util/Miscellany/IndexGenerator.h"
 #include <atomic>
+#include <bit>
 #include <cstring>
 #include <spirv_reflect.h>
 
@@ -231,6 +232,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     StorageBlockResource block{};
                     block.name = blockName;
                     block.binding = binding->binding;
+                    // glShaderStorageBlockBinding survives every rebuild of this cache: the
+                    // authoritative record of a rebound block lives on the program (it is what
+                    // GL_BUFFER_BINDING reports), and only the shader's declared binding is
+                    // recoverable from the SPIR-V. Without this, any unrelated state-version
+                    // bump would silently revert the block to its declared binding.
+                    const Int rebound = program.GetShaderStorageBlockBindingOverride(blockName);
+                    if (rebound >= 0) block.binding = static_cast<Uint32>(rebound);
                     block.dataSize = static_cast<GLint>(binding->block.size);
                     const GLuint blockIndex = static_cast<GLuint>(cache.storageBlocks.size());
                     AddBufferVariablesRecursive(binding->block, blockName, blockIndex, cache.bufferVariables,
@@ -255,18 +263,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return programObject.get();
         }
 
-        void CopyResourceName(const String& source, GLsizei bufSize, GLsizei* length, GLchar* name) {
-            const GLsizei writtenLength = static_cast<GLsizei>(source.size());
-            if (length) {
-                *length = writtenLength;
-            }
-            if (name && bufSize > 0) {
-                const GLsizei copyLength = std::min<GLsizei>(bufSize - 1, writtenLength);
-                std::memcpy(name, source.data(), static_cast<SizeT>(copyLength));
-                name[copyLength] = '\0';
-            }
-        }
-
         const Uint8* ResolveIndirectCommandBytes(const void* indirect, SizeT requiredBytes, const char* label) {
             auto drawBuffer = MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::DrawIndirect).GetBoundObject();
             if (drawBuffer) {
@@ -287,100 +283,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return reinterpret_cast<const Uint8*>(indirect);
         }
 
-        Vector<GLuint> GetUniformBlockActiveVariables(const MG_State::GLState::ProgramObject& program,
-                                                      GLuint blockIndex) {
-            Vector<GLuint> activeVariables;
-            const Uint uniformCount = program.GetUniformCount();
-            activeVariables.reserve(uniformCount);
-            for (Uint uniformIndex = 0; uniformIndex < uniformCount; ++uniformIndex) {
-                if (program.GetActiveUniformBlockIndex(uniformIndex) == static_cast<Int>(blockIndex)) {
-                    activeVariables.push_back(uniformIndex);
-                }
-            }
-            return activeVariables;
-        }
-
-        GLuint FindProgramInputIndex(const MG_State::GLState::ProgramObject& program, const String& name) {
-            const Int activeCount = program.GetActiveAttributesCount();
-            for (Int index = 0; index < activeCount; ++index) {
-                if (program.GetActiveAttribName(index) == name) {
-                    return static_cast<GLuint>(index);
-                }
-            }
-            return GL_INVALID_INDEX;
-        }
-
-        GLuint FindProgramOutputIndex(const MG_State::GLState::ProgramObject& program, const String& name) {
-            const Int activeCount = program.GetActiveFragmentOutputCount();
-            for (Int index = 0; index < activeCount; ++index) {
-                if (program.GetActiveFragmentOutputName(index) == name) {
-                    return static_cast<GLuint>(index);
-                }
-            }
-            return GL_INVALID_INDEX;
-        }
-
-        GLint GetProgramOutputLocation(const MG_State::GLState::ProgramObject& program, const String& name) {
-            const Int activeCount = program.GetActiveFragmentOutputCount();
-            for (Int index = 0; index < activeCount; ++index) {
-                if (program.GetActiveFragmentOutputName(index) == name) {
-                    return program.GetFragmentOutputLocation(index);
-                }
-            }
-            return -1;
-        }
-
-        GLint GetProgramResourceActiveCount(const MG_State::GLState::ProgramObject& program, GLenum programInterface,
-                                            const ProgramResourceCache& cache) {
-            switch (programInterface) {
-            case GL_SHADER_STORAGE_BLOCK:
-                return static_cast<GLint>(cache.storageBlocks.size());
-            case GL_BUFFER_VARIABLE:
-                return static_cast<GLint>(cache.bufferVariables.size());
-            case GL_UNIFORM_BLOCK:
-                return program.GetActiveUniformBlocksCount();
-            case GL_UNIFORM:
-                return static_cast<GLint>(program.GetUniformCount());
-            case GL_PROGRAM_INPUT:
-                return program.GetActiveAttributesCount();
-            case GL_PROGRAM_OUTPUT:
-                return program.GetActiveFragmentOutputCount();
-            default:
-                return 0;
-            }
-        }
-
-        GLint GetProgramResourceMaxNameLength(const MG_State::GLState::ProgramObject& program, GLenum programInterface,
-                                              const ProgramResourceCache& cache) {
-            switch (programInterface) {
-            case GL_SHADER_STORAGE_BLOCK: {
-                SizeT maxLength = 0;
-                for (const auto& block : cache.storageBlocks) maxLength = std::max(maxLength, block.name.size() + 1);
-                return static_cast<GLint>(maxLength);
-            }
-            case GL_BUFFER_VARIABLE: {
-                SizeT maxLength = 0;
-                for (const auto& var : cache.bufferVariables) maxLength = std::max(maxLength, var.name.size() + 1);
-                return static_cast<GLint>(maxLength);
-            }
-            case GL_UNIFORM_BLOCK:
-                return program.GetActiveUniformBlocksMaxNameLength() + 1;
-            case GL_UNIFORM:
-                return program.GetUniformMaxLength() + 1;
-            case GL_PROGRAM_INPUT:
-                return program.GetActiveAttributesMaxLength() + 1;
-            case GL_PROGRAM_OUTPUT: {
-                SizeT maxLength = 0;
-                const Int activeCount = program.GetActiveFragmentOutputCount();
-                for (Int index = 0; index < activeCount; ++index) {
-                    maxLength = std::max(maxLength, program.GetActiveFragmentOutputName(index).size() + 1);
-                }
-                return static_cast<GLint>(maxLength);
-            }
-            default:
-                return 0;
-            }
-        }
     } // namespace
 
     void ClearProgramResourceCaches() {
@@ -394,11 +296,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     GLuint GetShaderStorageBlockIndex(const MG_State::GLState::ProgramObject& program, const String& name) {
         auto& cache = GetProgramResourceCache(program);
-        const auto it = std::find_if(cache.storageBlocks.begin(), cache.storageBlocks.end(),
+        auto find = [&cache](const String& key) {
-            [&](const StorageBlockResource& block) { return block.name == name; });
+            return std::find_if(cache.storageBlocks.begin(), cache.storageBlocks.end(),
-        return it == cache.storageBlocks.end()
+                [&](const StorageBlockResource& block) { return block.name == key; });
-            ? GL_INVALID_INDEX
+        };
-            : static_cast<GLuint>(std::distance(cache.storageBlocks.begin(), it));
+        auto it = find(name);
+        if (it == cache.storageBlocks.end()) {
+            // Cache names are normalized (NormalizeDescriptorName drops the array suffix), so
+            // an arrayed block that GL enumerates per element - "B[0]", "B[1]" - is one entry
+            // here, spelled "B". Retry against the bare name before giving up.
+            const auto bracket = name.rfind('[');
+            if (bracket == String::npos || name.empty() || name.back() != ']') return GL_INVALID_INDEX;
+            it = find(name.substr(0, bracket));
+            if (it == cache.storageBlocks.end()) return GL_INVALID_INDEX;
+        }
+        return static_cast<GLuint>(std::distance(cache.storageBlocks.begin(), it));
     }
 
     GLuint GetShaderStorageBlockBinding(const MG_State::GLState::ProgramObject& program, GLuint blockIndex) {
@@ -873,357 +785,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
-    void GetProgramInterfaceiv(GLuint program, GLenum programInterface, GLenum pname, GLint* params) {
+    void ShaderStorageBlockBinding(GLuint program, const GLchar* storageBlockName, GLuint storageBlockBinding) {
-        if (!params) return;
         auto* programObject = TryGetDirectVulkanProgram(program);
-        if (!programObject) return;
+        if (!programObject || storageBlockName == nullptr) return;
-        auto& cache = GetProgramResourceCache(*programObject);
-        switch (pname) {
-        case GL_ACTIVE_RESOURCES:
-            *params = GetProgramResourceActiveCount(*programObject, programInterface, cache);
-            return;
-        case GL_MAX_NAME_LENGTH:
-            *params = GetProgramResourceMaxNameLength(*programObject, programInterface, cache);
-            return;
-        case GL_MAX_NUM_ACTIVE_VARIABLES:
-            if (programInterface == GL_SHADER_STORAGE_BLOCK) {
-                SizeT maxCount = 0;
-                for (const auto& block : cache.storageBlocks) {
-                    maxCount = std::max(maxCount, block.activeVariables.size());
-                }
-                *params = static_cast<GLint>(maxCount);
-            } else if (programInterface == GL_UNIFORM_BLOCK) {
-                GLint maxCount = 0;
-                const Int activeBlocks = programObject->GetActiveUniformBlocksCount();
-                for (Int index = 0; index < activeBlocks; ++index) {
-                    maxCount = std::max(maxCount, programObject->GetUniformBlockActiveUniformCount(index));
-                }
-                *params = maxCount;
-            } else {
-                *params = 0;
-            }
-            return;
-        default:
-            *params = 0;
-            return;
-        }
-    }
-
-    GLuint GetProgramResourceIndex(GLuint program, GLenum programInterface, const GLchar* name) {
-        if (!name) return GL_INVALID_INDEX;
-        auto* programObject = TryGetDirectVulkanProgram(program);
-        if (!programObject) return GL_INVALID_INDEX;
-        auto& cache = GetProgramResourceCache(*programObject);
-        const String resourceName = name;
-        if (programInterface == GL_SHADER_STORAGE_BLOCK) {
-            return GetShaderStorageBlockIndex(*programObject, name);
-        }
-        if (programInterface == GL_BUFFER_VARIABLE) {
-            const auto it = std::find_if(cache.bufferVariables.begin(), cache.bufferVariables.end(),
-                [&](const BufferVariableResource& var) { return var.name == resourceName; });
-            return it == cache.bufferVariables.end()
-                ? GL_INVALID_INDEX
-                : static_cast<GLuint>(std::distance(cache.bufferVariables.begin(), it));
-        }
-        if (programInterface == GL_UNIFORM_BLOCK) {
-            return programObject->GetUniformBlockIndex(name);
-        }
-        if (programInterface == GL_UNIFORM) {
-            const Int activeUniformIndex = programObject->GetActiveUniformIndex(resourceName);
-            return activeUniformIndex >= 0 ? static_cast<GLuint>(activeUniformIndex) : GL_INVALID_INDEX;
-        }
-        if (programInterface == GL_PROGRAM_INPUT) {
-            return FindProgramInputIndex(*programObject, resourceName);
-        }
-        if (programInterface == GL_PROGRAM_OUTPUT) {
-            return FindProgramOutputIndex(*programObject, resourceName);
-        }
-        return GL_INVALID_INDEX;
-    }
-
-    void GetProgramResourceName(GLuint program, GLenum programInterface, GLuint index, GLsizei bufSize,
-                                GLsizei* length, GLchar* name) {
-        auto* programObject = TryGetDirectVulkanProgram(program);
-        if (!programObject) return;
-        auto& cache = GetProgramResourceCache(*programObject);
-        if (programInterface == GL_SHADER_STORAGE_BLOCK && index < cache.storageBlocks.size()) {
-            CopyResourceName(cache.storageBlocks[index].name, bufSize, length, name);
-            return;
-        }
-        if (programInterface == GL_BUFFER_VARIABLE && index < cache.bufferVariables.size()) {
-            CopyResourceName(cache.bufferVariables[index].name, bufSize, length, name);
-            return;
-        }
-        if (programInterface == GL_UNIFORM_BLOCK && programObject->IsActiveUniformBlock(index)) {
-            CopyResourceName(programObject->GetUniformBlockName(index), bufSize, length, name);
-            return;
-        }
-        if (programInterface == GL_UNIFORM && index < programObject->GetUniformCount()) {
-            CopyResourceName(programObject->GetActiveUniformName(index), bufSize, length, name);
-            return;
-        }
-        if (programInterface == GL_PROGRAM_INPUT && index < static_cast<GLuint>(programObject->GetActiveAttributesCount())) {
-            CopyResourceName(programObject->GetActiveAttribName(index), bufSize, length, name);
-            return;
-        }
-        if (programInterface == GL_PROGRAM_OUTPUT &&
-            index < static_cast<GLuint>(programObject->GetActiveFragmentOutputCount())) {
-            CopyResourceName(programObject->GetActiveFragmentOutputName(index), bufSize, length, name);
-            return;
-        }
-        if (length) *length = 0;
-        if (name && bufSize > 0) name[0] = '\0';
-    }
-
-    void GetProgramResourceiv(GLuint program, GLenum programInterface, GLuint index, GLsizei propCount,
-                              const GLenum* props, GLsizei bufSize, GLsizei* length, GLint* params) {
-        auto* programObject = TryGetDirectVulkanProgram(program);
-        if (!programObject || !props || !params || bufSize <= 0) return;
-        auto& cache = GetProgramResourceCache(*programObject);
-        GLsizei written = 0;
-        auto writeValue = [&](GLint value) {
-            if (written < bufSize) {
-                params[written++] = value;
-            }
-        };
-
-        for (GLsizei propIndex = 0; propIndex < propCount; ++propIndex) {
-            const GLenum prop = props[propIndex];
-            if (programInterface == GL_SHADER_STORAGE_BLOCK && index < cache.storageBlocks.size()) {
-                const auto& block = cache.storageBlocks[index];
-                switch (prop) {
-                case GL_NAME_LENGTH:
-                    writeValue(static_cast<GLint>(block.name.size() + 1));
-                    break;
-                case GL_BUFFER_BINDING:
-                    writeValue(static_cast<GLint>(block.binding));
-                    break;
-                case GL_BUFFER_DATA_SIZE:
-                    writeValue(block.dataSize);
-                    break;
-                case GL_NUM_ACTIVE_VARIABLES:
-                    writeValue(static_cast<GLint>(block.activeVariables.size()));
-                    break;
-                case GL_ACTIVE_VARIABLES:
-                    for (const auto variable : block.activeVariables) writeValue(static_cast<GLint>(variable));
-                    break;
-                default:
-                    writeValue(0);
-                    break;
-                }
-            } else if (programInterface == GL_BUFFER_VARIABLE && index < cache.bufferVariables.size()) {
-                const auto& var = cache.bufferVariables[index];
-                switch (prop) {
-                case GL_NAME_LENGTH:
-                    writeValue(static_cast<GLint>(var.name.size() + 1));
-                    break;
-                case GL_TYPE:
-                    writeValue(GL_FLOAT);
-                    break;
-                case GL_ARRAY_SIZE:
-                    writeValue(1);
-                    break;
-                case GL_OFFSET:
-                    writeValue(var.offset);
-                    break;
-                case GL_BLOCK_INDEX:
-                    writeValue(static_cast<GLint>(var.blockIndex));
-                    break;
-                case GL_ARRAY_STRIDE:
-                case GL_MATRIX_STRIDE:
-                case GL_TOP_LEVEL_ARRAY_SIZE:
-                case GL_TOP_LEVEL_ARRAY_STRIDE:
-                case GL_IS_ROW_MAJOR:
-                    writeValue(0);
-                    break;
-                default:
-                    writeValue(0);
-                    break;
-                }
-            } else if (programInterface == GL_UNIFORM_BLOCK &&
-                       programObject->IsActiveUniformBlock(index)) {
-                const auto activeVariables = GetUniformBlockActiveVariables(*programObject, index);
-                switch (prop) {
-                case GL_NAME_LENGTH:
-                    writeValue(static_cast<GLint>(programObject->GetUniformBlockName(index).size() + 1));
-                    break;
-                case GL_BUFFER_BINDING:
-                    writeValue(static_cast<GLint>(programObject->GetUniformBlockBinding(index)));
-                    break;
-                case GL_BUFFER_DATA_SIZE:
-                    writeValue(static_cast<GLint>(programObject->GetUBOSizeAt(index)));
-                    break;
-                case GL_NUM_ACTIVE_VARIABLES:
-                    writeValue(static_cast<GLint>(activeVariables.size()));
-                    break;
-                case GL_ACTIVE_VARIABLES:
-                    for (const GLuint variableIndex : activeVariables) {
-                        writeValue(static_cast<GLint>(variableIndex));
-                    }
-                    break;
-                case GL_REFERENCED_BY_VERTEX_SHADER:
-                    writeValue(programObject->IsUniformBlockReferencedByStage(index, EShLangVertex) ? GL_TRUE
-                                                                                                     : GL_FALSE);
-                    break;
-                case GL_REFERENCED_BY_FRAGMENT_SHADER:
-                    writeValue(programObject->IsUniformBlockReferencedByStage(index, EShLangFragment) ? GL_TRUE
-                                                                                                       : GL_FALSE);
-                    break;
-                case GL_REFERENCED_BY_COMPUTE_SHADER:
-                    writeValue(programObject->IsUniformBlockReferencedByStage(index, EShLangCompute) ? GL_TRUE
-                                                                                                      : GL_FALSE);
-                    break;
-                case GL_REFERENCED_BY_GEOMETRY_SHADER:
-                case GL_REFERENCED_BY_TESS_CONTROL_SHADER:
-                case GL_REFERENCED_BY_TESS_EVALUATION_SHADER:
-                    writeValue(GL_FALSE);
-                    break;
-                default:
-                    writeValue(0);
-                    break;
-                }
-            } else if (programInterface == GL_UNIFORM && index < programObject->GetUniformCount()) {
-                const auto& uniformName = programObject->GetActiveUniformName(index);
-                const GLint location = programObject->GetUniformLocation(uniformName);
-                switch (prop) {
-                case GL_NAME_LENGTH:
-                    writeValue(static_cast<GLint>(uniformName.size() + 1));
-                    break;
-                case GL_TYPE:
-                    writeValue(static_cast<GLint>(programObject->GetActiveUniformType(index)));
-                    break;
-                case GL_ARRAY_SIZE:
-                    writeValue(programObject->GetActiveUniformArraySize(index));
-                    break;
-                case GL_BLOCK_INDEX:
-                    writeValue(programObject->GetActiveUniformBlockIndex(index));
-                    break;
-                case GL_LOCATION:
-                    writeValue(location);
-                    break;
-                case GL_OFFSET:
-                    writeValue(location >= 0 && programObject->IsValidUniformLocation(location)
-                                   ? static_cast<GLint>(programObject->GetUniformOffset(location))
-                                   : 0);
-                    break;
-                case GL_ARRAY_STRIDE:
-                case GL_MATRIX_STRIDE:
-                case GL_IS_ROW_MAJOR:
-                case GL_TOP_LEVEL_ARRAY_SIZE:
-                case GL_TOP_LEVEL_ARRAY_STRIDE:
-                case GL_REFERENCED_BY_VERTEX_SHADER:
-                case GL_REFERENCED_BY_FRAGMENT_SHADER:
-                case GL_REFERENCED_BY_COMPUTE_SHADER:
-                case GL_REFERENCED_BY_GEOMETRY_SHADER:
-                case GL_REFERENCED_BY_TESS_CONTROL_SHADER:
-                case GL_REFERENCED_BY_TESS_EVALUATION_SHADER:
-                    writeValue(0);
-                    break;
-                default:
-                    writeValue(0);
-                    break;
-                }
-            } else if (programInterface == GL_PROGRAM_INPUT &&
-                       index < static_cast<GLuint>(programObject->GetActiveAttributesCount())) {
-                const auto& resourceName = programObject->GetActiveAttribName(index);
-                switch (prop) {
-                case GL_NAME_LENGTH:
-                    writeValue(static_cast<GLint>(resourceName.size() + 1));
-                    break;
-                case GL_TYPE:
-                    writeValue(static_cast<GLint>(programObject->GetActiveAttribType(index)));
-                    break;
-                case GL_ARRAY_SIZE:
-                    writeValue(programObject->GetActiveAttribArraySize(index));
-                    break;
-                case GL_LOCATION:
-                    writeValue(programObject->GetAttributeLocation(resourceName));
-                    break;
-                case GL_REFERENCED_BY_VERTEX_SHADER:
-                    writeValue(GL_TRUE);
-                    break;
-                case GL_REFERENCED_BY_FRAGMENT_SHADER:
-                case GL_REFERENCED_BY_COMPUTE_SHADER:
-                case GL_REFERENCED_BY_GEOMETRY_SHADER:
-                case GL_REFERENCED_BY_TESS_CONTROL_SHADER:
-                case GL_REFERENCED_BY_TESS_EVALUATION_SHADER:
-                case GL_IS_PER_PATCH:
-                case GL_LOCATION_INDEX:
-                    writeValue(0);
-                    break;
-                default:
-                    writeValue(0);
-                    break;
-                }
-            } else if (programInterface == GL_PROGRAM_OUTPUT &&
-                       index < static_cast<GLuint>(programObject->GetActiveFragmentOutputCount())) {
-                const auto& resourceName = programObject->GetActiveFragmentOutputName(index);
-                switch (prop) {
-                case GL_NAME_LENGTH:
-                    writeValue(static_cast<GLint>(resourceName.size() + 1));
-                    break;
-                case GL_TYPE:
-                    writeValue(static_cast<GLint>(programObject->GetFragmentOutputType(index)));
-                    break;
-                case GL_ARRAY_SIZE:
-                    writeValue(programObject->GetActiveFragmentOutputArraySize(index));
-                    break;
-                case GL_LOCATION:
-                    writeValue(programObject->GetFragmentOutputLocation(index));
-                    break;
-                case GL_LOCATION_INDEX:
-                    writeValue(0);
-                    break;
-                case GL_REFERENCED_BY_FRAGMENT_SHADER:
-                    writeValue(GL_TRUE);
-                    break;
-                case GL_REFERENCED_BY_VERTEX_SHADER:
-                case GL_REFERENCED_BY_COMPUTE_SHADER:
-                case GL_REFERENCED_BY_GEOMETRY_SHADER:
-                case GL_REFERENCED_BY_TESS_CONTROL_SHADER:
-                case GL_REFERENCED_BY_TESS_EVALUATION_SHADER:
-                case GL_IS_PER_PATCH:
-                    writeValue(0);
-                    break;
-                default:
-                    writeValue(0);
-                    break;
-                }
-            } else {
-                writeValue(0);
-            }
-        }
-        if (length) *length = written;
-    }
-
-    GLint GetProgramResourceLocation(GLuint program, GLenum programInterface, const GLchar* name) {
-        auto* programObject = TryGetDirectVulkanProgram(program);
-        if (!programObject || !name) return -1;
-        if (programInterface == GL_UNIFORM) {
-            return programObject->GetUniformLocation(name);
-        }
-        if (programInterface == GL_PROGRAM_INPUT) {
-            return programObject->GetAttributeLocation(name);
-        }
-        if (programInterface == GL_PROGRAM_OUTPUT) {
-            return GetProgramOutputLocation(*programObject, name);
-        }
-        return -1;
-    }
-
-    GLint GetProgramResourceLocationIndex(GLuint program, GLenum programInterface, const GLchar* name) {
-        auto* programObject = TryGetDirectVulkanProgram(program);
-        if (!programObject || !name) return -1;
-        if (programInterface == GL_PROGRAM_OUTPUT) {
-            return GetProgramOutputLocation(*programObject, name) >= 0 ? 0 : -1;
-        }
-        return -1;
-    }
-
-    void ShaderStorageBlockBinding(GLuint program, GLuint storageBlockIndex, GLuint storageBlockBinding) {
-        auto* programObject = TryGetDirectVulkanProgram(program);
-        if (!programObject) return;
-        auto& cache = GetProgramResourceCache(*programObject);
         const Int maxBindings = pActiveBackendObject
             ? pActiveBackendObject->GetDynamicParameters().MaxShaderStorageBufferBindings
             : 0;
@@ -1233,13 +797,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 MakeUnique<GenericErrorInfo>("DirectVulkan", __func__, "Shader storage binding is out of range."));
             return;
         }
-        if (storageBlockIndex >= cache.storageBlocks.size()) {
+        // The frontend already validated that the name denotes an active block, and has
-            MG_State::pGLContext->RecordError(
+        // already recorded the new binding on the program - which is what reseeds this cache
-                ErrorCode::InvalidValue,
+        // whenever it is rebuilt. Writing the entry here as well keeps an ALREADY-BUILT cache
-                MakeUnique<GenericErrorInfo>("DirectVulkan", __func__, "Shader storage block index is not active."));
+        // (the common case: the very next draw reads it) from having to be thrown away.
-            return;
+        auto& cache = GetProgramResourceCache(*programObject);
-        }
+        const GLuint blockIndex = GetShaderStorageBlockIndex(*programObject, storageBlockName);
-        cache.storageBlocks[storageBlockIndex].binding = storageBlockBinding;
+        if (blockIndex == GL_INVALID_INDEX) return;
+        cache.storageBlocks[blockIndex].binding = storageBlockBinding;
     }
     void ReadPixels(GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels) {
         MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::ReadPixels called with null VulkanRenderer");
@@ -1392,31 +957,66 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         pVulkanRenderer->MultiDrawArrays(payload);
     }
 
+    // Shared body of glMultiDrawElements (basevertex == nullptr) and
+    // glMultiDrawElementsBaseVertex: identical calls except for the per-draw
+    // vertex offset, which VkMultiDrawIndexedInfoEXT / VkDrawIndexedIndirectCommand /
+    // vkCmdDrawIndexed all carry natively.
+    static void MultiDrawElementsImpl(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                                      GLsizei drawcount, const GLint* basevertex) {
+        if (drawcount <= 0) {
+            return;
+        }
+        MultiDrawIndexedCmd payload{};
+        payload.mode = mode;
+        payload.indexBufferView.indexType = type;
+
+        // Loop-invariant: the index type is fixed for the whole multi-draw, so resolve
+        // its byte size once instead of twice per sub-draw (a cross-TU switch that
+        // showed up in per-frame profiles of sodium-style 132x32 multi-draws). Index
+        // sizes are 1/2/4, so the per-sub-draw offset division below reduces to a
+        // shift - the hardware divide was the hottest instruction of this loop.
+        const SizeT indexSize = MG_Util::GetGLTypeSize(type);
+        if (indexSize == 0) {
+            MGLOG_E("MultiDrawElements skipped: unsupported index type 0x%x", type);
+            return;
+        }
+        const Uint32 indexSizeShift = static_cast<Uint32>(std::countr_zero(indexSize));
+
+        // TODO: allocate draw cmd buf elsewhere
+        static Vector<DrawIndexedCmdParam> params;
+        params.clear();
+        params.resize(drawcount);
+
+        for (GLsizei i = 0; i < drawcount; ++i) {
+            if (count[i] == 0) {
+                continue;
+            }
+
+            // TODO: this index view needs a redesign, now there's a lotta redundant uploads
+
+            payload.indexBufferView.indexByteOffset = 0;
+            payload.indexBufferView.indexByteSize =
+                    std::max(reinterpret_cast<SizeT>(indices[i]) + count[i] * indexSize,
+                             payload.indexBufferView.indexByteSize);
+
+            auto& param = params[i];
+
+            param.indexCount = count[i];
+            param.instanceCount = 1;
+            param.firstIndex = reinterpret_cast<SizeT>(indices[i]) >> indexSizeShift;
+            param.vertexOffset = basevertex != nullptr ? basevertex[i] : 0;
+            param.firstInstance = 0;
+        }
+        payload.drawCount = drawcount;
+        payload.pParams = params.data();
+        pVulkanRenderer->MultiDrawElements(payload);
+    }
+
     void MultiDrawElements(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
                            GLsizei drawcount) {
         MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::MultiDrawElements called with null VulkanRenderer");
         MOBILEGL_ASSERT(MG_State::pGLContext, "DirectVulkan::MultiDrawElements called with null GL context");
-
+        MultiDrawElementsImpl(mode, count, type, indices, drawcount, nullptr);
-        // Vector<DrawElementCmd> cmds;
-        // cmds.reserve(static_cast<SizeT>(drawcount));
-        // for (GLsizei i = 0; i < drawcount; ++i) {
-        //     if (count[i] == 0) {
-        //         continue;
-        //     }
-        //
-        //     DrawElementCmd payload{};
-        //     payload.mode = mode;
-        //     payload.firstVertex = 0;
-        //     payload.indexCount = count[i];
-        //     payload.indexType = type;
-        //     payload.indexByteOffset = reinterpret_cast<SizeT>(indices[i]);
-        //     cmds.push_back(payload);
-        // }
-        //
-        // if (cmds.empty()) {
-        //     return;
-        // }
-        // pVulkanRenderer->MultiDrawElements(cmds);
     }
 
     void DrawElementsBaseVertex(GLenum mode, GLsizei count, GLenum type, const GLvoid* indices, GLint basevertex) {
@@ -1444,40 +1044,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     void MultiDrawElementsBaseVertex(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
                                      GLsizei drawcount, const GLint* basevertex) {
-        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::MultiDrawElements called with null VulkanRenderer");
+        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::MultiDrawElementsBaseVertex called with null VulkanRenderer");
-        MOBILEGL_ASSERT(MG_State::pGLContext, "DirectVulkan::MultiDrawElements called with null GL context");
+        MOBILEGL_ASSERT(MG_State::pGLContext, "DirectVulkan::MultiDrawElementsBaseVertex called with null GL context");
-        MultiDrawIndexedCmd payload{};
+        MultiDrawElementsImpl(mode, count, type, indices, drawcount, basevertex);
-        payload.mode = mode;
-        payload.indexBufferView.indexType = type;
-
-        // TODO: allocate draw cmd buf elsewhere
-        static Vector<DrawIndexedCmdParam> params;
-        params.clear();
-        params.resize(drawcount);
-
-        for (GLsizei i = 0; i < drawcount; ++i) {
-            if (count[i] == 0) {
-                continue;
-            }
-
-            // TODO: this index view needs a redesign, now there's a lotta redundant uploads
-
-            payload.indexBufferView.indexByteOffset = 0;
-            payload.indexBufferView.indexByteSize =
-                    std::max(reinterpret_cast<SizeT>(indices[i]) + count[i] * MG_Util::GetGLTypeSize(type),
-                             payload.indexBufferView.indexByteSize);
-
-            auto& param = params[i];
-
-            param.indexCount = count[i];
-            param.instanceCount = 1;
-            param.firstIndex = reinterpret_cast<SizeT>(indices[i]) / MG_Util::GetGLTypeSize(type);
-            param.vertexOffset = basevertex[i];
-            param.firstInstance = 0;
-        }
-        payload.drawCount = drawcount;
-        payload.pParams = params.data();
-        pVulkanRenderer->MultiDrawElements(payload);
     }
 
     void BlitFramebuffer(GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1, GLint dstX0, GLint dstY0, GLint dstX1,

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.h

@@ -97,15 +97,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void GetIntegeri_v(GLenum target, GLuint index, GLint* data);
     void GetInteger64i_v(GLenum target, GLuint index, GLint64* data);
     void GetProgramiv(GLuint program, GLenum pname, GLint* params);
-    void GetProgramInterfaceiv(GLuint program, GLenum programInterface, GLenum pname, GLint* params);
+    void ShaderStorageBlockBinding(GLuint program, const GLchar* storageBlockName, GLuint storageBlockBinding);
-    GLuint GetProgramResourceIndex(GLuint program, GLenum programInterface, const GLchar* name);
-    void GetProgramResourceName(GLuint program, GLenum programInterface, GLuint index, GLsizei bufSize,
-                                GLsizei* length, GLchar* name);
-    void GetProgramResourceiv(GLuint program, GLenum programInterface, GLuint index, GLsizei propCount,
-                              const GLenum* props, GLsizei bufSize, GLsizei* length, GLint* params);
-    GLint GetProgramResourceLocation(GLuint program, GLenum programInterface, const GLchar* name);
-    GLint GetProgramResourceLocationIndex(GLuint program, GLenum programInterface, const GLchar* name);
-    void ShaderStorageBlockBinding(GLuint program, GLuint storageBlockIndex, GLuint storageBlockBinding);
     void ReadPixels(GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels);
     void GetTexImage(GLenum target, GLint level, GLenum format, GLenum type, GLvoid* pixels);
     void GetTextureImage(const SharedPtr<MG_State::GLState::ITextureObject>& texture, TextureUploadTarget uploadTarget,

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -2379,6 +2379,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return it->second;
         }
 
+        // Structural change: the insert below can move every entry of this
+        // open-addressing map, so all memoised entry pointers die here.
+        ++m_cacheStructureEpoch;
         auto& entry = m_cache[hash];
         entry.hash = hash;
         entry.lastUsedFrame = m_frameCounter;
@@ -2580,6 +2583,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // erase runs ~VkProgramObject (modules/layouts destroyed); notify after
                 // so an observer never observes a half-destroyed entry through a lookup.
                 // Observers only need the handle values to purge their keyed caches.
+                ++m_cacheStructureEpoch; // erase moves/kills entries: memoised pointers die
                 it = m_cache.erase(it);
                 if (m_evictionObserver != nullptr) {
                     m_evictionObserver->OnProgramEvicted(hash, descriptorSetLayout);

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -105,8 +105,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // position-invariance quirk (see PipelineFactory::ShouldSuppressDepthWrite).
             Bool fragmentReplacesDepth = false;
             // Frame-boundary counter value of the last GetOrCreateProgram hit; drives
-            // cache eviction (see OnFrameBoundary).
+            // cache eviction (see OnFrameBoundary). Mutable: the draw snapshot's memoised
-            Uint64 lastUsedFrame = 0;
+            // entry pointer re-stamps use through a const reference (StampProgramUse).
+            mutable Uint64 lastUsedFrame = 0;
 
             static inline VkDevice s_device = VK_NULL_HANDLE;
 
@@ -262,6 +263,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkProgramObject& GetOrCreateProgram(
             const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags);
 
+        // Bumped whenever m_cache's STRUCTURE changes (any insert or erase): the cache is
+        // an open-addressing map holding entries by value, so both moves existing entries.
+        // A caller that memoised a VkProgramObject* may keep dereferencing it only while
+        // this is unchanged; on a bump it must re-run GetOrCreateProgram.
+        Uint64 GetCacheStructureEpoch() const { return m_cacheStructureEpoch; }
+        // A memoised entry pointer bypasses GetOrCreateProgram, whose per-lookup stamp is
+        // what keeps an in-use entry out of OnFrameBoundary's idle sweep - so such a
+        // caller must re-stamp the entry itself, at least once per frame boundary.
+        void StampProgramUse(const VkProgramObject& entry) const { entry.lastUsedFrame = m_frameCounter; }
+
         // Observer may be null (no notifications). Not owned.
         void SetEvictionObserver(IEvictionObserver* observer) { m_evictionObserver = observer; }
         // Frame boundary hook: ages the program cache and evicts long-unused entries
@@ -312,6 +323,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         mutable ProgramLookupCache m_lastLookup;
         // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
         Uint64 m_frameCounter = 0;
+        // See GetCacheStructureEpoch(). Starts at 1 so a zero-initialized memo can never match.
+        Uint64 m_cacheStructureEpoch = 1;
         IEvictionObserver* m_evictionObserver = nullptr;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -116,6 +116,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_frameCount = frameCount;
         m_maxBindings = maxBindings;
         m_samplerResolveMemo.assign(m_maxBindings, SamplerResolveMemo{});
+        // Every entry is freshly constructed (all-invalid), so nothing needs sweeping until
+        // a resolve writes one.
+        m_samplerResolveMemoHighWater = 0;
         m_setsPerFrame = setsPerFrame;
         m_peakDescriptorSetsObserved = 0;
         m_textureManager = textureManager;
@@ -174,6 +177,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_minDynamicOffsetAlignment = 1;
         m_frameCount = 0;
         m_maxBindings = 0;
+        m_samplerResolveMemo.clear();
+        m_samplerResolveMemoHighWater = 0;
         m_setsPerFrame = 0;
         m_peakDescriptorSetsObserved = 0;
         m_textureManager = nullptr;
@@ -207,11 +212,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         for (auto& entry : m_descriptorReuseMemo) {
             entry.valid = false;
         }
+        m_fastRebindMemo.valid = false;
         m_lastBindValid = false;
         // Re-fingerprint the bound sampler set fresh this frame so any GL object address
-        // reuse cannot outlive a single frame (see SamplerResolveMemo).
+        // reuse cannot outlive a single frame (see SamplerResolveMemo). Only the entries a
-        for (auto& memo : m_samplerResolveMemo) {
+        // resolve has actually written can be valid, so the high-water mark bounds the
-            memo.valid = false;
+        // sweep - the vector itself is sized to the device's binding cap (256 here), which
+        // is ~30x more entries than any program declares.
+        const Uint32 touchedBindings =
+            std::min<Uint32>(m_samplerResolveMemoHighWater, static_cast<Uint32>(m_samplerResolveMemo.size()));
+        for (Uint32 binding = 0; binding < touchedBindings; ++binding) {
+            m_samplerResolveMemo[binding].valid = false;
+            m_samplerResolveMemo[binding].infoValid = false;
         }
     }
 
@@ -247,6 +259,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             for (auto& entry : m_descriptorReuseMemo) {
                 entry.valid = false;
             }
+            // The rebind memo's set may be among the freed ones.
+            m_fastRebindMemo.valid = false;
             MGLOG_D("UniformDescriptorBinder: freed %zu descriptor sets for destroyed layout", purgedSets);
         }
     }
@@ -254,9 +268,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     Bool UniformManager::ResolveSamplerDescriptor(VkCommandBuffer commandBuffer,
                                                             const MG_State::GLState::ProgramObject& program,
                                                             const ProgramFactory::VkProgramObject& programObj,
-                                                            Uint32 binding, VkDescriptorImageInfo& outImageInfo) const {
+                                                            Uint32 binding, VkDescriptorImageInfo& outImageInfo,
+                                                            Bool trustUnchangedHint) const {
         MOBILEGL_ASSERT(m_textureManager != nullptr, "ResolveSamplerDescriptor: texture manager is null");
         MOBILEGL_ASSERT(m_samplerManager != nullptr, "ResolveSamplerDescriptor: sampler manager is null");
+        // The caller proved every input of this binding's resolution unchanged since the
+        // last full resolve (which also filled the cache), so the whole chain below -
+        // texture/sampler resolution, completeness probe, sync, layout handling, sampler
+        // and view lookups - would recompute the identical descriptor.
+        if (trustUnchangedHint && binding < m_samplerResolveMemo.size() &&
+            m_samplerResolveMemo[binding].infoValid) {
+            outImageInfo = m_samplerResolveMemo[binding].info;
+            return true;
+        }
         MOBILEGL_ASSERT(binding < programObj.samplerNameByBinding.size(),
                         "ResolveSamplerDescriptor: sampler binding %u name lookup out of range", binding);
         // Raw-pointer resolve to skip the SharedPtr atomic refcount churn: the bound texture stays
@@ -365,6 +389,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 viewFormatMemo->viewFormatDomain = numericDomain;
                 viewFormatMemo->viewFormat = sampledViewFormat;
                 viewFormatMemo->viewFormatValid = true;
+                NoteSamplerResolveMemoTouched(binding);
             }
         }
         if (sampledViewFormat == VK_FORMAT_UNDEFINED) {
@@ -420,6 +445,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 memo.viewLevelCount = viewLevelCount;
                 memo.sampler = resolvedSampler;
                 memo.valid = true;
+                NoteSamplerResolveMemoTouched(binding);
             }
         } else {
             resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering,
@@ -430,7 +456,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             .imageView = sampledImageView,
             .imageLayout = resource->layout,
         };
-        return outImageInfo.sampler != VK_NULL_HANDLE;
+        if (outImageInfo.sampler == VK_NULL_HANDLE) {
+            return false;
+        }
+        if (binding < m_samplerResolveMemo.size()) {
+            m_samplerResolveMemo[binding].info = outImageInfo;
+            m_samplerResolveMemo[binding].infoValid = true;
+            NoteSamplerResolveMemoTouched(binding);
+        }
+        return true;
     }
 
     Bool UniformManager::ResolveSamplerDescriptorOverride(
@@ -808,10 +842,55 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return m_fallbackTexture2D;
     }
 
+    Bool UniformManager::ResolveSampledBinding(const MG_State::GLState::ProgramObject& program,
+                                               const ProgramFactory::VkProgramObject& programObj,
+                                               Uint32 binding,
+                                               MG_State::GLState::ITextureObject*& outTexture,
+                                               const MG_State::GLState::SamplerObject*& outSampler) const {
+        // Open-coded ResolveSamplerTextureRaw so the unit is resolved once for both the
+        // texture and the sampler override - this runs per binding per full-path draw,
+        // and program-alternating draw streams take the full path on every draw.
+        MOBILEGL_ASSERT(MG_State::pGLContext != nullptr, "ResolveSampledBinding: GL context is null");
+        MOBILEGL_ASSERT(binding < programObj.samplerUniformLocationByBinding.size(),
+                        "ResolveSampledBinding: sampler location binding %u out of range", binding);
+        MOBILEGL_ASSERT(binding < programObj.samplerTextureTargetByBinding.size(),
+                        "ResolveSampledBinding: sampler target binding %u out of range", binding);
+        const Int location = programObj.samplerUniformLocationByBinding[binding];
+        const Int unit = ResolveSamplerUnitIndex(program, location, binding);
+        auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
+        const TextureTarget preferredTarget = programObj.samplerTextureTargetByBinding[binding];
+        MG_State::GLState::ITextureObject* texture =
+            textureUnit.GetBindingSlot(preferredTarget).GetBoundObject().get();
+        // Undefined default texture (name 0, no image) resolves as "unbound", exactly
+        // like ResolveSamplerTextureRaw reports it.
+        if (MG_State::GLState::IsUndefinedDefaultTexture(texture)) {
+            texture = nullptr;
+        }
+        if (texture == nullptr) {
+            // ResolveSamplerDescriptor will substitute the fallback texture for this binding;
+            // include it in the sampled set so the pre-render-pass sync/transition pass covers
+            // its first use instead of leaving that work to happen inside an active pass.
+            if (preferredTarget != TextureTarget::Texture2D &&
+                preferredTarget != TextureTarget::TextureRectangle) {
+                return false;
+            }
+            texture = GetFallbackTexture(preferredTarget).get();
+        }
+        const auto& samplerOverride = textureUnit.GetSamplerObject();
+        outTexture = texture;
+        outSampler = samplerOverride ? samplerOverride.get()
+                                     : (texture != nullptr ? texture->GetSamplerObject().get() : nullptr);
+        return true;
+    }
+
     Bool UniformManager::CollectSampledTextures(const MG_State::GLState::ProgramObject& program,
                                                           const ProgramFactory::VkProgramObject& programObj,
-                                                          Vector<MG_State::GLState::ITextureObject*>& outTextures) {
+                                                          Vector<MG_State::GLState::ITextureObject*>& outTextures,
+                                                          Vector<SampledBindingRecord>* outBindingRecords) {
         outTextures.clear();
+        if (outBindingRecords != nullptr) {
+            outBindingRecords->clear();
+        }
 
         const Uint32 bindingCount =
             std::min<Uint32>(m_maxBindings, static_cast<Uint32>(programObj.bindingKinds.size()));
@@ -820,17 +899,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 continue;
             }
 
-            MG_State::GLState::ITextureObject* texture = ResolveSamplerTextureRaw(program, programObj, binding);
+            MG_State::GLState::ITextureObject* texture = nullptr;
-            if (!texture) {
+            const MG_State::GLState::SamplerObject* sampler = nullptr;
-                // ResolveSamplerDescriptor will substitute the fallback texture for this binding;
+            if (!ResolveSampledBinding(program, programObj, binding, texture, sampler)) {
-                // include it in the sampled set so the pre-render-pass sync/transition pass covers
+                continue;
-                // its first use instead of leaving that work to happen inside an active pass.
+            }
-                const TextureTarget preferredTarget = programObj.samplerTextureTargetByBinding[binding];
+            if (outBindingRecords != nullptr) {
-                if (preferredTarget != TextureTarget::Texture2D &&
+                outBindingRecords->push_back({texture != nullptr ? texture->GetLifetimeId() : 0,
-                    preferredTarget != TextureTarget::TextureRectangle) {
+                                              sampler != nullptr ? sampler->GetLifetimeId() : 0});
-                    continue;
-                }
-                texture = GetFallbackTexture(preferredTarget).get();
             }
 
             auto found = std::find(outTextures.begin(), outTextures.end(), texture);
@@ -841,6 +917,38 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
+    Bool UniformManager::SampledBindingsUnchanged(const MG_State::GLState::ProgramObject& program,
+                                                  const ProgramFactory::VkProgramObject& programObj,
+                                                  const Vector<SampledBindingRecord>& previousRecords) const {
+        SizeT recordIndex = 0;
+        // Iterate only the bindings this program declares (ascending), exactly like
+        // BindProgramUniformBuffers: this runs per draw whenever the texture bind
+        // generation moved, and walking all m_maxBindings slots to find the 1-8 real
+        // ones dominated it.
+        for (const Uint32 binding : programObj.activeBindings) {
+            if (binding >= m_maxBindings) {
+                break; // ascending, so nothing past the cap can follow
+            }
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            MG_State::GLState::ITextureObject* texture = nullptr;
+            const MG_State::GLState::SamplerObject* sampler = nullptr;
+            if (!ResolveSampledBinding(program, programObj, binding, texture, sampler)) {
+                continue;
+            }
+            if (recordIndex >= previousRecords.size()) {
+                return false;
+            }
+            const SampledBindingRecord& record = previousRecords[recordIndex++];
+            if (record.textureLifetimeId != (texture != nullptr ? texture->GetLifetimeId() : 0) ||
+                record.samplerLifetimeId != (sampler != nullptr ? sampler->GetLifetimeId() : 0)) {
+                return false;
+            }
+        }
+        return recordIndex == previousRecords.size();
+    }
+
     Bool UniformManager::CollectStorageImageTextures(
         const MG_State::GLState::ProgramObject& program,
         const ProgramFactory::VkProgramObject& programObj,
@@ -1155,12 +1263,101 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return VK_SUCCESS;
     }
 
+    Bool UniformManager::ResolveDynamicUboDescriptor(const MG_State::GLState::ProgramObject& program,
+                                                     const ProgramFactory::VkProgramObject& programObj,
+                                                     Uint32 binding, Uint32 arrayElement, Uint32 frameIndex,
+                                                     VkBuffer& outBuffer, VkDeviceSize& outRange,
+                                                     Uint32& outDynamicOffset) {
+        UboBindResult ubo{};
+        const Bool hasPayload = ResolveUniformBufferPayload(program, programObj, binding, arrayElement, ubo);
+        MOBILEGL_ASSERT(hasPayload && (ubo.directBindable || (ubo.payload != nullptr && ubo.payloadSize > 0)),
+                        "UniformDescriptorBinder::ResolveDynamicUboDescriptor failed: missing UBO payload on binding %u element %u",
+                        binding, arrayElement);
+        if (ubo.directBindable) {
+            // Zero-copy: bind the app's resident VkBuffer directly, no per-draw memcpy.
+            outBuffer = ubo.buffer;
+            outRange = ubo.range;
+            outDynamicOffset = static_cast<Uint32>(ubo.dynamicOffset);
+            return true;
+        }
+        // Global-UBO slice reuse (see GlobalUboSliceMemo): unchanged
+        // uniform bytes re-use the slice already uploaded this frame.
+        const Bool isGlobalUbo = programObj.globalUboBinding == static_cast<Int>(binding) && arrayElement == 0;
+        const Uint64 uboFrameSerial = m_bufferManager->GetFrameSerial();
+        const Uint64 uboProgramLifetimeId = program.GetLifetimeId();
+        const Uint32 uboContentVersion = program.GetUBOContentVersion();
+        if (isGlobalUbo) {
+            for (const auto& memo : m_globalUboMemo) {
+                if (memo.buffer != VK_NULL_HANDLE && memo.programLifetimeId == uboProgramLifetimeId &&
+                    memo.frameSerial == uboFrameSerial && memo.uboContentVersion == uboContentVersion &&
+                    memo.range == static_cast<VkDeviceSize>(ubo.payloadSize)) {
+                    outBuffer = memo.buffer;
+                    outRange = memo.range;
+                    outDynamicOffset = static_cast<Uint32>(memo.offset);
+                    return true;
+                }
+            }
+        }
+        BufferSlice slice{};
+        if (!m_bufferManager->UploadTransient(BufferKind::Uniform, frameIndex, ubo.payload, ubo.payloadSize,
+                                              m_minDynamicOffsetAlignment, slice)) {
+            MOBILEGL_ASSERT(false,
+                            "UniformDescriptorBinder::ResolveDynamicUboDescriptor failed: UBO upload failed on binding %u element %u",
+                            binding, arrayElement);
+            return false;
+        }
+        outBuffer = slice.buffer;
+        outRange = ubo.payloadSize;
+        outDynamicOffset = static_cast<Uint32>(slice.offset);
+        if (isGlobalUbo) {
+            m_globalUboMemo[m_globalUboMemoNext] =
+                GlobalUboSliceMemo{uboProgramLifetimeId, uboFrameSerial, uboContentVersion,
+                                   slice.buffer, slice.offset, static_cast<VkDeviceSize>(ubo.payloadSize)};
+            m_globalUboMemoNext = (m_globalUboMemoNext + 1) % kGlobalUboMemoSize;
+        }
+        return true;
+    }
+
+    void UniformManager::BindDescriptorSetDeduped(VkCommandBuffer commandBuffer, VkPipelineBindPoint bindPoint,
+                                                  VkPipelineLayout pipelineLayout, VkDescriptorSet descriptorSet,
+                                                  const Vector<Uint32>& dynamicOffsets) {
+        // Skip the driver call when this exact binding is already live on the
+        // command buffer (see the bind-dedup shadow in the header).
+        const Uint32 offsetCount = static_cast<Uint32>(dynamicOffsets.size());
+        Bool identicalBind = m_lastBindValid && m_lastBindSet == descriptorSet &&
+                             m_lastBindLayout == pipelineLayout && m_lastBindPoint == bindPoint &&
+                             m_lastBindOffsetCount == offsetCount && offsetCount <= kMaxShadowedDynamicOffsets;
+        if (identicalBind) {
+            for (Uint32 i = 0; i < offsetCount; ++i) {
+                if (m_lastBindOffsets[i] != dynamicOffsets[i]) {
+                    identicalBind = false;
+                    break;
+                }
+            }
+        }
+        if (!identicalBind) {
+            vkCmdBindDescriptorSets(commandBuffer, bindPoint, pipelineLayout, 0, 1,
+                                    &descriptorSet, offsetCount, dynamicOffsets.data());
+            if (offsetCount <= kMaxShadowedDynamicOffsets) {
+                m_lastBindValid = true;
+                m_lastBindSet = descriptorSet;
+                m_lastBindLayout = pipelineLayout;
+                m_lastBindPoint = bindPoint;
+                m_lastBindOffsetCount = offsetCount;
+                std::copy_n(dynamicOffsets.data(), offsetCount, m_lastBindOffsets);
+            } else {
+                m_lastBindValid = false;
+            }
+        }
+    }
+
     Bool UniformManager::BindProgramUniformBuffers(VkCommandBuffer commandBuffer,
                                                              const MG_State::GLState::ProgramObject& program,
                                                              const ProgramFactory::VkProgramObject& programObj,
                                                              Uint32 frameIndex,
                                                              VkPipelineBindPoint bindPoint,
-                                                             const SamplerBindingOverride* samplerBindingOverride) {
+                                                             const SamplerBindingOverride* samplerBindingOverride,
+                                                             Bool samplerDescriptorsUnchangedHint) {
         auto& frame = m_frames[frameIndex];
         if (frame.descriptorPools.empty()) {
             MGLOG_E("UniformDescriptorBinder::BindProgramUniformBuffers failed: frame descriptor pools are invalid");
@@ -1170,6 +1367,34 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             frame.activeDescriptorPoolIndex = 0;
         }
 
+        // Dynamic-offset-only rebind (see FastRebindMemo in the header): the last
+        // cacheable walk of this exact program selected a set whose contents are
+        // provably still what this walk would write - the hint covers every
+        // sampler binding, and an unchanged (buffer, range) for the single
+        // dynamic UBO covers the rest - except the dynamic offset, which rebinding
+        // the SAME set delivers without any descriptor write.
+        const Bool cacheable = (samplerBindingOverride == nullptr);
+        if (cacheable && samplerDescriptorsUnchangedHint && m_fastRebindMemo.valid &&
+            m_fastRebindMemo.frameIndex == frameIndex &&
+            m_fastRebindMemo.programLifetimeId == program.GetLifetimeId() &&
+            m_fastRebindMemo.programHash == programObj.hash) {
+            VkBuffer uboBuffer = VK_NULL_HANDLE;
+            VkDeviceSize uboRange = 0;
+            Uint32 uboDynamicOffset = 0;
+            if (ResolveDynamicUboDescriptor(program, programObj, m_fastRebindMemo.uboBinding, 0, frameIndex,
+                                            uboBuffer, uboRange, uboDynamicOffset) &&
+                uboBuffer == m_fastRebindMemo.uboBuffer && uboRange == m_fastRebindMemo.uboRange) {
+                auto& fastOffsets = m_dynamicOffsetsScratch;
+                fastOffsets.clear();
+                fastOffsets.push_back(uboDynamicOffset);
+                BindDescriptorSetDeduped(commandBuffer, bindPoint, programObj.pipelineLayout,
+                                         m_fastRebindMemo.set, fastOffsets);
+                return true;
+            }
+            // Any mismatch (arena wrap or growth, direct-bind retarget, upload
+            // failure) falls through to the full walk, which re-records the memo.
+        }
+
         // The descriptor set is chosen AFTER the writes are built (below), so a draw
         // whose resolved descriptor content matches the previous draw can reuse that
         // set and skip both AcquireDescriptorSet and vkUpdateDescriptorSets.
@@ -1201,6 +1426,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         texelBufferViews.reserve(m_maxBindings);
         dynamicOffsets.reserve(programObj.dynamicBindings.size() + uboArrayExtra);
 
+        // Eligibility probe for FastRebindMemo, filled by this walk: exactly one
+        // dynamic-UBO descriptor (no arrayed elements) and otherwise only
+        // combined-image samplers, so the whole set's content is pinned by the
+        // sampler hint plus one (buffer, range) compare.
+        Uint32 dynamicUboDescriptorCount = 0;
+        Uint32 fastRebindUboBinding = 0;
+        Bool fastRebindKindsEligible = true;
+
         // Iterate only the bindings this program declares. The old walk covered all 256 slots of
         // bindingKinds on every draw to find the 1-8 a real program uses.
         for (const Uint32 binding : programObj.activeBindings) {
@@ -1221,67 +1454,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     binding < programObj.bindingDescriptorCounts.size()
                         ? std::max<Uint32>(1, programObj.bindingDescriptorCounts[binding])
                         : 1u;
+                dynamicUboDescriptorCount += descriptorCount;
+                fastRebindUboBinding = binding;
                 const SizeT firstBufferInfoIndex = bufferInfos.size();
                 for (Uint32 element = 0; element < descriptorCount; ++element) {
-                    UboBindResult ubo{};
-                    const Bool hasPayload =
-                        ResolveUniformBufferPayload(program, programObj, binding, element, ubo);
-                    MOBILEGL_ASSERT(hasPayload && ubo.payload != nullptr && ubo.payloadSize > 0,
-                                    "UniformDescriptorBinder::BindProgramUniformBuffers failed: missing UBO payload on binding %u element %u",
-                                    binding, element);
-
                     VkDescriptorBufferInfo bufferInfo{};
                     // Keep offset 0 (sub-range selected via the dynamic offset) so the hashed bufferInfo
                     // is stable across draws and the descriptor-set reuse cache keeps hitting.
                     bufferInfo.offset = 0;
-                    Uint32 dynOffset;
+                    Uint32 dynOffset = 0;
-                    if (ubo.directBindable) {
+                    if (!ResolveDynamicUboDescriptor(program, programObj, binding, element, frameIndex,
-                        // Zero-copy: bind the app's resident VkBuffer directly, no per-draw memcpy.
+                                                     bufferInfo.buffer, bufferInfo.range, dynOffset)) {
-                        bufferInfo.buffer = ubo.buffer;
+                        return false;
-                        bufferInfo.range = ubo.range;
-                        dynOffset = static_cast<Uint32>(ubo.dynamicOffset);
-                    } else {
-                        // Global-UBO slice reuse (see GlobalUboSliceMemo): unchanged
-                        // uniform bytes re-use the slice already uploaded this frame.
-                        const Bool isGlobalUbo =
-                            programObj.globalUboBinding == static_cast<Int>(binding) && element == 0;
-                        const Uint64 uboFrameSerial = m_bufferManager->GetFrameSerial();
-                        const Uint64 uboProgramLifetimeId = program.GetLifetimeId();
-                        const Uint32 uboContentVersion = program.GetUBOContentVersion();
-                        Bool reusedSlice = false;
-                        if (isGlobalUbo) {
-                            for (const auto& memo : m_globalUboMemo) {
-                                if (memo.buffer != VK_NULL_HANDLE &&
-                                    memo.programLifetimeId == uboProgramLifetimeId &&
-                                    memo.frameSerial == uboFrameSerial &&
-                                    memo.uboContentVersion == uboContentVersion &&
-                                    memo.range == static_cast<VkDeviceSize>(ubo.payloadSize)) {
-                                    bufferInfo.buffer = memo.buffer;
-                                    bufferInfo.range = memo.range;
-                                    dynOffset = static_cast<Uint32>(memo.offset);
-                                    reusedSlice = true;
-                                    break;
-                                }
-                            }
-                        }
-                        if (!reusedSlice) {
-                            BufferSlice slice{};
-                            if (!m_bufferManager->UploadTransient(BufferKind::Uniform, frameIndex, ubo.payload,
-                                                                  ubo.payloadSize, m_minDynamicOffsetAlignment, slice)) {
-                                MOBILEGL_ASSERT(false, "UniformDescriptorBinder::BindProgramUniformBuffers failed: UBO upload failed on binding %u element %u",
-                                        binding, element);
-                                return false;
-                            }
-                            bufferInfo.buffer = slice.buffer;
-                            bufferInfo.range = ubo.payloadSize;
-                            dynOffset = static_cast<Uint32>(slice.offset);
-                            if (isGlobalUbo) {
-                                m_globalUboMemo[m_globalUboMemoNext] = GlobalUboSliceMemo{
-                                    uboProgramLifetimeId, uboFrameSerial, uboContentVersion,
-                                    slice.buffer, slice.offset, static_cast<VkDeviceSize>(ubo.payloadSize)};
-                                m_globalUboMemoNext = (m_globalUboMemoNext + 1) % kGlobalUboMemoSize;
-                            }
-                        }
                     }
                     bufferInfos.push_back(bufferInfo);
                     // Dynamic offsets are consumed in binding order, then array element order,
@@ -1304,6 +1488,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
 
                 texelBufferViews.push_back(bufferView);
+                fastRebindKindsEligible = false;
                 write.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER;
                 write.pTexelBufferView = &texelBufferViews.back();
                 writes.push_back(write);
@@ -1317,6 +1502,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
 
                 bufferInfos.push_back(bufferInfo);
+                fastRebindKindsEligible = false;
                 write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
                 write.pBufferInfo = &bufferInfos.back();
                 writes.push_back(write);
@@ -1329,6 +1515,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     return false;
                 }
                 imageInfos.push_back(imageInfo);
+                fastRebindKindsEligible = false;
                 write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
                 write.pImageInfo = &imageInfos.back();
                 writes.push_back(write);
@@ -1341,7 +1528,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     samplerBindingOverride->sampler != nullptr) {
                     hasImage = ResolveSamplerDescriptorOverride(*samplerBindingOverride, imageInfo);
                 } else {
-                    hasImage = ResolveSamplerDescriptor(commandBuffer, program, programObj, binding, imageInfo);
+                    hasImage = ResolveSamplerDescriptor(commandBuffer, program, programObj, binding, imageInfo,
+                                                        samplerDescriptorsUnchangedHint);
                 }
                 if (!hasImage) {
                     MGLOG_E(
@@ -1372,7 +1560,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // cursor only advances), so its written contents survive; the layout is part of
         // the signature so reuse never crosses programs. Sampler overrides (blits)
         // bypass and invalidate the cache.
-        const Bool cacheable = (samplerBindingOverride == nullptr);
         Uint64 signature = 0xcbf29ce484222325ULL;
         {
             const auto mix64 = [&signature](Uint64 word) {
@@ -1435,34 +1622,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         }
 
-        // Skip the driver call when this exact binding is already live on the
+        // (Re)record the dynamic-offset-only rebind memo. Recording on every
-        // command buffer (see the bind-dedup shadow in the header).
+        // cacheable walk (allocated or reused set alike - both hold exactly the
-        const Uint32 offsetCount = static_cast<Uint32>(dynamicOffsets.size());
+        // content just computed) keeps the single slot tracking the most recent
-        Bool identicalBind = m_lastBindValid && m_lastBindSet == descriptorSet &&
+        // program; a non-cacheable override walk drops it alongside the reuse
-                             m_lastBindLayout == programObj.pipelineLayout && m_lastBindPoint == bindPoint &&
+        // memo above.
-                             m_lastBindOffsetCount == offsetCount && offsetCount <= kMaxShadowedDynamicOffsets;
+        if (cacheable && fastRebindKindsEligible && dynamicUboDescriptorCount == 1) {
-        if (identicalBind) {
+            m_fastRebindMemo = FastRebindMemo{
-            for (Uint32 i = 0; i < offsetCount; ++i) {
+                /*valid=*/true,          frameIndex,      program.GetLifetimeId(), programObj.hash,
-                if (m_lastBindOffsets[i] != dynamicOffsets[i]) {
+                fastRebindUboBinding,    bufferInfos[0].buffer,
-                    identicalBind = false;
+                bufferInfos[0].range,    descriptorSet};
-                    break;
+        } else {
-                }
+            m_fastRebindMemo.valid = false;
-            }
-        }
-        if (!identicalBind) {
-            vkCmdBindDescriptorSets(commandBuffer, bindPoint, programObj.pipelineLayout, 0, 1,
-                                    &descriptorSet, offsetCount, dynamicOffsets.data());
-            if (offsetCount <= kMaxShadowedDynamicOffsets) {
-                m_lastBindValid = true;
-                m_lastBindSet = descriptorSet;
-                m_lastBindLayout = programObj.pipelineLayout;
-                m_lastBindPoint = bindPoint;
-                m_lastBindOffsetCount = offsetCount;
-                std::copy_n(dynamicOffsets.data(), offsetCount, m_lastBindOffsets);
-            } else {
-                m_lastBindValid = false;
-            }
         }
+
+        BindDescriptorSetDeduped(commandBuffer, bindPoint, programObj.pipelineLayout, descriptorSet,
+                                 dynamicOffsets);
         return true;
     }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -53,18 +53,42 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // caches - a live layout's entry must never be purged (its sets would be
         // unreachable pool slots), so there is deliberately no age-based sweep here.
         void OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout);
+        // One record per visited CombinedImageSampler binding (post fallback substitution,
+        // in binding order): the resolved texture and effective sampler, as never-reused
+        // lifetime ids so a freed-and-reallocated object at the same heap address can only
+        // MISS a comparison, never false-hit it (same ABA rule as SamplerResolveMemo).
+        struct SampledBindingRecord {
+            Uint64 textureLifetimeId = 0;
+            Uint64 samplerLifetimeId = 0;
+        };
         Bool CollectSampledTextures(const MG_State::GLState::ProgramObject& program,
                                     const ProgramFactory::VkProgramObject& programObj,
-                                    Vector<MG_State::GLState::ITextureObject*>& outTextures);
+                                    Vector<MG_State::GLState::ITextureObject*>& outTextures,
+                                    Vector<SampledBindingRecord>* outBindingRecords = nullptr);
+        // Shadow-compare for the SetupDraw fast path: re-runs the CollectSampledTextures
+        // walk and reports whether every visited binding still resolves to the recorded
+        // (texture, effective sampler) pair. A texture bind generation bump alone (e.g. a
+        // redundant glBindSampler, which always bumps it) does not prove the sampled set
+        // moved; this walk does, without rebuilding the set or falling off the fast path.
+        Bool SampledBindingsUnchanged(const MG_State::GLState::ProgramObject& program,
+                                      const ProgramFactory::VkProgramObject& programObj,
+                                      const Vector<SampledBindingRecord>& previousRecords) const;
         Bool CollectStorageImageTextures(const MG_State::GLState::ProgramObject& program,
                                          const ProgramFactory::VkProgramObject& programObj,
                                          Vector<MG_State::GLState::ITextureObject*>& outTextures) const;
+        // samplerDescriptorsUnchangedHint: the caller (SetupDraw fast path) proved that
+        // every input of every combined-image-sampler resolution is unchanged since the
+        // previous draw's resolve - same (texture, sampler) per binding, texture params
+        // sum, sampling-resolution generation (sampler params + texture shape), image
+        // epochs AND per-resource layout values - so the per-binding cached
+        // VkDescriptorImageInfo may be reused without re-running the resolve chain.
         Bool BindProgramUniformBuffers(VkCommandBuffer commandBuffer,
                                        const MG_State::GLState::ProgramObject& program,
                                        const ProgramFactory::VkProgramObject& programObj,
                                        Uint32 frameIndex,
                                        VkPipelineBindPoint bindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
-                                       const SamplerBindingOverride* samplerBindingOverride = nullptr);
+                                       const SamplerBindingOverride* samplerBindingOverride = nullptr,
+                                       Bool samplerDescriptorsUnchangedHint = false);
 
         // Pure format-policy helper kept public for host regression tests. Formatted storage
         // images use their shader qualifier; transformed float images use glBindImageTexture's
@@ -114,6 +138,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         static Bool ResolveSamplerTexture(const MG_State::GLState::ProgramObject& program,
                                    const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                    SharedPtr<MG_State::GLState::ITextureObject>& outTexture);
+        // Shared per-binding resolution for CollectSampledTextures and
+        // SampledBindingsUnchanged, so membership and comparison can never diverge:
+        // texture after the fallback substitution (may still be null when no fallback
+        // exists), effective sampler = unit override else the texture's own sampler.
+        // False = the binding is skipped (unbound with a non-2D fallback target).
+        Bool ResolveSampledBinding(const MG_State::GLState::ProgramObject& program,
+                                   const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
+                                   MG_State::GLState::ITextureObject*& outTexture,
+                                   const MG_State::GLState::SamplerObject*& outSampler) const;
         // Raw-pointer variant for the per-draw sampled-texture walk (CollectSampledTextures):
         // the bound texture stays alive through the draw via GL binding state, so callers that
         // only need the pointer skip the SharedPtr copy's atomic refcount churn.
@@ -121,9 +154,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj, Uint32 binding);
         SharedPtr<MG_State::GLState::ITextureObject> GetFallbackTexture(TextureTarget target) const;
+        // trustUnchangedHint: reuse this binding's cached VkDescriptorImageInfo outright
+        // (see BindProgramUniformBuffers' samplerDescriptorsUnchangedHint for the proof
+        // obligations the caller carries).
         Bool ResolveSamplerDescriptor(VkCommandBuffer commandBuffer, const MG_State::GLState::ProgramObject& program,
                                       const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
-                                      VkDescriptorImageInfo& outImageInfo) const;
+                                      VkDescriptorImageInfo& outImageInfo,
+                                      Bool trustUnchangedHint = false) const;
         Bool ResolveSamplerDescriptorOverride(const SamplerBindingOverride& samplerBindingOverride,
                                               VkDescriptorImageInfo& outImageInfo) const;
         Bool ResolveTexelBufferDescriptor(const MG_State::GLState::ProgramObject& program,
@@ -149,6 +186,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool ResolveUniformBufferPayload(const MG_State::GLState::ProgramObject& program,
                                          const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                          Uint32 arrayElement, UboBindResult& out) const;
+        // Shared resolution of one dynamic-UBO binding element into the
+        // (buffer, range, dynamicOffset) triple the descriptor consumes: direct
+        // bind, global-slice reuse, or transient upload. Used by the full walk
+        // and by the dynamic-offset-only rebind (see FastRebindMemo).
+        Bool ResolveDynamicUboDescriptor(const MG_State::GLState::ProgramObject& program,
+                                         const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
+                                         Uint32 arrayElement, Uint32 frameIndex, VkBuffer& outBuffer,
+                                         VkDeviceSize& outRange, Uint32& outDynamicOffset);
+        // The vkCmdBindDescriptorSets tail shared by the full walk and the
+        // dynamic-offset-only rebind: skips the driver call when this exact
+        // binding is already live on the command buffer (see the bind-dedup
+        // shadow below), otherwise binds and refreshes the shadow.
+        void BindDescriptorSetDeduped(VkCommandBuffer commandBuffer, VkPipelineBindPoint bindPoint,
+                                      VkPipelineLayout pipelineLayout, VkDescriptorSet descriptorSet,
+                                      const Vector<Uint32>& dynamicOffsets);
         Bool CreateDescriptorPool(Uint32 maxSets, VkDescriptorPool& outPool) const;
         Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex);
         VkResult AllocateDescriptorSetsFromActivePool(
@@ -196,6 +248,37 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         DescriptorReuseEntry m_descriptorReuseMemo[kDescriptorReuseMemoSize];
         Uint32 m_descriptorReuseMemoNext = 0;
 
+        // Dynamic-offset-only rebind (see BindProgramUniformBuffers): records the
+        // descriptor set selected by the last cacheable full walk of a program
+        // whose active bindings are exactly one dynamic UBO (single descriptor)
+        // plus combined-image samplers. When the next call proves every sampler
+        // descriptor input unchanged (samplerDescriptorsUnchangedHint) and the
+        // UBO re-resolves to the SAME VkBuffer+range - only the dynamic offset
+        // moved, the per-draw glUniform case - the walk collapses to: resolve one
+        // offset, rebind the recorded set with new pDynamicOffsets (Vulkan allows
+        // rebinding the same set with different dynamic offsets).
+        // Invalidation inventory: BeginFrame clears it (the frame's sets are
+        // recycled) and the frameIndex field guards cross-frame confusion on top;
+        // OnDescriptorSetLayoutDestroyed clears it (the set may be freed); a
+        // sampler-override walk clears it (mirrors m_descriptorReuseMemo); a
+        // program relink bumps the backend state version and thus programObj.hash
+        // so the key misses; the program lifetime id is never reused, so a
+        // deleted-and-recreated program misses; a texture/sampler/binding change
+        // drops the hint upstream; an arena wrap or growth resolves a different
+        // VkBuffer and misses. AcquireDescriptorSet's per-frame cursor only
+        // advances, so the recorded set is never re-written within its frame.
+        struct FastRebindMemo {
+            Bool valid = false;
+            Uint32 frameIndex = 0;
+            Uint64 programLifetimeId = 0;
+            ProgramFactory::HashType programHash = 0;
+            Uint32 uboBinding = 0;
+            VkBuffer uboBuffer = VK_NULL_HANDLE;
+            VkDeviceSize uboRange = 0;
+            VkDescriptorSet set = VK_NULL_HANDLE;
+        };
+        FastRebindMemo m_fastRebindMemo;
+
         // vkCmdBindDescriptorSets dedup: consecutive draws with a static uniform
         // block resolve to the same set AND the same dynamic offsets, so the
         // driver call can be skipped outright. Command-buffer-scope state; reset
@@ -253,7 +336,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             SamplerNumericDomain viewFormatDomain = SamplerNumericDomain::Unknown;
             VkFormat viewFormat = VK_FORMAT_UNDEFINED;
             Bool viewFormatValid = false;
+            // Whole resolved descriptor from this binding's last full resolve. Reused
+            // ONLY under ResolveSamplerDescriptor's trustUnchangedHint, whose caller
+            // proves every resolve input unchanged; cleared with the per-frame reset
+            // (the cached VkSampler outlives a frame only via a fresh resolve, which
+            // also re-stamps it against VkSamplerManager's frame-boundary sweep).
+            VkDescriptorImageInfo info{};
+            Bool infoValid = false;
         };
         mutable Vector<SamplerResolveMemo> m_samplerResolveMemo;
+        // Exclusive upper bound on the entries of m_samplerResolveMemo that any resolve
+        // has ever written. The vector is sized to the DEVICE binding cap (256 on desktop
+        // NVIDIA), but a program declares 1-8 bindings, so the per-frame reset below was
+        // memsetting ~22 KB of never-touched entries every frame - a measurable slice of
+        // the per-frame fixed cost on draw-light frames. Every site that can turn any of
+        // an entry's *Valid flags on raises this mark first, so entries at or above it are
+        // provably still in their constructed (all-invalid) state and clearing them is a
+        // no-op. Never lowered except by Initialize/Shutdown, which rebuild the vector.
+        mutable Uint32 m_samplerResolveMemoHighWater = 0;
+        void NoteSamplerResolveMemoTouched(Uint32 binding) const {
+            if (binding >= m_samplerResolveMemoHighWater) {
+                m_samplerResolveMemoHighWater = binding + 1;
+            }
+        }
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp

@@ -33,14 +33,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             XXHASH_VERIFY(XXH64_update(m_hashState, &attr.IsBgra, sizeof(attr.IsBgra)));
             XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Divisor, sizeof(attr.Divisor)));
 
-            // The buffer's heap address is an identity component of the key: a freed
+            // The bound buffer's IDENTITY is a component of the key, and it has to be the
-            // buffer's reused address can alias an old cache entry, but only under a
+            // buffer's never-reused lifetime id - NOT its heap address, which this used to
-            // byte-identical attribute layout - and the entry payload is a pure function
+            // hash. An address is recycled by the allocator, so a deleted-and-recreated
-            // of the hashed inputs, with the draw path re-resolving bindingBufferKeys
+            // buffer reproduces it; combined with a byte-identical attribute layout that
-            // against the live VAO attribute pointers, so an aliased hit returns exactly
+            // reproduces the WHOLE content hash, and the hash is what
-            // what a rebuild would. Address drift only grows the map; the OnFrameBoundary
+            // TryBindResolvedVertexBindings accepts as proof that a memoised binding still
-            // aging sweep bounds that.
+            // reads the buffer it was resolved from. It did not: a destroyed buffer's GPU
-            const SizeT bufferKey = reinterpret_cast<SizeT>(attr.Buffer.get());
+            // slice was bound for its successor's draw, which is how a transform-feedback
+            // capture came back holding a dead VAO's vertex data (0,0,0,1 - the previous
+            // test's positions) instead of its own.
+            // Zero for client memory (no buffer), which is a distinct identity of its own.
+            const Uint64 bufferKey = attr.Buffer ? attr.Buffer->GetLifetimeId() : 0;
             XXHASH_VERIFY(XXH64_update(m_hashState, &bufferKey, sizeof(bufferKey)));
         }
 
@@ -71,6 +75,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         const BackendVertexInputState& entry = GetOrCreateVertexInputState(vao, GetOrComputeHash(vao));
         vao.SetBackendStateMemo(&entry, m_evictionEpoch);
+        // Also mirror the layout identity and the two per-draw masks into the VAO's aux
+        // memo (pure VALUES derived from the VAO configuration, so config-version
+        // guarding alone is sound). The draw fast path reads them from the VAO object it
+        // already touched instead of chasing into this entry - see PackVertexInputAuxMemo.
+        vao.SetBackendAuxMemo(entry.layoutHash,
+                              PackVertexInputAuxMasks(entry.unsupportedAttribMask, entry.attributeLocationMask));
         return entry;
     }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -28,11 +28,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         struct BackendVertexInputState {
             HashType hash = 0;
             // Hash of the resolved Vulkan vertex layout only (bindings, attributes,
-            // unsupported mask) - NO buffer identities. `hash` mixes buffer heap
+            // unsupported mask) - NO buffer identities. `hash` mixes each bound
-            // addresses so per-chunk VBOs mint a fresh identity per buffer; keying
+            // buffer's never-reused LIFETIME ID, so per-chunk VBOs mint a fresh
-            // pipelines on that minted one VkPipeline per chunk section for an
+            // identity per buffer; keying pipelines on that minted one VkPipeline per
-            // identical layout, defeating pipeline reuse and the per-draw memo.
+            // chunk section for an identical layout, defeating pipeline reuse and the
-            // Pipelines depend only on the layout, so they key on this instead.
+            // per-draw memo. Pipelines depend only on the layout, so they key on this
+            // instead.
             HashType layoutHash = 0;
             // Frame boundary of the last cache hit; entries idle past the
             // OnFrameBoundary retirement age are evicted (CPU heap only).
@@ -71,6 +72,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ~VertexInputStateFactory() = default;
         VertexInputStateFactory(const VertexInputStateFactory&) = delete;
 
+        // The VAO aux-memo payload GetOrCreateVertexInputState(vao) stamps: aux0 is the
+        // entry's layoutHash, aux1 packs (unsupportedAttribMask << 32) | attributeLocationMask.
+        // Readers that find the aux memo valid can use these without resolving the entry.
+        static Uint64 PackVertexInputAuxMasks(Uint32 unsupportedAttribMask, Uint32 attributeLocationMask) {
+            return (static_cast<Uint64>(unsupportedAttribMask) << 32) | attributeLocationMask;
+        }
+
         HashType ComputeHash(const MG_State::GLState::VertexArrayObject& vao) const;
         // Memoized ComputeHash: reuses the VAO's cached hash while its config version
         // is unchanged. Use this on per-draw paths.
@@ -79,8 +87,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const MG_State::GLState::VertexArrayObject& vao, HashType hash);
         const BackendVertexInputState& GetOrCreateVertexInputState(const MG_State::GLState::VertexArrayObject& vao);
         // Frame boundary hook: ages the cache and evicts entries not hit for many
-        // frames. The key mixes buffer heap addresses, so buffer/VAO churn keeps
+        // frames. The key mixes each bound buffer's never-reused lifetime id, so
-        // minting fresh keys; without eviction the map grows for the whole session.
+        // buffer/VAO churn keeps minting fresh keys - and does so by construction,
+        // not by luck: a recreated buffer can no longer land back on its dead
+        // predecessor's key. Without eviction the map grows for the whole session.
         // Entries hold no Vulkan handles (pipeline creation copies the descriptions)
         // and the draw path's entry reference never spans a frame boundary, so
         // eviction here needs no GPU-idle proof. Self-gated: one counter bump and

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.cpp

@@ -258,6 +258,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void VkBufferManager::ReleaseAllLiveResources() {
         for (auto& weak : m_liveResources) {
             if (auto resource = weak.lock()) {
+                BumpSliceEpoch(*resource);
                 resource->buffer.Destroy();
                 resource->storageSize = 0;
                 resource->usageFlags = 0;
@@ -272,6 +273,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool VkBufferManager::CreateResidentStorage(VkBufferResource& resource, VkDeviceSize size,
                                                 VkBufferUsageFlags usage, VkMemoryPropertyFlags requiredFlags) {
+        // The only place a resident VkBuffer handle is minted, so every resident slice
+        // change funnels through here (callers release the old handle first).
+        BumpSliceEpoch(resource);
         // Staged range copies write resident storage with vkCmdCopyBuffer.
         usage |= VK_BUFFER_USAGE_TRANSFER_DST_BIT;
         const Bool created = resource.buffer.Create({
@@ -358,6 +362,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (!resource) {
             return; // lazy: AcquireResidentSlice performs a full upload on creation
         }
+        // A respecify can change the size, the usage hint (so the resident/streamed
+        // route), and the contents at once; retire every memo before deciding what to
+        // do about the storage.
+        BumpSliceEpoch(*resource);
         // Any cached streaming slice refers to the previous contents.
         resource->transientFrameSerial = 0;
         if (!resource->buffer.IsValid()) {
@@ -390,6 +398,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (!resource) {
             return;
         }
+        // Drops the streaming memo below and may end in a storage swap or a deferred
+        // full re-upload, so no memoised slice survives this.
+        BumpSliceEpoch(*resource);
         resource->transientFrameSerial = 0;
         if (!resource->buffer.IsValid() || resource->pendingFullUpload) {
             return;
@@ -422,6 +433,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (!resource) {
             return;
         }
+        BumpSliceEpoch(*resource);
         resource->transientFrameSerial = 0;
         if (!resource->buffer.IsValid() || resource->pendingFullUpload) {
             return;
@@ -481,6 +493,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             TrackLiveResource(resource);
         }
 
+        // Bumped for the request, not just for the storage it may create. This is the
+        // one call the frontend makes when a buffer becomes persistently mapped for
+        // writing (BufferObject::AcquireMemoryRange), and a map the backend declines
+        // keeps mutating its shadow with no further API call - so it is what lets
+        // GetSliceEpochCounter stand for "no buffer needs a persistent-map range push".
+        BumpSliceEpoch(*resource);
+
         // Idempotent: an already-backed buffer returns the same mapped base.
         if (resource->persistentMapped && resource->buffer.IsValid() && resource->storageSize == size) {
             return resource->buffer.GetMappedData();
@@ -608,8 +627,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         } else if (resource->transientChangeSerial == changeSerial && resource->transientSize == size &&
                    resource->transientFrameSerial != 0) {
             if (++resource->unchangedStreak >= kStreamedPromotionStreak) {
+                // Promotion moves the buffer off the arena and onto resident storage.
                 resource->promotedResident = true;
                 resource->promotedChangeSerial = changeSerial;
+                BumpSliceEpoch(*resource);
                 if (AcquireResidentSlice(kind, bufferObject, outSlice)) {
                     return true;
                 }
@@ -619,6 +640,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource->unchangedStreak = 0;
         }
 
+        // A fresh arena allocation: a different slice than the last call handed back,
+        // and (below) the point where a promoted buffer's resident storage is released.
+        // The stable-promotion exit above returns before this, so a buffer the app has
+        // stopped touching keeps one slice for as long as it keeps its resident storage.
+        BumpSliceEpoch(*resource);
         if (!m_transientUploadArena.Upload(m_currentFrameIndex, bufferObject->MappedData(), size, 16,
                                            outSlice)) {
             return false;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.h

@@ -57,6 +57,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // never orphaned or recreated. Draw-time acquire binds it directly, no re-upload.
         Bool persistentMapped = false;
 
+        // Bumped from a manager-wide counter every time anything that decides which
+        // BufferSlice an Acquire*Slice call hands back changes: storage created or
+        // released, a full re-upload becoming due, a promotion/demotion between
+        // resident and streamed storage, or a new per-frame arena slice. Callers that
+        // memoise a resolved slice compare this to prove the memo still describes the
+        // buffer. The counter is manager-wide (never per-resource) so a freshly
+        // created resource - including one that replaces a destroyed resource at the
+        // same address - can never reproduce a value some memo already holds. 0 means
+        // "no slice has ever been handed out", which no memo can match.
+        Uint64 sliceEpoch = 0;
+
         // Cached transient (streaming) slice for the current frame.
         BufferSlice transientSlice{};
         Uint64 transientFrameSerial = 0;
@@ -132,6 +143,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void OnResourceDestroyed(SharedPtr<MG_State::GLState::BackendBufferResource>&& resource);
 
         Uint64 GetFrameSerial() const { return m_frameSerial; }
+        // Highest value handed to any VkBufferResource::sliceEpoch. Unchanged since a
+        // memo was taken means no buffer this manager owns changed which slice it hands
+        // back, and none was persistently mapped, in between - so a memo of resolved
+        // slices needs no per-buffer re-check. See AcquirePersistentMap for the mapping half.
+        Uint64 GetSliceEpochCounter() const { return m_sliceEpochCounter; }
         // Highest frame serial whose GPU work is known complete; serials at or
         // below it may be considered signaled. Drives IsResourceBusy and the
         // backend GL fence objects.
@@ -158,6 +174,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void DestroyAllDeferredReleases();
         void TrackLiveResource(const SharedPtr<VkBufferResource>& resource);
         void ReleaseAllLiveResources();
+        // See VkBufferResource::sliceEpoch.
+        void BumpSliceEpoch(VkBufferResource& resource) { resource.sliceEpoch = ++m_sliceEpochCounter; }
 
         VkBufferManagerInitInfo m_initInfo{};
         BufferArena m_transientUploadArena;
@@ -170,5 +188,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint32 m_currentFrameIndex = 0;
         Uint64 m_frameSerial = 1;
         Uint64 m_completedSerialFloor = 0;
+        // Never reset (not even by Shutdown): a value handed to a resource must stay
+        // unique for the process, or a memo taken before a re-initialize could match
+        // a different resource's state after it.
+        Uint64 m_sliceEpochCounter = 0;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferObject.cpp

@@ -176,16 +176,4 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
-    BufferSlice VkBufferObject::GetSlice(VkDeviceSize offset, VkDeviceSize size) const {
-        MOBILEGL_ASSERT(offset <= m_size, "VkBufferObject::GetSlice offset out of range");
-        const VkDeviceSize resolvedSize = (size == VK_WHOLE_SIZE) ? (m_size - offset) : size;
-        MOBILEGL_ASSERT(offset + resolvedSize <= m_size, "VkBufferObject::GetSlice range out of bounds");
-
-        BufferSlice slice{};
-        slice.buffer = m_buffer;
-        slice.offset = offset;
-        slice.size = resolvedSize;
-        slice.mapped = (m_mappedData != nullptr) ? static_cast<Uint8*>(m_mappedData) + offset : nullptr;
-        return slice;
-    }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferObject.h

@@ -48,7 +48,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkBuffer GetHandle() const { return m_buffer; }
         VkDeviceSize GetSize() const { return m_size; }
-        BufferSlice GetSlice(VkDeviceSize offset = 0, VkDeviceSize size = VK_WHOLE_SIZE) const;
+        // Inline: runs on the per-draw acquire path (a resident buffer bind is a
+        // GetSlice per binding), where an out-of-line call was measurable.
+        BufferSlice GetSlice(VkDeviceSize offset = 0, VkDeviceSize size = VK_WHOLE_SIZE) const {
+            MOBILEGL_ASSERT(offset <= m_size, "VkBufferObject::GetSlice offset out of range");
+            const VkDeviceSize resolvedSize = (size == VK_WHOLE_SIZE) ? (m_size - offset) : size;
+            MOBILEGL_ASSERT(offset + resolvedSize <= m_size, "VkBufferObject::GetSlice range out of bounds");
+
+            BufferSlice slice{};
+            slice.buffer = m_buffer;
+            slice.offset = offset;
+            slice.size = resolvedSize;
+            slice.mapped = (m_mappedData != nullptr) ? static_cast<Uint8*>(m_mappedData) + offset : nullptr;
+            return slice;
+        }
         void* GetMappedData() const { return m_mappedData; }
         Bool IsMapped() const { return m_mappedData != nullptr; }
         Bool IsValid() const { return m_allocator != nullptr && m_buffer != VK_NULL_HANDLE && m_allocation != nullptr; }

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -15,6 +15,7 @@
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
 
 #include <Config.h>
+#include <algorithm>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
@@ -620,12 +621,34 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         TextureResource::s_device = m_device;
         TextureResource::s_allocator = m_allocator;
 
+        // Own pool for the recycled upload-batch command buffers. Parking a
+        // dozen reset-but-alive command buffers in the renderer's shared pool
+        // interleaves their retained chunks with the frame command buffers
+        // allocated/freed there every frame; isolating them keeps both pools'
+        // internal allocators dense.
+        VkCommandPoolCreateInfo uploadPoolInfo{VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO};
+        uploadPoolInfo.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT |
+                               VK_COMMAND_POOL_CREATE_TRANSIENT_BIT;
+        uploadPoolInfo.queueFamilyIndex = initInfo.graphicsQueueFamilyIndex;
+        VK_VERIFY(vkCreateCommandPool(m_device, &uploadPoolInfo, nullptr, &m_uploadCommandPool),
+                  "vkCreateCommandPool(texture upload batch)");
+
         return true;
     }
 
     void VkTextureManager::Shutdown() {
         if (m_device != VK_NULL_HANDLE) {
+            // A still-open (never-submitted) batch is discarded, not submitted:
+            // the renderer has already drained the device and the data has no
+            // observer. Submitted batches are waited and recycled, then the
+            // pools they recycled into are destroyed.
+            DiscardPendingUploadBatch();
             ReclaimCompletedUploads(/*waitAll=*/true);
+            DestroyUploadPools();
+            if (m_uploadCommandPool != VK_NULL_HANDLE) {
+                vkDestroyCommandPool(m_device, m_uploadCommandPool, nullptr);
+                m_uploadCommandPool = VK_NULL_HANDLE;
+            }
         }
         DestroyDeferredReleases();
         ++m_resourceEraseEpoch;  // every memoized resource pointer dies with the map
@@ -1859,6 +1882,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         resource.syncedTextureParamsVersion = 0;
 
         if (preservedResource) {
+            // The preserve copy reads the OLD image on its own immediately-
+            // submitted-and-waited command buffer; a batched upload into that
+            // image still sitting in the open batch must reach the queue first
+            // or the copy carries pre-upload texels forward.
+            FlushPendingUploads();
             const Bool preserved = PreserveTextureContentsOnRecreate(
                 m_device, m_commandPool, m_graphicsQueue, *preservedResource, resource);
             MOBILEGL_ASSERT(preserved,
@@ -1869,6 +1897,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     void VkTextureManager::DeferResourceRelease(TextureResource&& resource) {
+        // The deferred-release queues are drained under fence/queue-idle proofs
+        // that only cover SUBMITTED work; a recorded-but-unsubmitted upload
+        // batch referencing this image would escape them. Push the batch onto
+        // the queue first so every later proof covers it. Rare (only recreate/
+        // erase of an image uploaded this very frame), so the flush is cheap.
+        if (m_uploadBatchOpen && resource.image != VK_NULL_HANDLE &&
+            std::find(m_uploadBatchImages.begin(), m_uploadBatchImages.end(), resource.image) !=
+                m_uploadBatchImages.end()) {
+            FlushPendingUploads();
+        }
         if (resource.image == VK_NULL_HANDLE && resource.fullView == VK_NULL_HANDLE &&
             resource.sampledView == VK_NULL_HANDLE &&
             resource.perMipViews.empty() && resource.perMipSampledViews.empty() &&
@@ -1919,14 +1957,211 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             } else if (vkGetFenceStatus(m_device, entry.fence) != VK_SUCCESS) {
                 break;
             }
-            vkDestroyFence(m_device, entry.fence, nullptr);
+            // Recycle, don't destroy: the fence resets into the fence pool,
-            vkFreeCommandBuffers(m_device, m_commandPool, 1, &entry.commandBuffer);
+            // the command buffer resets into the CB pool (m_uploadCommandPool
-            vmaDestroyBuffer(m_allocator, entry.stagingBuffer, entry.stagingAllocation);
+            // carries RESET_COMMAND_BUFFER_BIT), and the staging blocks
+            // return to the block pool for the next batch to bump-allocate.
+            // This is where the mc_tex_stream win comes from: the per-upload
+            // fence create/destroy + command-buffer alloc/free ioctl traffic
+            // was the measured 41%-in-kernel cost, not the submit itself.
+            if (vkResetFences(m_device, 1, &entry.fence) == VK_SUCCESS) {
+                m_freeUploadFences.push_back(entry.fence);
+            } else {
+                vkDestroyFence(m_device, entry.fence, nullptr);
+            }
+            if (vkResetCommandBuffer(entry.commandBuffer, 0) == VK_SUCCESS) {
+                m_freeUploadCommandBuffers.push_back(entry.commandBuffer);
+            } else {
+                vkFreeCommandBuffers(m_device, m_uploadCommandPool, 1, &entry.commandBuffer);
+            }
+            for (auto& block : entry.stagingBlocks) {
+                RecycleUploadStagingBlock(Move(block));
+            }
+            entry.stagingBlocks.clear();
         }
         m_pendingUploadReclaims.erase(m_pendingUploadReclaims.begin(),
                                       m_pendingUploadReclaims.begin() + static_cast<std::ptrdiff_t>(completed));
     }
 
+    void VkTextureManager::RecycleUploadStagingBlock(UploadStagingBlock&& block) {
+        if (block.buffer == VK_NULL_HANDLE) {
+            return;
+        }
+        // Bound the idle pool: a one-off giant upload (initial atlas define)
+        // must not pin its staging memory forever.
+        constexpr VkDeviceSize kMaxFreeUploadStagingBytes = 32u * 1024u * 1024u;
+        if (m_allocator == nullptr || m_freeUploadStagingBytes + block.capacity > kMaxFreeUploadStagingBytes) {
+            vmaDestroyBuffer(m_allocator, block.buffer, block.allocation);
+            return;
+        }
+        block.cursor = 0;
+        m_freeUploadStagingBytes += block.capacity;
+        m_freeUploadStagingBlocks.push_back(Move(block));
+    }
+
+    VkCommandBuffer VkTextureManager::EnsureUploadBatchOpen() {
+        if (m_uploadBatchOpen) {
+            return m_uploadBatchCommandBuffer;
+        }
+        if (!m_freeUploadCommandBuffers.empty()) {
+            m_uploadBatchCommandBuffer = m_freeUploadCommandBuffers.back();
+            m_freeUploadCommandBuffers.pop_back();
+        } else {
+            VkCommandBufferAllocateInfo allocInfo{};
+            allocInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
+            allocInfo.commandPool = m_uploadCommandPool;
+            allocInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
+            allocInfo.commandBufferCount = 1;
+            VK_VERIFY(vkAllocateCommandBuffers(m_device, &allocInfo, &m_uploadBatchCommandBuffer),
+                      "vkAllocateCommandBuffers(texture upload batch)");
+        }
+        VkCommandBufferBeginInfo beginInfo{};
+        beginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
+        beginInfo.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
+        VK_VERIFY(vkBeginCommandBuffer(m_uploadBatchCommandBuffer, &beginInfo),
+                  "vkBeginCommandBuffer(texture upload batch)");
+        m_uploadBatchOpen = true;
+        return m_uploadBatchCommandBuffer;
+    }
+
+    Uint8* VkTextureManager::AcquireUploadStagingSpace(VkDeviceSize size, VkBuffer& outBuffer,
+                                                       VkDeviceSize& outBaseOffset) {
+        // 16 covers every uncompressed texel size in use (1..16 bytes) and the
+        // bufferOffset multiple-of-4 rule; per-item offsets inside the span
+        // keep the pre-batching tight packing.
+        constexpr VkDeviceSize kUploadStagingAlignment = 16;
+        constexpr VkDeviceSize kUploadStagingBlockSize = 1u * 1024u * 1024u;
+        UploadStagingBlock* current = m_uploadBatchBlocks.empty() ? nullptr : &m_uploadBatchBlocks.back();
+        VkDeviceSize alignedCursor = 0;
+        if (current != nullptr) {
+            alignedCursor = (current->cursor + (kUploadStagingAlignment - 1)) & ~(kUploadStagingAlignment - 1);
+            if (alignedCursor + size > current->capacity) {
+                current = nullptr;
+            }
+        }
+        if (current == nullptr) {
+            UploadStagingBlock block;
+            for (SizeT i = 0; i < m_freeUploadStagingBlocks.size(); ++i) {
+                if (m_freeUploadStagingBlocks[i].capacity >= size) {
+                    block = Move(m_freeUploadStagingBlocks[i]);
+                    m_freeUploadStagingBytes -= block.capacity;
+                    m_freeUploadStagingBlocks.erase(m_freeUploadStagingBlocks.begin() +
+                                                    static_cast<std::ptrdiff_t>(i));
+                    break;
+                }
+            }
+            if (block.buffer == VK_NULL_HANDLE) {
+                VkBufferCreateInfo bufferInfo{};
+                bufferInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
+                bufferInfo.size = std::max(kUploadStagingBlockSize, size);
+                bufferInfo.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
+                bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+                VmaAllocationCreateInfo stagingAllocationInfo{};
+                stagingAllocationInfo.usage = VMA_MEMORY_USAGE_AUTO_PREFER_HOST;
+                stagingAllocationInfo.flags = VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT |
+                                              VMA_ALLOCATION_CREATE_MAPPED_BIT;
+                stagingAllocationInfo.requiredFlags =
+                    VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
+                VmaAllocationInfo allocationResult{};
+                VK_VERIFY(vmaCreateBuffer(m_allocator, &bufferInfo, &stagingAllocationInfo, &block.buffer,
+                                          &block.allocation, &allocationResult),
+                          "vmaCreateBuffer(texture upload staging block)");
+                block.mapped = static_cast<Uint8*>(allocationResult.pMappedData);
+                block.capacity = bufferInfo.size;
+                MOBILEGL_ASSERT(block.mapped != nullptr,
+                                "AcquireUploadStagingSpace: staging block is not persistently mapped");
+            }
+            block.cursor = 0;
+            m_uploadBatchBlocks.push_back(Move(block));
+            current = &m_uploadBatchBlocks.back();
+            alignedCursor = 0;
+        }
+        outBuffer = current->buffer;
+        outBaseOffset = alignedCursor;
+        current->cursor = alignedCursor + size;
+        return current->mapped + alignedCursor;
+    }
+
+    void VkTextureManager::FlushPendingUploads() {
+        if (!m_uploadBatchOpen) {
+            return;
+        }
+        VK_VERIFY(vkEndCommandBuffer(m_uploadBatchCommandBuffer), "vkEndCommandBuffer(texture upload batch)");
+
+        VkFence uploadFence = VK_NULL_HANDLE;
+        if (!m_freeUploadFences.empty()) {
+            uploadFence = m_freeUploadFences.back();
+            m_freeUploadFences.pop_back();
+        } else {
+            VkFenceCreateInfo fenceInfo{};
+            fenceInfo.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
+            VK_VERIFY(vkCreateFence(m_device, &fenceInfo, nullptr, &uploadFence), "vkCreateFence(texture upload)");
+        }
+
+        VkSubmitInfo submitInfo{};
+        submitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
+        submitInfo.commandBufferCount = 1;
+        submitInfo.pCommandBuffers = &m_uploadBatchCommandBuffer;
+        VK_VERIFY(vkQueueSubmit(m_graphicsQueue, 1, &submitInfo, uploadFence), "vkQueueSubmit(texture upload batch)");
+
+        PendingUploadReclaim reclaim;
+        reclaim.fence = uploadFence;
+        reclaim.commandBuffer = m_uploadBatchCommandBuffer;
+        reclaim.stagingBlocks = Move(m_uploadBatchBlocks);
+        m_pendingUploadReclaims.push_back(Move(reclaim));
+        m_uploadBatchCommandBuffer = VK_NULL_HANDLE;
+        m_uploadBatchOpen = false;
+        m_uploadBatchBlocks.clear();
+        m_uploadBatchImages.clear();
+        m_uploadBatchStagingBytes = 0;
+
+        ReclaimCompletedUploads();
+        // Backstop for pathological upload storms: bound in-flight staging
+        // memory by blocking on the oldest batch only once the list is deep.
+        constexpr SizeT kMaxPendingTextureUploads = 16;
+        if (m_pendingUploadReclaims.size() > kMaxPendingTextureUploads) {
+            VK_VERIFY(vkWaitForFences(m_device, 1, &m_pendingUploadReclaims.front().fence, VK_TRUE, UINT64_MAX),
+                      "vkWaitForFences(texture upload backstop)");
+            ReclaimCompletedUploads();
+        }
+    }
+
+    void VkTextureManager::DiscardPendingUploadBatch() {
+        if (!m_uploadBatchOpen) {
+            return;
+        }
+        // The batch was never submitted, so the command buffer is in the
+        // recording state, not pending - freeing it is legal.
+        vkFreeCommandBuffers(m_device, m_uploadCommandPool, 1, &m_uploadBatchCommandBuffer);
+        m_uploadBatchCommandBuffer = VK_NULL_HANDLE;
+        m_uploadBatchOpen = false;
+        for (auto& block : m_uploadBatchBlocks) {
+            RecycleUploadStagingBlock(Move(block));
+        }
+        m_uploadBatchBlocks.clear();
+        m_uploadBatchImages.clear();
+        m_uploadBatchStagingBytes = 0;
+    }
+
+    void VkTextureManager::DestroyUploadPools() {
+        for (auto& block : m_freeUploadStagingBlocks) {
+            if (block.buffer != VK_NULL_HANDLE) {
+                vmaDestroyBuffer(m_allocator, block.buffer, block.allocation);
+            }
+        }
+        m_freeUploadStagingBlocks.clear();
+        m_freeUploadStagingBytes = 0;
+        if (!m_freeUploadCommandBuffers.empty()) {
+            vkFreeCommandBuffers(m_device, m_uploadCommandPool, static_cast<Uint32>(m_freeUploadCommandBuffers.size()),
+                                 m_freeUploadCommandBuffers.data());
+            m_freeUploadCommandBuffers.clear();
+        }
+        for (const VkFence fence : m_freeUploadFences) {
+            vkDestroyFence(m_device, fence, nullptr);
+        }
+        m_freeUploadFences.clear();
+    }
+
     void VkTextureManager::DestroyDeferredReleases() {
         for (auto& deferredReleases : m_deferredReleases) {
             deferredReleases.clear();
@@ -2070,6 +2305,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             IntVec3 regionLo = {0, 0, 0};
             IntVec3 regionSize = {0, 0, 0};
             SizeT texelBytes = 0;
+            // Scatter refinement of the single dirty box: when the storage's rect
+            // list reports the writes' true footprint (~100 sprites whose union box
+            // spans the whole atlas), each rect is staged tightly and copied with
+            // its own VkBufferImageCopy in ONE vkCmdCopyBufferToImage. Empty means
+            // "stage the one box above". Only set while subRegion.
+            Vector<MG_State::GLState::MipmapDirtyRegion> rects;
         };
 
         Vector<UploadItem> uploadItems;
@@ -2133,6 +2374,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                     static_cast<SizeT>(uploadItem.regionSize.y()) *
                                                     static_cast<SizeT>(uploadItem.regionSize.z()) *
                                                     uploadItem.texelBytes;
+                        // Scatter refinement: the storage only hands out its rect list
+                        // when the rects' summed area is materially smaller than the
+                        // union box (0 otherwise), so taking it always stages fewer
+                        // bytes than the box - the very amplification this path exists
+                        // to avoid paying twice.
+                        MG_State::GLState::MipmapDirtyRegion
+                            dirtyRects[MG_State::GLState::MipmapStorage::kMaxDirtyRects];
+                        const SizeT dirtyRectCount = mipmapTexture.GetStorageDirtyRects(
+                            target, level, dirtyRects, MG_State::GLState::MipmapStorage::kMaxDirtyRects);
+                        if (dirtyRectCount >= 2) {
+                            uploadItem.rects.assign(dirtyRects, dirtyRects + dirtyRectCount);
+                            SizeT rectTexels = 0;
+                            for (const auto& rect : uploadItem.rects) {
+                                rectTexels += rect.TexelCount();
+                            }
+                            uploadItem.uploadByteSize = rectTexels * uploadItem.texelBytes;
+                        }
                     }
                 }
                 if (formatInfo.expandRgbToRgba) {
@@ -2271,61 +2529,66 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         }
 
+        // Rare mid-frame hazard, kept at parity with the old per-upload
+        // submits: this image already has an upload recorded in the OPEN batch
+        // and has since been referenced by the frame's open recording (drawn).
+        // Appending here would merge both uploads into the same pre-frame
+        // submission the old code split into two; flush first so the second
+        // upload lands in its own later submission, exactly like before.
+        if (m_uploadBatchOpen && WasTouchedThisRecording(outResource) &&
+            std::find(m_uploadBatchImages.begin(), m_uploadBatchImages.end(), outResource.image) !=
+                m_uploadBatchImages.end()) {
+            FlushPendingUploads();
+        }
+        // Bound the staging bytes a single batch can pin before its fence can
+        // reclaim them.
+        constexpr VkDeviceSize kMaxBatchStagingBytes = 64u * 1024u * 1024u;
+        if (m_uploadBatchOpen && m_uploadBatchStagingBytes + stagingSize > kMaxBatchStagingBytes) {
+            FlushPendingUploads();
+        }
+
+        VkCommandBuffer commandBuffer = EnsureUploadBatchOpen();
         VkBuffer stagingBuffer = VK_NULL_HANDLE;
-        VmaAllocation stagingAllocation = nullptr;
+        VkDeviceSize stagingBase = 0;
-
+        Uint8* mapped = AcquireUploadStagingSpace(stagingSize, stagingBuffer, stagingBase);
-        VkBufferCreateInfo bufferInfo{};
-        bufferInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
-        bufferInfo.size = stagingSize;
-        bufferInfo.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
-        bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
-        VmaAllocationCreateInfo stagingAllocationInfo{};
-        stagingAllocationInfo.usage = VMA_MEMORY_USAGE_AUTO_PREFER_HOST;
-        stagingAllocationInfo.flags = VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;
-        stagingAllocationInfo.requiredFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-        VK_VERIFY(vmaCreateBuffer(m_allocator, &bufferInfo, &stagingAllocationInfo, &stagingBuffer, &stagingAllocation, nullptr),
-                  "vmaCreateBuffer(staging texture)");
-
-        void* mapped = nullptr;
-        VK_VERIFY(vmaMapMemory(m_allocator, stagingAllocation, &mapped), "vmaMapMemory(staging texture)");
         for (const auto& item : uploadItems) {
-            Uint8* dst = static_cast<Uint8*>(mapped) + item.offset;
+            Uint8* dst = mapped + item.offset;
             if (!item.subRegion) {
                 std::memcpy(dst, item.source, item.uploadByteSize);
                 continue;
             }
-            // Tight-pack the dirty box: the shadow keeps whole-level rows, the
+            // Tight-pack the dirty box(es): the shadow keeps whole-level rows, the
-            // staging slice holds only the region (bufferRowLength stays 0).
+            // staging slice holds only the region (bufferRowLength stays 0). Multi-
+            // rect items pack their rects back to back in list order; the copy loop
+            // below recomputes the same running offsets.
             const SizeT levelRowBytes = static_cast<SizeT>(item.texelSize.x()) * item.texelBytes;
             const SizeT levelSliceBytes = static_cast<SizeT>(item.texelSize.y()) * levelRowBytes;
-            const SizeT regionRowBytes = static_cast<SizeT>(item.regionSize.x()) * item.texelBytes;
             const Uint8* src = static_cast<const Uint8*>(item.source);
-            for (Int z = 0; z < item.regionSize.z(); ++z) {
+            const auto packBox = [&](Uint8* out, const IntVec3& lo, const IntVec3& boxSize) {
-                for (Int y = 0; y < item.regionSize.y(); ++y) {
+                const SizeT boxRowBytes = static_cast<SizeT>(boxSize.x()) * item.texelBytes;
-                    const Uint8* srcRow = src +
+                for (Int z = 0; z < boxSize.z(); ++z) {
-                                          static_cast<SizeT>(item.regionLo.z() + z) * levelSliceBytes +
+                    for (Int y = 0; y < boxSize.y(); ++y) {
-                                          static_cast<SizeT>(item.regionLo.y() + y) * levelRowBytes +
+                        const Uint8* srcRow = src + static_cast<SizeT>(lo.z() + z) * levelSliceBytes +
-                                          static_cast<SizeT>(item.regionLo.x()) * item.texelBytes;
+                                              static_cast<SizeT>(lo.y() + y) * levelRowBytes +
-                    std::memcpy(dst + (static_cast<SizeT>(z) * item.regionSize.y() + y) * regionRowBytes,
+                                              static_cast<SizeT>(lo.x()) * item.texelBytes;
-                                srcRow, regionRowBytes);
+                        std::memcpy(out + (static_cast<SizeT>(z) * static_cast<SizeT>(boxSize.y()) + y) *
+                                              boxRowBytes,
+                                    srcRow, boxRowBytes);
+                    }
                 }
+                return static_cast<SizeT>(boxSize.x()) * static_cast<SizeT>(boxSize.y()) *
+                       static_cast<SizeT>(boxSize.z()) * item.texelBytes;
+            };
+            if (!item.rects.empty()) {
+                for (const auto& rect : item.rects) {
+                    dst += packBox(dst, rect.lo,
+                                   IntVec3{rect.hi.x() - rect.lo.x(), rect.hi.y() - rect.lo.y(),
+                                           rect.hi.z() - rect.lo.z()});
+                }
+                continue;
             }
+            packBox(dst, item.regionLo, item.regionSize);
         }
-        vmaUnmapMemory(m_allocator, stagingAllocation);
-
-        VkCommandBufferAllocateInfo allocInfo{};
-        allocInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
-        allocInfo.commandPool = m_commandPool;
-        allocInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
-        allocInfo.commandBufferCount = 1;
-
-        VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
-        VK_VERIFY(vkAllocateCommandBuffers(m_device, &allocInfo, &commandBuffer), "vkAllocateCommandBuffers(texture)");
-
-        VkCommandBufferBeginInfo beginInfo{};
-        beginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
-        beginInfo.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
-        VK_VERIFY(vkBeginCommandBuffer(commandBuffer, &beginInfo), "vkBeginCommandBuffer(texture)");
 
         const VkImageAspectFlags aspectMask = GetAspectMaskForFormat(outResource.format);
         VkPipelineStageFlags uploadSrcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
@@ -2348,9 +2611,49 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                             outResource.viewType == VK_IMAGE_VIEW_TYPE_2D_ARRAY ||
                                             outResource.viewType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
         for (const auto& item : uploadItems) {
+            if (!item.rects.empty()) {
+                // Multi-rect item: one VkBufferImageCopy per rect, all submitted in a
+                // single vkCmdCopyBufferToImage. The rect list is pairwise disjoint by
+                // construction, so no two copies write the same texels. Multi-rect
+                // implies subRegion, which implies a plain color aspect - the combined
+                // depth-stencil split below can never see one of these.
+                VkBufferImageCopy rectCopies[MG_State::GLState::MipmapStorage::kMaxDirtyRects];
+                Uint32 rectCopyCount = 0;
+                VkDeviceSize runningOffset = item.offset;
+                for (const auto& rect : item.rects) {
+                    const IntVec3 rectSize = {rect.hi.x() - rect.lo.x(), rect.hi.y() - rect.lo.y(),
+                                              rect.hi.z() - rect.lo.z()};
+                    const Uint32 rectDepth = static_cast<Uint32>(std::max(rectSize.z(), 1));
+                    VkBufferImageCopy rectCopy{};
+                    rectCopy.bufferOffset = stagingBase + runningOffset;
+                    rectCopy.bufferRowLength = 0;
+                    rectCopy.bufferImageHeight = 0;
+                    rectCopy.imageSubresource.aspectMask = aspectMask;
+                    rectCopy.imageSubresource.mipLevel = item.level;
+                    rectCopy.imageSubresource.baseArrayLayer = item.baseArrayLayer;
+                    rectCopy.imageSubresource.layerCount = 1;
+                    rectCopy.imageOffset = {rect.lo.x(), rect.lo.y(),
+                                            depthSelectsArrayLayer ? 0 : rect.lo.z()};
+                    rectCopy.imageExtent = {static_cast<Uint32>(rectSize.x()),
+                                            static_cast<Uint32>(rectSize.y()),
+                                            depthSelectsArrayLayer ? 1u : rectDepth};
+                    if (depthSelectsArrayLayer) {
+                        // The GL "depth" axis addresses array layers here, so a partial
+                        // z-range narrows the layer span rather than the extent.
+                        rectCopy.imageSubresource.baseArrayLayer =
+                            item.baseArrayLayer + static_cast<Uint32>(rect.lo.z());
+                        rectCopy.imageSubresource.layerCount = rectDepth;
+                    }
+                    rectCopies[rectCopyCount++] = rectCopy;
+                    runningOffset += static_cast<VkDeviceSize>(rect.TexelCount() * item.texelBytes);
+                }
+                vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image,
+                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, rectCopyCount, rectCopies);
+                continue;
+            }
             const Uint32 depthOrLayers = item.texelSize.z() > 0 ? static_cast<Uint32>(item.texelSize.z()) : 1u;
             VkBufferImageCopy copy{};
-            copy.bufferOffset = item.offset;
+            copy.bufferOffset = stagingBase + item.offset;
             copy.bufferRowLength = 0;
             copy.bufferImageHeight = 0;
             copy.imageSubresource.aspectMask = aspectMask;
@@ -2383,7 +2686,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 depthCopy.imageSubresource.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
                 VkBufferImageCopy stencilCopy = copy;
                 stencilCopy.imageSubresource.aspectMask = VK_IMAGE_ASPECT_STENCIL_BIT;
-                stencilCopy.bufferOffset = item.offset + static_cast<VkDeviceSize>(texelCount) * 4;
+                stencilCopy.bufferOffset = stagingBase + item.offset + static_cast<VkDeviceSize>(texelCount) * 4;
                 const VkBufferImageCopy copies[2] = {depthCopy, stencilCopy};
                 vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image,
                                        VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 2, copies);
@@ -2406,36 +2709,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to sampled read-only layout failed");
         outResource.layout = finalLayout;
 
-        VK_VERIFY(vkEndCommandBuffer(commandBuffer), "vkEndCommandBuffer(texture)");
+        // Ordering argument (replaces the old immediate per-texture submit):
-
+        // this upload is RECORDED into the shared batch command buffer, which
-        VkSubmitInfo submitInfo{};
+        // FlushPendingUploads submits - with one vkQueueSubmit and one pooled
-        submitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
+        // fence for the whole batch - strictly BEFORE any other submission on
-        submitInfo.commandBufferCount = 1;
+        // the same queue whose commands could consume the image: the renderer
-        submitInfo.pCommandBuffers = &commandBuffer;
+        // flushes at every frame-command-buffer submit (mid-frame flush,
-
+        // readback, Present), and the texture manager flushes before the
-        VkFenceCreateInfo fenceInfo{};
+        // preserve-on-recreate copy and before deferring an image the batch
-        fenceInfo.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
+        // references. The frame command buffer therefore still lands behind
-        VkFence uploadFence = VK_NULL_HANDLE;
+        // the uploads on the queue, so a texture uploaded and then immediately
-        VK_VERIFY(vkCreateFence(m_device, &fenceInfo, nullptr, &uploadFence), "vkCreateFence(texture upload)");
+        // sampled in the same frame sees its data exactly as it did when each
-
+        // upload was its own submit. No fence is waited here, for the same
-        VK_VERIFY(vkQueueSubmit(m_graphicsQueue, 1, &submitInfo, uploadFence), "vkQueueSubmit(texture)");
+        // reason as before: the batch queues behind the previous frame's
-        // Do NOT wait the fence here: this submit sits behind the previous
+        // rendering, and a synchronous wait would drain the GPU; the staging
-        // frame's rendering on the queue, so a synchronous wait stalls the CPU
+        // blocks/command buffer are parked on the reclaim list at flush time
-        // until the GPU drains - a per-frame vkQueueWaitIdle for any workload
+        // and recycled once the batch fence signals.
-        // with animated textures. Ordering against the current frame's draws is
+        if (std::find(m_uploadBatchImages.begin(), m_uploadBatchImages.end(), outResource.image) ==
-        // already guaranteed (its command buffer is submitted later, at
+            m_uploadBatchImages.end()) {
-        // present), so only the transient objects need to survive execution;
+            m_uploadBatchImages.push_back(outResource.image);
-        // park them until the fence signals.
-        m_pendingUploadReclaims.push_back({uploadFence, commandBuffer, stagingBuffer, stagingAllocation});
-        ReclaimCompletedUploads();
-        // Backstop for pathological upload storms: bound in-flight staging
-        // memory by blocking on the oldest upload only once the list is deep.
-        constexpr SizeT kMaxPendingTextureUploads = 16;
-        if (m_pendingUploadReclaims.size() > kMaxPendingTextureUploads) {
-            VK_VERIFY(vkWaitForFences(m_device, 1, &m_pendingUploadReclaims.front().fence, VK_TRUE, UINT64_MAX),
-                      "vkWaitForFences(texture upload backstop)");
-            ReclaimCompletedUploads();
         }
+        m_uploadBatchStagingBytes += stagingSize;
 
         if (!ok) {
             MGLOG_D("%s: texture upload cmd failed", __func__);
@@ -2445,6 +2739,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             mipmapTexture.MarkStorageDirty(item.target, item.level, false);
         }
         outResource.layout = finalLayout;
+        // Large batches flush right away instead of riding until the frame
+        // submit: a big copy amortizes its own vkQueueSubmit, submitting it
+        // early lets the GPU overlap the copy with the rest of the frame's
+        // CPU recording (measurably faster than a frame-tail burst), and the
+        // frame-tail burst pattern was observed to leave the GPU in a
+        // latency state that taxes whatever runs next. Small uploads keep
+        // accumulating, so a lightmap+sprite frame still costs one submit.
+        constexpr VkDeviceSize kEagerUploadFlushBytes = 128u * 1024u;
+        if (m_uploadBatchStagingBytes >= kEagerUploadFlushBytes) {
+            FlushPendingUploads();
+        }
         return true;
     }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -65,6 +65,11 @@ public:
         VkPipelineStageFlags sampledReadStageMask = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT |
                                                     VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
                                                     VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
+        // Family of `graphicsQueue`; the manager creates its own command pool
+        // on it for the recycled upload-batch command buffers, so their parked
+        // allocations never sit in (and fragment) the renderer's shared pool
+        // that frame command buffers churn through every frame.
+        Uint32 graphicsQueueFamilyIndex = 0;
     };
 
     struct TextureResource {
@@ -308,6 +313,14 @@ public:
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
     void BeginFrame(Uint32 frameIndex);
+    // Submits the accumulated texture-upload batch (one command buffer, one
+    // vkQueueSubmit, one pooled fence) if any uploads are pending. MUST run
+    // before any other vkQueueSubmit on the shared graphics queue whose
+    // commands may consume an image the batch writes - the frame command
+    // buffer submit (mid-frame flush, readback, Present) and the
+    // preserve-on-recreate copy are the existing callers. No-op when the
+    // batch is empty.
+    void FlushPendingUploads();
     // Drains every frame slot's deferred image/view releases. Only valid when
     // the caller has proven every queue submission complete; used by the
     // present-less frame-boundary drain.
@@ -453,6 +466,9 @@ private:
     VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
     VmaAllocator m_allocator = nullptr;
     VkCommandPool m_commandPool = VK_NULL_HANDLE;
+    // Dedicated pool for the recycled upload-batch command buffers (see
+    // InitInfo::graphicsQueueFamilyIndex).
+    VkCommandPool m_uploadCommandPool = VK_NULL_HANDLE;
     VkQueue m_graphicsQueue = VK_NULL_HANDLE;
     Bool m_imageFormatListSupported = false;
     Uint32 m_currentFrameIndex = 0;
@@ -506,15 +522,58 @@ private:
     std::unordered_map<VkFormat, VkSampleCountFlags> m_multisampleCountsByFormat;
     Vector<Vector<TextureResource>> m_deferredReleases;
     Vector<Vector<VkImageView>> m_deferredViewReleases;
+
+    // --- Batched upload machinery ---
+    // Uploads within a frame are recorded into ONE shared command buffer and
+    // submitted with ONE vkQueueSubmit at FlushPendingUploads (the renderer
+    // flushes before every frame-command-buffer submit). Staging memory comes
+    // from a pool of persistently-mapped, reusable blocks instead of a
+    // vmaCreateBuffer per upload.
+    struct UploadStagingBlock {
+        VkBuffer buffer = VK_NULL_HANDLE;
+        VmaAllocation allocation = nullptr;
+        Uint8* mapped = nullptr; // persistently mapped for the block's lifetime
+        VkDeviceSize capacity = 0;
+        VkDeviceSize cursor = 0; // bump cursor while the block backs the open batch
+    };
+    // Opens the batch command buffer lazily (allocates/reuses + begins recording).
+    VkCommandBuffer EnsureUploadBatchOpen();
+    // Bump-allocates `size` staging bytes for the open batch, growing onto a
+    // new/pooled block when the current one cannot fit. Returns the write
+    // pointer; outBuffer/outBaseOffset locate the space for copy commands.
+    Uint8* AcquireUploadStagingSpace(VkDeviceSize size, VkBuffer& outBuffer, VkDeviceSize& outBaseOffset);
+    void RecycleUploadStagingBlock(UploadStagingBlock&& block);
+    // Drops a recorded-but-unsubmitted batch on the floor. Shutdown only: the
+    // device is being torn down, so the lost texel data is unobservable.
+    void DiscardPendingUploadBatch();
+    void DestroyUploadPools();
+
+    Vector<UploadStagingBlock> m_freeUploadStagingBlocks;
+    VkDeviceSize m_freeUploadStagingBytes = 0;
+    Vector<VkCommandBuffer> m_freeUploadCommandBuffers;
+    Vector<VkFence> m_freeUploadFences;
+    Bool m_uploadBatchOpen = false;
+    VkCommandBuffer m_uploadBatchCommandBuffer = VK_NULL_HANDLE;
+    // Blocks whose staging bytes the open batch's copies reference (last =
+    // the block the bump cursor is currently allocating from).
+    Vector<UploadStagingBlock> m_uploadBatchBlocks;
+    // Images the open batch writes; consulted for the rare re-upload-after-
+    // draw flush and by DeferResourceRelease (an unsubmitted command buffer
+    // referencing a deferred-released image would escape every fence-based
+    // destruction proof, so the batch is flushed before the image is parked).
+    Vector<VkImage> m_uploadBatchImages;
+    VkDeviceSize m_uploadBatchStagingBytes = 0;
+
     // Texture uploads are submitted out-of-band but NOT waited on (waiting
     // behind the queue serialized the CPU against the previous frame's GPU
-    // work every time an animated atlas re-uploaded). Their transient objects
+    // work every time an animated atlas re-uploaded). Each flushed batch's
-    // are parked here and reclaimed once the upload fence signals.
+    // transients are parked here and RECYCLED (fence reset to the fence pool,
+    // command buffer reset to the CB pool, staging blocks back to the block
+    // pool) once the batch fence signals.
     struct PendingUploadReclaim {
         VkFence fence = VK_NULL_HANDLE;
         VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
-        VkBuffer stagingBuffer = VK_NULL_HANDLE;
+        Vector<UploadStagingBlock> stagingBlocks;
-        VmaAllocation stagingAllocation = nullptr;
     };
     Vector<PendingUploadReclaim> m_pendingUploadReclaims;
 };

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -327,6 +327,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool RecreateSwapchain();
 
     private:
+        // Tiered emission for an already-set-up multi-draw batch (state bound, index
+        // buffer bound for the indexed form). Tier 1: VK_EXT_multi_draw. Tier 2: one
+        // vkCmdDraw(Indexed)Indirect over a transient command array. Tier 3: unrolled
+        // vkCmdDraw(Indexed) loop. Tier eligibility is per-batch (uniform instance
+        // state for tier 1, firstInstance/feature legality for tier 2); every tier
+        // consumes the same param span, so contiguous-run merging done by the caller
+        // benefits all of them.
+        void EmitMultiDrawIndexed(VkCommandBuffer commandBuffer, const DrawIndexedCmdParam* pParams, Uint32 drawCount);
+        void EmitMultiDraw(VkCommandBuffer commandBuffer, const DrawCmdParam* pParams, Uint32 drawCount);
+
         struct BlitUniformData {
             float srcRect[4] = {0.f, 0.f, 1.f, 1.f};
             float dstRect[4] = {0.f, 0.f, 1.f, 1.f};
@@ -477,6 +487,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool m_indexTypeUint8ExtensionEnabled = false;
         Bool m_logicOpFeatureEnabled = false;
         Bool m_multiDrawIndirectFeatureEnabled = false;
+        // drawIndirectFirstInstance gates indirect commands whose firstInstance != 0;
+        // cached at device creation because the tier-2 multi-draw path (a transient
+        // VkDrawIndexedIndirectCommand array) is illegal for such a sub-draw without it.
+        Bool m_drawIndirectFirstInstanceFeatureEnabled = false;
+        // VK_EXT_multi_draw: native batched submission for the CPU-side glMultiDraw*
+        // families (tier 1 of the multi-draw dispatch).
+        Bool m_multiDrawExtensionEnabled = false;
+        Uint32 m_maxMultiDrawCount = 0;
+        // Multi-draw dispatch tiers, resolved once at device creation from device support
+        // clamped by MOBILEGL_MAGMA_MULTIDRAW_MODE (a preference, never a demand):
+        //   tier 1 (ext):      one vkCmdDrawMulti(Indexed)EXT           - m_multiDrawAllowExt
+        //   tier 2 (indirect): one vkCmdDraw(Indexed)Indirect batch     - m_multiDrawAllowIndirect
+        //   tier 3 (unroll):   one vkCmdDraw(Indexed) per sub-draw      - always available
+        // m_multiDrawForceUnrollIndirect additionally forces the GPU-parameter
+        // glMultiDraw*Indirect paths onto their per-command loop (mode=unroll only).
+        Bool m_multiDrawAllowExt = false;
+        Bool m_multiDrawAllowIndirect = false;
+        Bool m_multiDrawForceUnrollIndirect = false;
         Bool m_samplerAnisotropyFeatureEnabled = false;
         Bool m_shaderDrawParametersExtensionEnabled = false;
         Bool m_shaderDrawParametersFeatureEnabled = false;
@@ -508,6 +536,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                                  VkDeviceSize countBufferOffset, Uint32 maxDrawCount,
                                                                  Uint32 stride);
         static inline PFNDrawIndexedIndirectCountFunc s_vkCmdDrawIndexedIndirectCount = nullptr;
+        // VK_EXT_multi_draw entry points, loaded at device creation when the extension
+        // (and its multiDraw feature) is enabled; null otherwise.
+        static inline PFN_vkCmdDrawMultiEXT s_vkCmdDrawMultiEXT = nullptr;
+        static inline PFN_vkCmdDrawMultiIndexedEXT s_vkCmdDrawMultiIndexedEXT = nullptr;
 
         // VK_EXT_transform_feedback (GL transform feedback capture)
         Bool m_transformFeedbackFeatureEnabled = false;
@@ -531,7 +563,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Per-pipeline provoking-vertex mode. capturesXfbFromGeometryStage must be a LINK-TIME
         // property of the program, never the dynamic "is transform feedback active" flag: the
         // 8-entry m_pipelineMemo and the SetupDrawSnapshot fast path key on programObj.hash and
-        // GetRenderStateParametersVersion(), neither of which moves when glBeginTransformFeedback is
+        // the pipeline-state value hash, neither of which moves when glBeginTransformFeedback is
         // called, so a dynamic input here would hand back a stale VkPipeline.
         VkProvokingVertexModeEXT SelectProvokingVertexMode(VkPrimitiveTopology topology,
                                                           Bool capturesXfbFromGeometryStage) const;
@@ -623,7 +655,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 programHash = 0;
             Uint64 vertexInputHash = 0;
             Uint64 renderPassHash = 0;
-            Uint renderStateVersion = 0;
+            // VALUE hash of the pipeline-relevant fixed-function state (see
+            // ComputePipelineStateHash), not the monotonic pipeline-state version:
+            // the version never repeats, so a per-draw GL_BLEND toggle would miss
+            // all entries forever even though the state alternates between two
+            // values the memo already holds.
+            Uint64 pipelineStateHash = 0;
             ProgramFactory::CompileOptionFlags transformFlags = {};
             VkPipeline pipeline = VK_NULL_HANDLE;
         };
@@ -631,11 +668,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         PipelineMemoEntry m_pipelineMemo[kPipelineMemoSize];
         Uint32 m_pipelineMemoCount = 0;
         Uint32 m_pipelineMemoNext = 0;
+        // Hash of every fixed-function GL state the pipeline payload reads that the
+        // memo key's other fields (mode / program / vertex input / render pass /
+        // transform flags) do not already pin down. Equal hash under an equal rest
+        // of key => byte-identical PipelineCreatePayload. Cached per pipeline-state
+        // version: the version is monotonic and bumps on every pipeline-state
+        // change, so an unchanged (version, colorAttachmentCount) proves the state
+        // bytes are unchanged and the hash can be reused without re-reading them.
+        Uint64 ComputePipelineStateHash(Uint32 colorAttachmentCount) const;
+        Uint m_pipelineStateHashVersion = 0;
+        Uint32 m_pipelineStateHashColorCount = 0;
+        Uint64 m_pipelineStateHash = 0;
+        Bool m_pipelineStateHashValid = false;
+        // GetShaderTransformFlags memo. NOT pure in the pre-transform alone: the
+        // function also reads whether the bound DRAW framebuffer is the default one
+        // (only the default framebuffer gets the Y-flip and rotation bits - an FBO
+        // pass renders unflipped). Keyed on BOTH inputs; missing the FBO bit shipped
+        // an upside-down default-framebuffer pass after any render-to-texture
+        // (minecraft-1.17-main-menu retrace, whole frame flipped).
+        VkSurfaceTransformFlagBitsKHR m_baseTransformFlagsPreTransform =
+            VK_SURFACE_TRANSFORM_FLAG_BITS_MAX_ENUM_KHR;
+        Bool m_baseTransformFlagsIsDefaultFbo = false;
+        Bool m_baseTransformFlagsKeyValid = false;
+        Uint32 m_baseTransformFlagsCache = 0;
+        // isDefaultFbo must be the default-ness of the CURRENTLY bound draw framebuffer;
+        // every caller already has it in hand from its own guards.
+        Uint32 GetBaseTransformFlagsRaw(Bool isDefaultFbo);
         // Drops every memoized pipeline handle. Required at command-buffer
-        // boundaries and whenever any pipeline may have been destroyed.
+        // boundaries and whenever any pipeline may have been destroyed. Also drops
+        // the cached pipeline-state hash: the same boundaries can retire the GL
+        // context whose monotonic version the cache is keyed on.
         void InvalidatePipelineMemo() {
             m_pipelineMemoCount = 0;
             m_pipelineMemoNext = 0;
+            m_pipelineStateHashValid = false;
         }
         UnorderedMap<ProgramFactory::HashType, VkPipeline> m_computePipelines;
         UniquePtr<ProgramFactory> m_programFactory;
@@ -692,6 +758,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 programLifetimeId = 0;
             Uint32 programVersion = 0;
             const void* vao = nullptr;
+            // Same rule as VaoDrawMemo::vaoLifetimeId: (address, config version) is not an
+            // identity, because a recycled address can arrive carrying a config version
+            // the dead VAO also had (two mutations to configure one attribute is the
+            // common shape), and "the VAO did not move" would then skip the layout
+            // re-resolve for a different VAO.
+            Uint64 vaoLifetimeId = 0;
             Uint32 vaoConfigVersion = 0;
             const void* drawFbo = nullptr;
             Uint16 fboVersion = 0;
@@ -707,14 +779,77 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 renderbufferImageEpoch = 0;
             Uint64 sampledContentSum = 0;
             Uint64 sampledParamsSum = 0;
+            // Guards the sampler-descriptor reuse hint: bumped by any sampler-object
+            // parameter or texture shape change (see GetSamplingResolutionGeneration),
+            // none of which the sums above cover.
+            Uint64 samplingResolutionGeneration = 0;
+            // Render-pass flavor input (DepthTest || StencilTest at snapshot time).
+            // A pipeline-state change that leaves this equal cannot change which
+            // render pass GetOrCreateRenderPass would pick, so the fast path may
+            // re-resolve just the pipeline against the active pass; a change that
+            // flips it must fall back to the full path's pass selection.
+            Bool drawUsesDepthStencil = false;
             IntVec2 renderPassExtent = {0, 0};
+            // colorAttachmentCount of the snapshotting draw's render pass: the
+            // pipeline-state hash input, so the fast path can refresh that hash and
+            // probe the pipeline memo after a state change without re-fetching the
+            // render-pass entry (the pass itself is pinned by renderPassHash above).
+            Uint32 renderPassColorCount = 0;
             VkPipeline pipeline = VK_NULL_HANDLE;
+            // layoutHash of the snapshotting draw's vertex-input state. The pipeline and
+            // the vertex-input pre-flight depend on the VAO only through this (plus the
+            // program, pinned separately), so a changed VAO whose aux memo carries the
+            // same layoutHash re-uses the snapshot's pipeline and pre-flight verdict
+            // outright - the VAO-cycling case Minecraft chunk rendering hits every draw.
+            Uint64 vaoLayoutHash = 0;
+            // Memoised ProgramFactory entry of the snapshotting draw, valid while
+            // (programLifetimeId, programVersion, resolvedTransformFlags) match - all
+            // checked above - AND the factory's cache structure epoch is unchanged (the
+            // cache is open-addressing and holds entries by value, so any insert/erase
+            // moves them). The fast path must re-stamp use through StampProgramUse when
+            // it bypasses GetOrCreateProgram, or the idle sweep could evict a live entry.
+            const ProgramFactory::VkProgramObject* programObj = nullptr;
+            Uint64 programFactoryEpoch = 0;
+            // Per-entry copies of the snapshotting draw's sampled set (the scratch
+            // vectors below hold only the LAST full-path draw's set, which with more
+            // than one snapshot entry is not necessarily this entry's program).
+            // sampledTextures/sampledResources carry the same epoch-guarded pointer
+            // lifetime rules as the scratch originals: textureEraseEpoch (checked
+            // every probe) declines the entry before any erased resource pointer
+            // could be dereferenced. sampledLayouts is the layout VALUE each
+            // resource held when this entry's descriptors were built (the
+            // descriptor-reuse hint needs the SAME layout, not just a sampleable
+            // one), and sampledBindingRecords feeds SampledBindingsUnchanged when
+            // the bind generation moved.
+            Vector<MG_State::GLState::ITextureObject*> sampledTextures;
+            Vector<VkTextureManager::TextureResource*> sampledResources;
+            Vector<VkImageLayout> sampledLayouts;
+            Vector<UniformManager::SampledBindingRecord> sampledBindingRecords;
         };
-        SetupDrawSnapshot m_setupDrawSnapshot;
+        // Program-keyed snapshot entries: program ping-pong (Sodium switches programs
+        // mid-frame every few draws) would otherwise evict the single snapshot on
+        // every switch and send every draw through the full path. Entries are found
+        // by programLifetimeId (MRU-first probe); every other guard stays per-probe,
+        // so a stale entry declines itself exactly like the old single snapshot did.
+        static constexpr Uint32 kSetupDrawSnapshotCount = 4;
+        SetupDrawSnapshot m_setupDrawSnapshots[kSetupDrawSnapshotCount];
+        Uint32 m_setupDrawSnapshotMru = 0;    // last entry that hit or was filled
+        Uint32 m_setupDrawSnapshotVictim = 0; // round-robin fill cursor when all entries are live
+        void InvalidateSetupDrawSnapshots() {
+            for (auto& snapshot : m_setupDrawSnapshots) {
+                snapshot.valid = false;
+            }
+        }
 
         // Per-draw scratch buffers (clear keeps capacity) — these paths run for every
         // draw call and must not allocate.
         Vector<MG_State::GLState::ITextureObject*> m_sampledTexturesScratch;
+        // Per-binding (texture, effective sampler) lifetime-id records from the same
+        // CollectSampledTextures walk that filled m_sampledTexturesScratch. The fast
+        // path shadow-compares against them (SampledBindingsUnchanged) when the
+        // texture bind generation moved, so a redundant glBindSampler/glBindTexture
+        // storm that resolves to the same bindings keeps the fast path.
+        Vector<UniformManager::SampledBindingRecord> m_sampledBindingRecordsScratch;
         // Parallel to m_sampledTexturesScratch, refilled by every SetupDraw's
         // first sampled-texture loop: the resolved backend resources, so the
         // post-transition loop can skip re-resolving textures whose layout is
@@ -779,6 +914,138 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         UnorderedMap<ConvertedVertexStreamKey, ConvertedVertexStream, ConvertedVertexStreamKeyHash>
             m_convertedVertexStreams;
 
+        // One VAO's resolved vkCmdBindVertexBuffers arguments, reusable by a later draw
+        // that would resolve them to the same thing. Consecutive draws in a chunk-renderer
+        // frame keep the program and the vertex layout and only swap the VAO, so a
+        // per-VAO memo turns the second and later draws through each VAO into a validate
+        // plus (usually skipped) rebind.
+        //
+        // Only whole-buffer bindings are memoised. Client-memory and format-converted
+        // streams re-upload from a range that depends on the draw's own vertex/index
+        // range, and synthetic bindings carry glVertexAttrib* values that are not part
+        // of any key here; a layout using any of them is never stored.
+        // Field order is hit-path cache locality, hot to cold: the per-draw validate
+        // reads the scalars and the EBO memo head, then only the first bindingCount
+        // elements of vkBuffers/vkOffsets; the per-binding revalidation arrays at the
+        // tail are touched once per frame at most.
+        struct ResolvedVertexBindings {
+            // Must equal DynamicStateShadow::kMaxShadowedVertexBindings (static_assert in
+            // the .cpp): past that width the bind shadow cannot skip a redundant bind
+            // either, so a wider layout resolves per draw. Minecraft-shaped layouts use four.
+            static constexpr Uint32 kMaxBindings = 8;
+
+            // Frame serial of the last completed resolve OR cross-frame revalidation.
+            // Zero until a resolve completes, and reset to zero before one starts, so a
+            // resolve that bails out midway cannot leave a half-filled entry matchable.
+            // Unlike the original frame-scoped memo, an entry whose buffers are all
+            // resident and unmapped is revalidated across frames (per-binding slice
+            // epoch compares) instead of re-resolved - see TryBindResolvedVertexBindings.
+            Uint64 frameSerial = 0;
+            // Identity of the resolved Vulkan layout: the VAO's content hash
+            // (VertexInputStateFactory::GetOrComputeHash - the same value the factory
+            // keys its entries on) fixes bindings.size(), each binding's base offset,
+            // which bindings are client/converted, and (through the mixed-in buffer
+            // addresses) which buffer each binding reads. Compared against the VAO's
+            // own hash memo on the hit path, so a hit never touches the factory entry.
+            VertexInputStateFactory::HashType vertexInputHash = 0;
+            // The program's vertex input layout: decides the synthetic-binding set and
+            // hence the total binding count.
+            Uint32 activeAttribMask = 0;
+            Uint32 bindingCount = 0;
+            // VkBufferManager::GetSliceEpochCounter() at resolve time. Still equal means
+            // no buffer anywhere changed its slice or was persistently mapped since, which
+            // settles every per-binding question below in one compare.
+            Uint64 sliceEpochCounter = 0;
+            // Any bound buffer already carrying a host map when the slice was resolved.
+            // Such a buffer can mutate its shadow with no API call, so it has to be
+            // re-pushed per draw and the one-compare path above cannot apply.
+            Bool anyBufferMapped = true;
+
+            // Resident element-buffer slice memo (skips the per-draw AcquireResidentSlice
+            // for the VAO's EBO, which cold-chases 500+ distinct resources in a
+            // chunk-cycling frame). Self-validating exactly like the bindings above: a hit
+            // requires the LIVE bound EBO pointer to equal indexBuffer AND either an
+            // unmoved manager-wide slice-epoch counter (nothing anywhere changed slices
+            // or gained a host map, the same one-compare rescue the vertex half uses) or
+            // that buffer's resource still carrying indexSliceEpoch (epochs are minted
+            // from a process-lifetime counter, so a recycled address can never
+            // revalidate). Restart-substituted and streamed EBOs are never stored.
+            // indexFrameSerial tracks the last frame the resource's GPU-use serial was
+            // stamped through this memo; 0 means no index memo. Independent of the
+            // vertex half: both are (pointer, epoch)-validated, so neither can serve
+            // stale state for the other.
+            const MG_State::GLState::BufferObject* indexBuffer = nullptr;
+            Uint64 indexSliceEpoch = 0;
+            // GetSliceEpochCounter() when the resource's epoch was last verified; only
+            // meaningful while indexFrameSerial matches the current frame serial.
+            Uint64 indexSliceEpochCounter = 0;
+            VkBuffer indexVkBuffer = VK_NULL_HANDLE;
+            VkDeviceSize indexSliceOffset = 0;
+            Uint64 indexFrameSerial = 0;
+
+            // Bound per draw (first bindingCount elements).
+            VkBuffer vkBuffers[kMaxBindings] = {};
+            VkDeviceSize vkOffsets[kMaxBindings] = {};
+            // Per binding: the VAO attribute location its buffer comes from, that buffer,
+            // and the buffer's VkBufferManager slice epoch when the slice was resolved.
+            // Only read by the per-frame revalidation and the something-moved fallback.
+            Uint8 attributeLocations[kMaxBindings] = {};
+            const MG_State::GLState::BufferObject* buffers[kMaxBindings] = {};
+            Uint64 sliceEpochs[kMaxBindings] = {};
+        };
+        // One direct-mapped slot of the per-VAO draw-memo table below. A slot belongs to
+        // the object whose (vaoKey, vaoLifetimeId) pair it carries: the address alone
+        // only picks the slot, and the never-reused lifetime id is what proves the slot
+        // is THIS VAO's, so the successor allocated onto a destroyed VAO's address
+        // always misses. That identity check is load-bearing and the content-hash
+        // validations below do NOT stand in for it - a recycled address under a
+        // byte-identical configuration reproduces the content hash exactly, which is
+        // how a destroyed VAO's resolved bindings were once handed to its successor's
+        // draw. The slot is still never dereferenced through vaoKey, and every fact it
+        // carries is still validated against live state before use:
+        //  - layoutHash/layoutAuxMasks are valid only while contentHash equals the LIVE
+        //    VAO's own hash memo (which the VAO's config version guards), so a config
+        //    change or a buffer rebind misses even for the same object.
+        //  - bindings revalidates per draw exactly as before (frame serial, content
+        //    hash, per-binding live buffer pointers and slice epochs).
+        struct alignas(64) VaoDrawMemo {
+            const MG_State::GLState::VertexArrayObject* vaoKey = nullptr;
+            // The VAO's never-reused lifetime id, checked alongside vaoKey. The pointer
+            // ALONE is not an identity: a deleted VAO's heap address is handed straight
+            // back by the next glGenVertexArrays-shaped allocation, and the successor then
+            // matched this slot and inherited the dead object's memos. Both stated
+            // defences failed with it, because both reduce to the content hash and the
+            // content hash's buffer-identity component was itself a recycled heap address.
+            Uint64 vaoLifetimeId = 0;
+            // The VAO content hash (VertexInputStateFactory::GetOrComputeHash) the two
+            // layout facts below were derived from; 0 while nothing valid is stored.
+            Uint64 contentHash = 0;
+            Bool layoutFactsValid = false;
+            // The resolved layout identity + packed (unsupported, location) masks -
+            // the exact values GetBackendAuxMemo used to serve, moved here so the
+            // per-draw probe stays inside this table's one hot line instead of
+            // touching a second cold line of every cycled VAO object.
+            Uint64 layoutHash = 0;
+            Uint64 layoutAuxMasks = 0;
+            ResolvedVertexBindings bindings;
+        };
+        // Fixed-size, allocated on first use, never rehashed or swept: entries are
+        // recycled in place on slot collisions (two-slot probe, older frame serial
+        // evicted), and stale entries self-invalidate through the compares above. A
+        // fixed table also makes every VaoDrawMemo/ResolvedVertexBindings pointer
+        // stable for the duration of a draw, which the EBO memo handoff
+        // (m_currentDrawResolvedEntry) relies on.
+        static constexpr Uint32 kVaoDrawMemoSlotCount = 2048; // power of two
+        Vector<VaoDrawMemo> m_vaoDrawMemoTable;
+        // Finds the slot holding `vao`, or recycles the older of its two candidate
+        // slots into an empty memo keyed on `vao`. Never returns null.
+        VaoDrawMemo* LookupVaoDrawMemo(const MG_State::GLState::VertexArrayObject* vao);
+        // The current draw's memo entry, set by UploadAndBindVertexBuffers and consumed
+        // by the same draw's UploadAndBindIndexBuffer (the EBO memo lives in the same
+        // entry). Valid ONLY within that window: the next draw's lookup can recycle the
+        // slot. Null when the draw's layout is not memoisable.
+        ResolvedVertexBindings* m_currentDrawResolvedEntry = nullptr;
+
         void CreateInstance();
         VkResult SetupDebugMessenger();
         VkResult DestroyDebugMessenger();
@@ -809,10 +1076,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj);
 
+        // The per-draw dynamic-state tail (viewport, scissor, blend constants, depth
+        // bias, line width, stencil), gated behind one render-state-parameters-version
+        // compare per command buffer - see the gate fields in DynamicStateShadow.
+        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo);
+
         Bool UploadAndBindVertexBuffers(VkCommandBuffer commandBuffer, const MG_State::GLState::VertexArrayObject& vao,
                                         const ProgramFactory::VkProgramObject& programObj,
                                         const DrawCmdParam& drawParams,
                                         const IndexBufferView* pIndexBufferView);
+        // Binds `entry`'s memoised buffers when every input it was resolved from is
+        // still live and unchanged, else returns false and leaves nothing bound.
+        // vaoContentHash is the VAO's memoised content hash (GetBackendHashMemo), which
+        // pins the layout AND the bound buffers without resolving the factory entry.
+        // Non-const entry: a cross-frame revalidation refreshes its serial/epoch stamps.
+        Bool TryBindResolvedVertexBindings(VkCommandBuffer commandBuffer,
+                                           const MG_State::GLState::VertexArrayObject& vao,
+                                           ResolvedVertexBindings& entry,
+                                           Uint64 vaoContentHash,
+                                           Uint32 activeAttribMask, Uint64 frameSerial);
         Bool UploadAndBindIndexBuffer(FrameContext::FrameData& frame,
                                      const MG_State::GLState::VertexArrayObject& vao,
                                       const IndexBufferView* pIndexBufferView = nullptr);

MobileGL/MG_Benchmark/Driver/DriverBench.c

@@ -12,8 +12,10 @@
  * libEGL.so.1 for the native driver, or a libMobileGL.so path for either
  * MobileGL backend selected with MOBILEGL_BACKEND_TYPE), creates a desktop-GL
  * context on a small pbuffer, renders into its own FBO and paces frames with
- * glFinish. No window system is required beyond what the provider itself
+ * glFinish. No window system is required: the default display is tried first
- * needs - see run_driver_bench.sh.
+ * so a desktop run reaches the real driver, and a headless box (CI, a build
+ * server) falls back to EGL_MESA_platform_surfaceless - see
+ * run_driver_bench.sh.
  *
  * Every case models one hot pattern from captured Minecraft traces:
  *   draw_tiny        back-to-back glDrawElements, shared state (chunk batch)
@@ -67,6 +69,7 @@ typedef unsigned int EGLenum;
 #define EGL_CONTEXT_MINOR_VERSION 0x30FB
 #define EGL_CONTEXT_OPENGL_PROFILE_MASK 0x30FD
 #define EGL_CONTEXT_OPENGL_CORE_PROFILE_BIT 0x00000001
+#define EGL_PLATFORM_SURFACELESS_MESA 0x31DD
 
 /* ---- GL constants ---- */
 #define GL_COLOR_BUFFER_BIT 0x00004000
@@ -89,6 +92,11 @@ typedef unsigned int EGLenum;
 #define GL_NEAREST 0x2600
 #define GL_NEAREST_MIPMAP_LINEAR 0x2702
 #define GL_DEPTH_TEST 0x0B71
+#define GL_BLEND 0x0BE2
+#define GL_SRC_ALPHA 0x0302
+#define GL_ONE_MINUS_SRC_ALPHA 0x0303
+#define GL_ONE 1
+#define GL_ZERO 0
 #define GL_VERTEX_SHADER 0x8B31
 #define GL_FRAGMENT_SHADER 0x8B30
 #define GL_COMPILE_STATUS 0x8B81
@@ -133,6 +141,9 @@ static void* g_provider;
 GLF(void, glClear, (unsigned))
 GLF(void, glClearColor, (float, float, float, float))
 GLF(void, glEnable, (GLenum))
+GLF(void, glDisable, (GLenum))
+GLF(void, glBlendFuncSeparate, (GLenum, GLenum, GLenum, GLenum))
+GLF(void, glDrawBuffers, (GLsizei, const GLenum*))
 GLF(void, glViewport, (GLint, GLint, GLsizei, GLsizei))
 GLF(const unsigned char*, glGetString, (GLenum))
 GLF(GLenum, glGetError, (void))
@@ -280,7 +291,21 @@ static void run_case(const char* name, case_fn body, long a, long b, long opsPer
     if (glGetError() != GL_NO_ERROR) fprintf(stderr, "WARN: GL error after %s\n", name);
 }
 
-/* a = draws per frame */
+/* A display that needs no window system. eglGetPlatformDisplay is EGL 1.5
+ * core and eglGetPlatformDisplayEXT is the EGL_EXT_platform_base spelling
+ * older loaders ship; both are client entry points, so they resolve before
+ * any display exists. Only the attribute-list types differ between the two
+ * and this passes none, so one cast covers both. */
+static EGLDisplay surfaceless_display(void) {
+    void* fn = dlsym(g_provider, "eglGetPlatformDisplay");
+    if (!fn) fn = g_eglGetProcAddress("eglGetPlatformDisplay");
+    if (!fn) fn = dlsym(g_provider, "eglGetPlatformDisplayEXT");
+    if (!fn) fn = g_eglGetProcAddress("eglGetPlatformDisplayEXT");
+    if (!fn) return NULL;
+    return ((EGLDisplay(*)(EGLenum, void*, const void*))fn)(EGL_PLATFORM_SURFACELESS_MESA,
+                                                            EGL_DEFAULT_DISPLAY, NULL);
+}
+
 /* ---- EGL bootstrap: one provider library, pbuffer, desktop-GL context ---- */
 static int boot_egl(void) {
     const char* libpath = getenv("DRIVERBENCH_EGL_LIB");
@@ -304,14 +329,30 @@ static int boot_egl(void) {
     ESYM(eglGetError)
     g_eglGetProcAddress = (void* (*)(const char*))p_eglGetProcAddress;
 
-    EGLDisplay dpy = ((EGLDisplay(*)(void*))p_eglGetDisplay)(EGL_DEFAULT_DISPLAY);
+    EGLint (*getError)(void) = (EGLint(*)(void))p_eglGetError;
-    if (!dpy) { fprintf(stderr, "FAIL: eglGetDisplay\n"); return 1; }
+    EGLBoolean (*initialize)(EGLDisplay, EGLint*, EGLint*) =
+        (EGLBoolean(*)(EGLDisplay, EGLint*, EGLint*))p_eglInitialize;
+
+    /* The default display first: it is the one a windowed app would get, and
+     * on a desktop it is the one that reaches the real GPU - which is the
+     * driver this bench exists to measure. It does need a window system,
+     * though; Mesa's default platform is X11, so with no $DISPLAY (CI, a
+     * build server, ssh without forwarding) eglInitialize fails. Fall back to
+     * EGL_MESA_platform_surfaceless rather than give up: every case draws into
+     * the FBO built by build_resources(), so no window is needed for any of
+     * the work being timed. */
     EGLint maj = 0, min = 0;
-    if (!((EGLBoolean(*)(EGLDisplay, EGLint*, EGLint*))p_eglInitialize)(dpy, &maj, &min)) {
+    const char* how = "default display";
-        fprintf(stderr, "FAIL: eglInitialize (0x%x)\n", ((EGLint(*)(void))p_eglGetError)());
+    EGLDisplay dpy = ((EGLDisplay(*)(void*))p_eglGetDisplay)(EGL_DEFAULT_DISPLAY);
-        return 1;
+    if (!dpy || !initialize(dpy, &maj, &min)) {
+        dpy = surfaceless_display();
+        how = "surfaceless display";
+        if (!dpy || !initialize(dpy, &maj, &min)) {
+            fprintf(stderr, "FAIL: eglInitialize (0x%x)\n", getError());
+            return 1;
+        }
     }
-    fprintf(stderr, "EGL %d.%d via %s\n", maj, min, libpath);
+    fprintf(stderr, "EGL %d.%d via %s (%s)\n", maj, min, libpath, how);
 
     // Desktop GL first (that is what MobileGL exposes and what the cases are
     // written against), GLES 3 second so the same binary can measure a device's
@@ -348,7 +389,10 @@ static int boot_egl(void) {
                                        EGL_RENDERABLE_TYPE, EGL_OPENGL_ES3_BIT, EGL_NONE};
         ncfg = 0;
         if (!chooseConfig(dpy, esCfgAttribs, &cfg, 1, &ncfg) || ncfg < 1) {
-            const EGLint relaxed[] = {EGL_SURFACE_TYPE, EGL_PBUFFER_BIT, EGL_RED_SIZE, 8, EGL_NONE};
+            // EGL_SURFACE_TYPE 0 matches any config: a stack that offers no
+            // pbuffer at all is still usable through the surfaceless context
+            // path below.
+            const EGLint relaxed[] = {EGL_SURFACE_TYPE, 0, EGL_RED_SIZE, 8, EGL_NONE};
             if (!chooseConfig(dpy, relaxed, &cfg, 1, &ncfg) || ncfg < 1) {
                 fprintf(stderr, "FAIL: eglChooseConfig\n");
                 return 1;
@@ -358,20 +402,22 @@ static int boot_egl(void) {
         ctx = createContext(dpy, cfg, EGL_NO_CONTEXT, esCtxAttribs);
     }
     if (ctx == EGL_NO_CONTEXT) {
-        fprintf(stderr, "FAIL: eglCreateContext (0x%x)\n", ((EGLint(*)(void))p_eglGetError)());
+        fprintf(stderr, "FAIL: eglCreateContext (0x%x)\n", getError());
         return 1;
     }
 
+    /* The pbuffer only exists to have something to make current - nothing is
+     * ever drawn to it. Where there is no pbuffer config, EGL_NO_SURFACE is
+     * exactly what EGL_KHR_surfaceless_context takes, so the same call covers
+     * both. */
     const EGLint pbAttribs[] = {EGL_WIDTH, 64, EGL_HEIGHT, 64, EGL_NONE};
     EGLSurface surf = ((EGLSurface(*)(EGLDisplay, EGLConfig, const EGLint*))p_eglCreatePbufferSurface)(
         dpy, cfg, pbAttribs);
-    if (surf == EGL_NO_SURFACE) {
+    if (surf == EGL_NO_SURFACE)
-        fprintf(stderr, "FAIL: eglCreatePbufferSurface (0x%x)\n", ((EGLint(*)(void))p_eglGetError)());
+        fprintf(stderr, "no pbuffer (0x%x), using a surfaceless context\n", getError());
-        return 1;
-    }
     if (!((EGLBoolean(*)(EGLDisplay, EGLSurface, EGLSurface, EGLContext))p_eglMakeCurrent)(dpy, surf,
                                                                                           surf, ctx)) {
-        fprintf(stderr, "FAIL: eglMakeCurrent (0x%x)\n", ((EGLint(*)(void))p_eglGetError)());
+        fprintf(stderr, "FAIL: eglMakeCurrent (0x%x)\n", getError());
         return 1;
     }
 
@@ -385,6 +431,7 @@ static int boot_egl(void) {
         if (!name) { fprintf(stderr, "FAIL: resolve %s\n", #name); return 1; }     \
     } while (0)
     RESOLVE(glClear); RESOLVE(glClearColor); RESOLVE(glEnable); RESOLVE(glViewport);
+    RESOLVE(glDisable); RESOLVE(glBlendFuncSeparate); RESOLVE(glDrawBuffers);
     RESOLVE(glGetString); RESOLVE(glGetError); RESOLVE(glFinish); RESOLVE(glFlush);
     RESOLVE(glGenBuffers); RESOLVE(glBindBuffer); RESOLVE(glBufferData); RESOLVE(glBufferSubData);
     RESOLVE(glGenVertexArrays); RESOLVE(glBindVertexArray); RESOLVE(glEnableVertexAttribArray);

MobileGL/MG_Benchmark/Driver/DriverBenchCases.inc

@@ -34,6 +34,9 @@ static GLuint g_uboRing;
 static GLint g_uboAlign = 256;
 static size_t g_uboSlot = 256;
 static GLuint g_sampler;
+/* Two small offscreen targets for the 26.2-style render-pass churn case. */
+static GLuint g_passFbo[2];
+static GLuint g_passColor[2];
 static float g_mvp[16] = {0.002f, 0, 0, 0, 0, 0.002f, 0, 0, 0, 0, -0.001f, 0, -1.f, -1.f, 0.f, 1.f};
 
 /* Minecraft chunk vertex: pos 3f, color 4ub, uv 2f, packed light 2s -> 32 B */
@@ -162,10 +165,13 @@ static void setup_vao(GLuint vao, GLuint vbo, GLuint ibo) {
     glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ibo);
 }
 
+static GLuint g_mainFbo;
+
 static void build_resources(void) {
     /* offscreen render target: 1280x720 RBO FBO, like CTS fbo surface mode */
     GLuint fbo, rboColor, rboDepth;
     glGenFramebuffers(1, &fbo);
+    g_mainFbo = fbo;
     glBindFramebuffer(GL_FRAMEBUFFER, fbo);
     glGenRenderbuffers(1, &rboColor);
     glBindRenderbuffer(GL_RENDERBUFFER, rboColor);
@@ -257,6 +263,21 @@ static void build_resources(void) {
     glBufferData(GL_UNIFORM_BUFFER, 4 * 1024 * 1024, g_scratch, GL_DYNAMIC_DRAW);
     glBindBuffer(GL_UNIFORM_BUFFER, 0);
 
+    for (int i = 0; i < 2; ++i) {
+        glGenFramebuffers(1, &g_passFbo[i]);
+        glBindFramebuffer(GL_FRAMEBUFFER, g_passFbo[i]);
+        glGenRenderbuffers(1, &g_passColor[i]);
+        glBindRenderbuffer(GL_RENDERBUFFER, g_passColor[i]);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, 256, 256);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, g_passColor[i]);
+        if (glCheckFramebufferStatus(GL_FRAMEBUFFER) != GL_FRAMEBUFFER_COMPLETE) {
+            bench_gl_failed("pass FBO incomplete", "");
+            return;
+        }
+    }
+    /* back to the main offscreen target the harness set up */
+    glBindFramebuffer(GL_FRAMEBUFFER, g_mainFbo);
+
     glGenSamplers(1, &g_sampler);
     glSamplerParameteri(g_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
     glSamplerParameteri(g_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
@@ -516,6 +537,72 @@ static void case_mc_sampler_churn(int frame, long a, long b) {
 }
 
 
+/* 26.2 switches render targets constantly: 132 glBindFramebuffer and 198
+ * glDrawBuffers per frame. Pass switching is where a Vulkan backend pays for
+ * render-pass breaks, so this case is the one to watch on Magma. a = passes. */
+static void case_mc_pass_switch(int frame, long a, long b) {
+    (void)frame; (void)b;
+    static const GLenum kColor0[1] = {GL_COLOR_ATTACHMENT0};
+    glBindVertexArray(g_vao[0]);
+    for (long i = 0; i < a; ++i) {
+        glBindFramebuffer(GL_FRAMEBUFFER, g_passFbo[i & 1]);
+        glDrawBuffers(1, kColor0);
+        glViewport(0, 0, 256, 256);
+        glDrawElements(GL_TRIANGLES, g_quadsPerSection * 6, GL_UNSIGNED_INT, 0);
+        glDrawElements(GL_TRIANGLES, g_quadsPerSection * 6, GL_UNSIGNED_INT, 0);
+    }
+    glBindFramebuffer(GL_FRAMEBUFFER, g_mainFbo);
+    glViewport(0, 0, 1280, 720);
+}
+
+/* Blaze3D toggles blend around batches: 46 glEnable/glDisable pairs and 28
+ * glBlendFuncSeparate per vanilla frame. a = toggle pairs. */
+static void case_mc_state_toggle(int frame, long a, long b) {
+    (void)frame; (void)b;
+    glBindVertexArray(g_vao[0]);
+    for (long i = 0; i < a; ++i) {
+        glEnable(GL_BLEND);
+        glBlendFuncSeparate(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_ONE, GL_ZERO);
+        glDrawElements(GL_TRIANGLES, g_quadsPerSection * 6, GL_UNSIGNED_INT, 0);
+        glDisable(GL_BLEND);
+        glDrawElements(GL_TRIANGLES, g_quadsPerSection * 6, GL_UNSIGNED_INT, 0);
+    }
+}
+
+/* 26.2 re-sets texture parameters relentlessly - 612 glTexParameteri per frame,
+ * almost always to the value already in place. Measures redundant-param
+ * filtering. a = parameter writes. */
+static void case_mc_tex_param(int frame, long a, long b) {
+    (void)frame; (void)b;
+    glBindVertexArray(g_vao[0]);
+    glBindTexture(GL_TEXTURE_2D, g_texAtlas);
+    for (long i = 0; i < a; i += 4) {
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST_MIPMAP_LINEAR);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    }
+    glDrawElements(GL_TRIANGLES, g_quadsPerSection * 6, GL_UNSIGNED_INT, 0);
+}
+
+/* Sodium switches programs mid-frame far more than vanilla: 62 glUseProgram and
+ * 60 mat4 uploads per frame. a = program switches. */
+static void case_mc_use_program(int frame, long a, long b) {
+    (void)frame; (void)b;
+    glBindVertexArray(g_vao[0]);
+    for (long i = 0; i < a; ++i) {
+        if (i & 1) {
+            glUseProgram(g_progEntity);
+            glUniformMatrix4fv(g_uMvpEntity, 1, 0, g_mvp);
+        } else {
+            glUseProgram(g_progChunk);
+            glUniformMatrix4fv(g_uMvpChunk, 1, 0, g_mvp);
+        }
+        glDrawElements(GL_TRIANGLES, g_quadsPerSection * 6, GL_UNSIGNED_INT, 0);
+    }
+    glUseProgram(g_progChunk);
+}
+
 /* ---- the case table both harnesses iterate --------------------------------
  * a/b are the case's own knobs; opsPerFrame is what one bench frame is
  * normalised by, so ns_per_op compares across renderers. The mc_* rates are
@@ -536,6 +623,10 @@ static const BenchCaseDesc kBenchCases[] = {
     {"mc_tex_stream", case_mc_tex_stream, 95, 0, 95},
     {"mc_uniform_lookup", case_mc_uniform_lookup, 41, 0, 41},
     {"mc_sampler_churn", case_mc_sampler_churn, 306, 0, 306},
+    {"mc_pass_switch", case_mc_pass_switch, 132, 0, 132},
+    {"mc_state_toggle", case_mc_state_toggle, 46, 0, 46},
+    {"mc_tex_param", case_mc_tex_param, 612, 0, 612},
+    {"mc_use_program", case_mc_use_program, 62, 0, 62},
     {"draw_tiny", case_draw_tiny, 2048, 0, 2048},
     {"draw_uniform", case_draw_uniform, 2048, 0, 2048},
     {"draw_multi_vao", case_draw_multi_vao, 2048, 0, 2048},

MobileGL/MG_Impl/GLImpl/Buffer/GL_Buffer.cpp

@@ -9,6 +9,7 @@
 #include "GL_Buffer.h"
 #include "Validators.h"
 #include "../Texture/GL_Texture.h"
+#include "../Getter/GL_Getter.h"
 #include <MG_Util/Converters/GLToMG/TextureEnumConverter.h>
 #include <MG_Util/Metrics/TextureMetrics.h>
 #include <Config.h>
@@ -861,6 +862,10 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         Range1D mappedRange = bufferObject->GetMappedRange();
         auto mappingAccess = bufferObject->GetMappingAccess();
+        // GL 4.6 6.5: the error is on OVERLAP with the mapped range, i.e. a half-open
+        // intersection test. There used to be a second test below this one asking only
+        // `offset + size >= mappedRange.start`, which rejects every write that starts
+        // before a mapped tail as well - it made a legal disjoint glBufferSubData fail.
         if (bufferObject->IsMapped() && !(mappingAccess & BufferMappingAccessBit::Persistent) &&
             (offset < mappedRange.end) && (offset + size > mappedRange.start)) {
             MG_State::pGLContext->RecordError(
@@ -871,18 +876,6 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        if (bufferObject->IsMapped() && !(mappingAccess & BufferMappingAccessBit::Persistent)) {
-            Range1D mappedRange = bufferObject->GetMappedRange();
-            if (offset + size >= mappedRange.start) {
-                MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidOperation,
-                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "BufferSubData_State",
-                                                 "Cannot modify a mapped buffer object unless it was "
-                                                 "mapped with GL_MAP_PERSISTENT_BIT."));
-                return;
-            }
-        }
-
         bufferObject->UploadSubData({(void*)data, (SizeT)size}, offset);
     }
 
@@ -1013,6 +1006,11 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void BufferStorage_State(GLenum target, GLsizeiptr size, const void* data, GLbitfield flags) {
+        // Error precedence: "no buffer is bound to target" outranks a bad size or bad
+        // flags, so the binding has to be resolved before either is validated.
+        auto bufferObject = GetBoundBufferObject(target, BufferOp::BufferStorage);
+        if (!bufferObject) return;
+
         if (size <= 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
@@ -1021,8 +1019,6 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (!ValidateStorageFlags(flags, BufferOp::BufferStorage)) return;
 
-        auto bufferObject = GetBoundBufferObject(target, BufferOp::BufferStorage);
-        if (!bufferObject) return;
         if (bufferObject->IsImmutableStorage()) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
@@ -1057,6 +1053,10 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void NamedBufferStorage_State(GLuint buffer, GLsizeiptr size, const void* data, GLbitfield flags) {
+        // Same precedence as BufferStorage_State: the buffer-name error comes first.
+        auto bufferObject = GetNamedBufferObject(buffer, BufferOp::NamedBufferStorage);
+        if (!bufferObject) return;
+
         if (size <= 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
@@ -1065,8 +1065,6 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (!ValidateStorageFlags(flags, BufferOp::NamedBufferStorage)) return;
 
-        auto bufferObject = GetNamedBufferObject(buffer, BufferOp::NamedBufferStorage);
-        if (!bufferObject) return;
         if (bufferObject->IsImmutableStorage()) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
@@ -1486,12 +1484,71 @@ namespace MobileGL::MG_Impl::GLImpl {
         GetBufferBindingSlot(bufferTarget).Bind(bufferObject);
     }
 
+    // GL 4.6 core 6.1.1: the constraints glBindBufferRange puts on the (offset, size) pair.
+    // Every one of them is INVALID_VALUE, and all of them are checked before a single piece
+    // of state is written - a rejected bind must leave the binding point exactly as it was.
+    // They apply only to a non-zero buffer: buffer 0 detaches the binding point and ignores
+    // offset and size, which is also how glBindBuffersRange spells "reset this element"
+    // (a NULL buffers array, or a zero entry inside one).
+    static Bool ValidateBufferRangeOffsetAndSize(GLenum target, GLintptr offset, GLsizeiptr size,
+                                                 const char* funcName) {
+        if (size <= 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                             std::format("size ({}) must be greater than zero.", size)));
+            return false;
+        }
+        if (offset < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                             std::format("offset ({}) must not be negative.", offset)));
+            return false;
+        }
+        // GL_UNIFORM_BUFFER and GL_SHADER_STORAGE_BUFFER each constrain the offset to their own
+        // implementation-defined alignment, which glGetIntegerv already answers.
+        GLenum alignmentQuery = GL_NONE;
+        if (target == GL_SHADER_STORAGE_BUFFER) {
+            alignmentQuery = GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT;
+        } else if (target == GL_UNIFORM_BUFFER) {
+            alignmentQuery = GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT;
+        }
+        if (alignmentQuery != GL_NONE) {
+            GLint alignment = 0;
+            GetIntegerv(alignmentQuery, &alignment);
+            if (alignment > 0 && (offset % static_cast<GLintptr>(alignment)) != 0) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", funcName,
+                        std::format("offset ({}) must be a multiple of {} ({}).", offset,
+                                    MG_Util::ConvertGLEnumToString(alignmentQuery), alignment)));
+                return false;
+            }
+        }
+        // A transform feedback capture binding is addressed in 32-bit components, so BOTH the
+        // offset and the size must be multiples of 4.
+        if (target == GL_TRANSFORM_FEEDBACK_BUFFER && ((offset % 4) != 0 || (size % 4) != 0)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", funcName,
+                    std::format("offset ({}) and size ({}) must both be multiples of 4 for "
+                                "GL_TRANSFORM_FEEDBACK_BUFFER.",
+                                offset, size)));
+            return false;
+        }
+        return true;
+    }
+
     void BindBufferRange_State(GLenum target, GLuint index, GLuint buffer, GLintptr offset, GLsizeiptr size) {
         MGLOG_D("%s: target = %s, index = %u, buffer = %u, offset = %d, size = %d", __func__,
                 MG_Util::ConvertGLEnumToString(target).c_str(), index, buffer, offset, size);
         BufferTarget bufferTarget = MG_Util::ConvertGLEnumToBufferTarget(target);
         if (!BufferImpl::ValidateBufferBindingPointTarget(bufferTarget)) return;
         if (!BufferImpl::ValidateBufferBindingPointIndex(bufferTarget, index)) return;
+        if (buffer != 0 && !ValidateBufferRangeOffsetAndSize(target, offset, size, __func__)) return;
         if (bufferTarget == BufferTarget::TransformFeedback && MG_State::pGLContext->IsTransformFeedbackActive()) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
@@ -1665,15 +1722,32 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     // ARB_multi_bind: defined by the spec as equivalent to a loop over the single-bind entry
-    // points (with buffer 0 resetting the binding point).
+    // points (with buffer 0 resetting the binding point) - but only AFTER an up-front check
+    // of the whole [first, first + count) range. Looping straight into the single-bind entry
+    // points reports the single-bind INVALID_VALUE for an out-of-range index instead of the
+    // multi-bind INVALID_OPERATION, and binds the in-range prefix before failing.
+    static Bool ValidateMultiBindBufferRange(GLenum target, GLuint first, GLsizei count, const char* funcName) {
+        BufferTarget bufferTarget = MG_Util::ConvertGLEnumToBufferTarget(target);
+        if (!BufferImpl::ValidateBufferBindingPointTarget(bufferTarget)) return false;
+        return BufferImpl::ValidateBufferBindingPointRange(bufferTarget, first, count, funcName);
+    }
+
     void BindBuffersBase(GLenum target, GLuint first, GLsizei count, const GLuint* buffers) {
+        if (!ValidateMultiBindBufferRange(target, first, count, __func__)) return;
         for (GLsizei i = 0; i < count; ++i) {
             BindBufferBase_State(target, first + i, buffers ? buffers[i] : 0);
         }
     }
 
+    // The (offset, size) constraints are the one part of glBindBuffersRange that stays
+    // per-element: ARB_multi_bind checks them separately for each binding point, leaves that
+    // point unchanged on failure, and still applies the remaining elements - which is exactly
+    // what looping into BindBufferRange_State does. Only the [first, first + count) range is
+    // an up-front, all-or-nothing check. Elements that name buffer 0 (or a NULL buffers array)
+    // reset the binding point through BindBufferBase_State and carry no offset/size to check.
     void BindBuffersRange(GLenum target, GLuint first, GLsizei count, const GLuint* buffers, const GLintptr* offsets,
                           const GLsizeiptr* sizes) {
+        if (!ValidateMultiBindBufferRange(target, first, count, __func__)) return;
         for (GLsizei i = 0; i < count; ++i) {
             if (!buffers || buffers[i] == 0) {
                 BindBufferBase_State(target, first + i, 0);

MobileGL/MG_Impl/GLImpl/Buffer/Validators.cpp

@@ -53,18 +53,46 @@ namespace MobileGL::MG_Impl::GLImpl::BufferImpl {
         return true;
     }
 
+    namespace {
+        // The GL-visible number of indexed binding points for `target`.
+        SizeT GetBufferBindingPointLimit(BufferTarget target) {
+            SizeT pointCount = MG_State::pGLContext->GetBufferBindingPointCount(target);
+            if (target == BufferTarget::ShaderStorage && MG_Backend::pActiveBackendObject) {
+                const Int backendCount =
+                    MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxShaderStorageBufferBindings;
+                pointCount = std::min(pointCount, static_cast<SizeT>(std::max(backendCount, 0)));
+            }
+            if (target == BufferTarget::TransformFeedback) {
+                // GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_ATTRIBS bounds the indexed capture
+                // binding points in GL 3.3 (no ARB_transform_feedback3).
+                pointCount = std::min<SizeT>(pointCount, 4);
+            }
+            return pointCount;
+        }
+    } // namespace
+
+    Bool ValidateBufferBindingPointRange(BufferTarget target, Uint first, GLsizei count, const char* funcName) {
+        if (count < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue, MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl/BufferImpl", funcName,
+                                                                      "count must be non-negative."));
+            return false;
+        }
+        const SizeT pointCount = GetBufferBindingPointLimit(target);
+        if (static_cast<Uint64>(first) + static_cast<Uint64>(count) > static_cast<Uint64>(pointCount)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl/BufferImpl", funcName,
+                    std::format("first + count ({} + {}) exceeds the {} indexed binding points of target {}.", first,
+                                count, pointCount, MG_Util::ConvertBufferTargetToString(target))));
+            return false;
+        }
+        return true;
+    }
+
     Bool ValidateBufferBindingPointIndex(BufferTarget target, Uint index) {
-        SizeT pointCount = MG_State::pGLContext->GetBufferBindingPointCount(target);
+        const SizeT pointCount = GetBufferBindingPointLimit(target);
-        if (target == BufferTarget::ShaderStorage && MG_Backend::pActiveBackendObject) {
-            const Int backendCount =
-                MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxShaderStorageBufferBindings;
-            pointCount = std::min(pointCount, static_cast<SizeT>(std::max(backendCount, 0)));
-        }
-        if (target == BufferTarget::TransformFeedback) {
-            // GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_ATTRIBS bounds the indexed capture
-            // binding points in GL 3.3 (no ARB_transform_feedback3).
-            pointCount = std::min<SizeT>(pointCount, 4);
-        }
 
         if (index < pointCount) {
             return true;

MobileGL/MG_Impl/GLImpl/Buffer/Validators.h

@@ -17,4 +17,8 @@ namespace MobileGL::MG_Impl::GLImpl::BufferImpl {
     Bool ValidateBufferMappingAccess(Flags<BufferMappingAccessBit> accessBits);
     Bool ValidateBufferBindingPointTarget(BufferTarget target);
     Bool ValidateBufferBindingPointIndex(BufferTarget target, Uint index);
+    // ARB_multi_bind: glBindBuffersBase/Range validate the whole [first, first + count) range
+    // up front and report INVALID_OPERATION, where a single out-of-range index would be
+    // INVALID_VALUE. Naively looping the single-bind entry points reports the wrong class.
+    Bool ValidateBufferBindingPointRange(BufferTarget target, Uint first, GLsizei count, const char* funcName);
 } // namespace MobileGL::MG_Impl::GLImpl::BufferImpl

MobileGL/MG_Impl/GLImpl/Drawing/GL_Drawing.cpp

@@ -474,6 +474,21 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
         if (!ValidateCurrentProgramForCompute(__func__)) return;
+        // GL 4.6 core 19: each num_groups_* must be within GL_MAX_COMPUTE_WORK_GROUP_COUNT
+        // for its dimension. GetIntegeri_v already floors that at the spec minimum.
+        const GLuint numGroups[3] = {numGroupsX, numGroupsY, numGroupsZ};
+        for (GLuint dimension = 0; dimension < 3; ++dimension) {
+            GLint maxGroups = 0;
+            GetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_COUNT, dimension, &maxGroups);
+            if (numGroups[dimension] > static_cast<GLuint>(std::max(maxGroups, 0))) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                                 "num_groups exceeds GL_MAX_COMPUTE_WORK_GROUP_COUNT for dimension " +
+                                                     std::to_string(dimension) + "."));
+                return;
+            }
+        }
         dispatchCompute(numGroupsX, numGroupsY, numGroupsZ);
     }
 
@@ -487,6 +502,24 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
         if (!ValidateCurrentProgramForCompute(__func__)) return;
+        // GL 4.6 core 19: `indirect` is a byte offset into GL_DISPATCH_INDIRECT_BUFFER -
+        // negative or misaligned is INVALID_VALUE, nothing bound is INVALID_OPERATION.
+        if (indirect < 0 || (indirect % 4) != 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "indirect must be non-negative and a multiple of 4."));
+            return;
+        }
+        const auto& indirectBuffer =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::DispatchIndirect).GetBoundObject();
+        if (!indirectBuffer) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "No buffer is bound to GL_DISPATCH_INDIRECT_BUFFER."));
+            return;
+        }
         dispatchComputeIndirect(indirect);
     }
 

MobileGL/MG_Impl/GLImpl/Exporting/Definitions.cpp

@@ -977,8 +977,8 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexed, GLuint index, GLint left, GL
 DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexedv, index, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeArrayv, first, count, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, GetFloati_v, GLenum target, GLuint index, GLfloat* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetFloati_v, target, index, data)
+DECLARE_GL_FUNCTION_HEAD(void, GetFloati_v, GLenum target, GLuint index, GLfloat* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetFloati_v, target, index, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, GetDoublei_v, GLenum target, GLuint index, GLdouble* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetDoublei_v, target, index, data)
+DECLARE_GL_FUNCTION_HEAD(void, GetDoublei_v, GLenum target, GLuint index, GLdouble* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetDoublei_v, target, index, data)
 DECLARE_GL_FUNCTION_HEAD(void, DrawArraysInstancedBaseInstance, GLenum mode, GLint first, GLsizei count, GLsizei instancecount, GLuint baseinstance) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawArraysInstancedBaseInstance, mode, first, count, instancecount, baseinstance)
 DECLARE_GL_FUNCTION_HEAD(void, DrawElementsInstancedBaseInstance, GLenum mode, GLsizei count, GLenum type, const void* indices, GLsizei instancecount, GLuint baseinstance) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawElementsInstancedBaseInstance, mode, count, type, indices, instancecount, baseinstance)
 DECLARE_GL_FUNCTION_HEAD(void, DrawElementsInstancedBaseVertexBaseInstance, GLenum mode, GLsizei count, GLenum type, const void* indices, GLsizei instancecount, GLint basevertex, GLuint baseinstance) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawElementsInstancedBaseVertexBaseInstance, mode, count, type, indices, instancecount, basevertex, baseinstance)
@@ -1273,7 +1273,7 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, MultiTexCoord4ivARB, GLenum target, const GL
 DECLARE_GL_FUNCTION_STUB_HEAD(void, MultiTexCoord4sARB, GLenum target, GLshort s, GLshort t, GLshort r, GLshort q) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, MultiTexCoord4sARB, target, s, t, r, q)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, MultiTexCoord4svARB, GLenum target, const GLshort* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, MultiTexCoord4svARB, target, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetQueryObjectivARB, GLuint id, GLenum pname, GLint* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetQueryObjectivARB, id, pname, params)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, MaxShaderCompilerThreadsARB, GLuint count) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, MaxShaderCompilerThreadsARB, count)
+DECLARE_GL_FUNCTION_HEAD(void, MaxShaderCompilerThreadsARB, GLuint count) DECLARE_GL_FUNCTION_END_NO_RETURN(void, MaxShaderCompilerThreadsARB, count)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, PointParameterfARB, GLenum pname, GLfloat param) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, PointParameterfARB, pname, param)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, PointParameterfvARB, GLenum pname, const GLfloat* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, PointParameterfvARB, pname, params)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetnTexImageARB, GLenum target, GLint level, GLenum format, GLenum type, GLsizei bufSize, void* img) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetnTexImageARB, target, level, format, type, bufSize, img)
@@ -1381,7 +1381,7 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, WindowPos3ivARB, const GLint* v) DECLARE_GL_
 DECLARE_GL_FUNCTION_STUB_HEAD(void, WindowPos3sARB, GLshort x, GLshort y, GLshort z) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, WindowPos3sARB, x, y, z)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, WindowPos3svARB, const GLshort* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, WindowPos3svARB, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, BlendBarrierKHR, void) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, BlendBarrierKHR, )
-DECLARE_GL_FUNCTION_STUB_HEAD(void, MaxShaderCompilerThreadsKHR, GLuint count) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, MaxShaderCompilerThreadsKHR, count)
+DECLARE_GL_FUNCTION_HEAD(void, MaxShaderCompilerThreadsKHR, GLuint count) DECLARE_GL_FUNCTION_END_NO_RETURN(void, MaxShaderCompilerThreadsKHR, count)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, MultiTexCoord1bOES, GLenum texture, GLbyte s) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, MultiTexCoord1bOES, texture, s)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, MultiTexCoord1bvOES, GLenum texture, const GLbyte* coords) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, MultiTexCoord1bvOES, texture, coords)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, MultiTexCoord2bOES, GLenum texture, GLbyte s, GLbyte t) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, MultiTexCoord2bOES, texture, s, t)

MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp

@@ -106,9 +106,11 @@ namespace MobileGL::MG_Impl::GLImpl {
         // `capabilityTargetIndex` is the row of the cache the attachment actually lives in;
         // kFormatCapabilityTargetCount asks about the format in general. Asking per target matters
         // because a capability recorded for one of them says nothing about the others: DirectGLES
-        // widens three-channel formats to four channels to keep them renderable as *multisample*
+        // decides each target's substitution against that target's own probe, and a buffer texture
-        // storage, and a format that survives only through that substitution is still texture-only
+        // never gets one at all. This is also where the three-channel widening becomes visible to
-        // on every ordinary target.
+        // the application - a GL_RGB8_SNORM colour attachment on a driver with no renderable
+        // three-channel format answers COMPLETE because the backend stores it as GL_RGBA16F and
+        // recorded FramebufferRenderable in CaveatCaps.
         Bool IsColorInternalFormatRenderable(TextureInternalFormat format, SizeT capabilityTargetIndex) {
             const SizeT formatIndex = static_cast<SizeT>(format);
             if (MG_Backend::pActiveBackendObject && formatIndex < MG_Backend::kFormatCapabilityFormatCount) {

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -23,6 +23,7 @@
 #include <MG_Util/Converters/MGToGL/RenderStateEnumConverter.h>
 #include <MG_State/GLState/FramebufferState/FramebufferObject.h>
 #include <MG_Util/Texture/TextureFormatProcessor.h>
+#include <MG_Util/Async/ShaderCompilePool.h>
 #include <MG_Backend/BackendObjects.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
@@ -50,6 +51,14 @@ namespace MobileGL::MG_Impl::GLImpl {
         constexpr GLint kFrontendMaxTessControlAtomicCounters = 0;
         constexpr GLint kFrontendMaxTessEvaluationAtomicCounters = 0;
         constexpr GLint kFrontendMaxVertexAtomicCounters = 0;
+        // One atomic counter is a uint, and a buffer never has to hold more counters than the
+        // combined limit the frontend advertises. GL 4.6 table 23.63 floors this at 32 bytes.
+        constexpr GLint kFrontendMaxAtomicCounterBufferSize =
+            kFrontendMaxCombinedAtomicCounters * static_cast<GLint>(sizeof(GLuint));
+        // KHR_debug minima (GL 4.6 table 23.66); the debug entry points are stubs, but the
+        // limits they advertise still have to be legal.
+        constexpr GLint kFrontendMaxDebugGroupStackDepth = 64;
+        constexpr GLint kFrontendMaxDebugLoggedMessages = 1;
         constexpr GLint kFrontendMaxVertexUniformComponents = 4096;
         constexpr GLint kFrontendMaxVertexUniformVectors = 128;
         constexpr GLint kFrontendMaxVertexUniformBlocks = 14;
@@ -263,6 +272,60 @@ namespace MobileGL::MG_Impl::GLImpl {
             return true;
         }
 
+        // GL_TEXTURE_BINDING_* is per-texture-unit state: glGetIntegerv answers for the
+        // active unit, glGetIntegeri_v answers for unit `index`. Both need the same
+        // pname -> target decode, so it lives here instead of being spelled out twice.
+        bool TryDecodeTextureUnitBindingPname(GLenum pname, TextureTarget& outTarget) {
+            switch (pname) {
+            case GL_TEXTURE_BINDING_1D: outTarget = TextureTarget::Texture1D; return true;
+            case GL_TEXTURE_BINDING_1D_ARRAY: outTarget = TextureTarget::Texture1DArray; return true;
+            case GL_TEXTURE_BINDING_2D: outTarget = TextureTarget::Texture2D; return true;
+            case GL_TEXTURE_BINDING_2D_ARRAY: outTarget = TextureTarget::Texture2DArray; return true;
+            case GL_TEXTURE_BINDING_2D_MULTISAMPLE: outTarget = TextureTarget::Texture2DMultisample; return true;
+            case GL_TEXTURE_BINDING_2D_MULTISAMPLE_ARRAY:
+                outTarget = TextureTarget::Texture2DMultisampleArray;
+                return true;
+            case GL_TEXTURE_BINDING_3D: outTarget = TextureTarget::Texture3D; return true;
+            case GL_TEXTURE_BINDING_BUFFER: outTarget = TextureTarget::TextureBuffer; return true;
+            case GL_TEXTURE_BINDING_CUBE_MAP: outTarget = TextureTarget::TextureCubeMap; return true;
+            case GL_TEXTURE_BINDING_CUBE_MAP_ARRAY: outTarget = TextureTarget::TextureCubeMapArray; return true;
+            case GL_TEXTURE_BINDING_RECTANGLE: outTarget = TextureTarget::TextureRectangle; return true;
+            default: return false;
+            }
+        }
+
+        GLint QueryTextureBindingOnUnit(Int unit, TextureTarget target) {
+            auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
+            const auto& obj = textureUnit.GetBindingSlot(target).GetBoundObject();
+            return obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
+        }
+
+        GLint QuerySamplerBindingOnUnit(Int unit) {
+            const auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
+            const auto& sampler = textureUnit.GetSamplerObject();
+            return sampler ? static_cast<GLint>(sampler->GetExternalIndex()) : 0;
+        }
+
+        // The ARB_viewport_array indexed rectangles. MobileGL keeps exactly one viewport, one
+        // scissor box and one depth range, so every in-range index answers with that single
+        // value - but it has to come from the frontend state the non-indexed getters read.
+        // The generic path at the bottom of GetIntegeri_v is a raw backend passthrough that
+        // has no case for these, so routing them through it returned zeros.
+        Bool IsIndexedViewportQuery(GLenum target) {
+            return target == GL_VIEWPORT || target == GL_SCISSOR_BOX || target == GL_DEPTH_RANGE;
+        }
+
+        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it.
+        Bool ValidateViewportQueryIndex(GLuint index, const char* caller) {
+            GLint maxViewports = 0;
+            GetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+            if (index < static_cast<GLuint>(std::max(maxViewports, 1))) return true;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller, "Viewport index is out of range."));
+            return false;
+        }
+
         void CopyIntsToBooleans(const GLint* src, SizeT count, GLboolean* dst) {
             for (SizeT i = 0; i < count; ++i) {
                 dst[i] = src[i] ? GL_TRUE : GL_FALSE;
@@ -670,7 +733,37 @@ namespace MobileGL::MG_Impl::GLImpl {
             }
         }
 
+        // Per-texture-unit bindings: GL 4.6 core table 23.19 makes every GL_TEXTURE_BINDING_*
+        // and GL_SAMPLER_BINDING indexed by texture unit. Without this they fell through to
+        // the raw backend passthrough at the bottom, which knows nothing about the
+        // frontend's binding state.
+        if (TextureTarget textureBindingTarget = TextureTarget::Unknown;
+            TryDecodeTextureUnitBindingPname(target, textureBindingTarget) || target == GL_SAMPLER_BINDING) {
+            GLint maxUnits = 0;
+            GetIntegerv(GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, &maxUnits);
+            maxUnits = std::min<GLint>(maxUnits, MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS);
+            if (index >= static_cast<GLuint>(std::max(maxUnits, 0))) {
+                *data = 0;
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Texture unit index is out of range."));
+                return;
+            }
+            *data = target == GL_SAMPLER_BINDING
+                ? QuerySamplerBindingOnUnit(static_cast<Int>(index))
+                : QueryTextureBindingOnUnit(static_cast<Int>(index), textureBindingTarget);
+            return;
+        }
+
         switch (target) {
+        // ARB_viewport_array queries the indexed rectangles through glGetIntegeri_v as well
+        // (gl4cMultiBindTests and the viewport_array group both do). The frontend keeps one
+        // viewport and one scissor box, so every in-range index reports that one.
+        case GL_VIEWPORT:
+        case GL_SCISSOR_BOX:
+            if (!ValidateViewportQueryIndex(index, __func__)) return;
+            GetIntegerv(target, data);
+            return;
         // The vertex buffer binding points of the vertex array object that is bound. Indexed by
         // binding point, not by attribute (GL 4.6 core 10.3.1).
         case GL_VERTEX_BINDING_BUFFER:
@@ -782,6 +875,46 @@ namespace MobileGL::MG_Impl::GLImpl {
         getIntegeri(target, index, data);
     }
 
+    // GL_ARB_viewport_array's typed indexed getters. They were no-op stubs, which left the
+    // caller's output buffer holding whatever was on the stack. The multi-component indexed
+    // rectangles are answered from the frontend's own viewport/scissor/depth-range state, via
+    // the non-indexed getter of the matching type - GL_DEPTH_RANGE is float state, so putting
+    // it through the integer query would round it to 0/1. Everything else MobileGL answers
+    // indexed is scalar integer-domain state, where converting the integer query is exact.
+    void GetFloati_v(GLenum target, GLuint index, GLfloat* data) {
+        if (!data) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "data pointer cannot be null"));
+            return;
+        }
+        if (IsIndexedViewportQuery(target)) {
+            if (!ValidateViewportQueryIndex(index, __func__)) return;
+            GetFloatv(target, data);
+            return;
+        }
+        GLint ints[4] = {};
+        GetIntegeri_v(target, index, ints);
+        data[0] = static_cast<GLfloat>(ints[0]);
+    }
+
+    void GetDoublei_v(GLenum target, GLuint index, GLdouble* data) {
+        if (!data) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "data pointer cannot be null"));
+            return;
+        }
+        if (IsIndexedViewportQuery(target)) {
+            if (!ValidateViewportQueryIndex(index, __func__)) return;
+            GetDoublev(target, data);
+            return;
+        }
+        GLint ints[4] = {};
+        GetIntegeri_v(target, index, ints);
+        data[0] = static_cast<GLdouble>(ints[0]);
+    }
+
     void GetInteger64i_v(GLenum target, GLuint index, GLint64* data) {
         if (!data) {
             MG_State::pGLContext->RecordError(
@@ -958,6 +1091,13 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // Per-texture-unit bindings: the non-indexed query reports the active unit.
+        if (TextureTarget textureBindingTarget = TextureTarget::Unknown;
+            TryDecodeTextureUnitBindingPname(pname, textureBindingTarget)) {
+            *params = QueryTextureBindingOnUnit(MG_State::pGLContext->GetActiveTextureUnit(), textureBindingTarget);
+            return;
+        }
+
         switch (pname) {
         case GL_ACTIVE_TEXTURE:
             *params = MG_State::pGLContext->GetActiveTextureUnit() + GL_TEXTURE0;
@@ -1069,12 +1209,32 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
             return;
         }
+        case GL_MAX_SHADER_COMPILER_THREADS_KHR:
+            // GL_KHR_parallel_shader_compile (GL_MAX_SHADER_COMPILER_THREADS_ARB is the same
+            // 0x91B0). The number of threads MobileGL's compile pool would actually use, so
+            // an application sizing its own submission batches gets a real answer.
+            //
+            // Zero when asynchronous compilation is off, which is the honest reply and the
+            // one the extension defines for an implementation with no compiler threads: the
+            // extension string is withdrawn in that configuration too, so a conforming
+            // application never reaches this query, and one that asks anyway is told there
+            // are none rather than being handed a thread count nothing will use.
+            *params = MG_Util::Async::AsyncShaderCompileEnabled()
+                          ? static_cast<GLint>(MG_Util::Async::ShaderCompilePool::Get().GetThreadCount())
+                          : 0;
+            return;
         case GL_MAX_DEBUG_GROUP_STACK_DEPTH:
-            *params = 0; // debug-group entrypoints are stubbed
+            // KHR_debug floors this at 64 even when the group entry points are stubs: the
+            // limit describes how deep glPushDebugGroup may nest, and 0 is not a legal answer.
+            *params = kFrontendMaxDebugGroupStackDepth;
             return;
         case GL_MAX_DEBUG_MESSAGE_LENGTH:
             *params = 1024; // debug-message entrypoints are stubbed, but KHR_debug requires a valid limit
             return;
+        case GL_MAX_DEBUG_LOGGED_MESSAGES:
+            // Size of the message log ring; KHR_debug requires at least 1.
+            *params = kFrontendMaxDebugLoggedMessages;
+            return;
         case GL_DEBUG_GROUP_STACK_DEPTH:
             *params = 0; // debug-group entrypoints are stubbed
             return;
@@ -1516,13 +1676,9 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_SAMPLE_MASK_VALUE:
             *params = static_cast<GLint>(MG_State::pGLContext->GetSampleMaskValue());
             return;
-        case GL_SAMPLER_BINDING: {
+        case GL_SAMPLER_BINDING:
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
+            *params = QuerySamplerBindingOnUnit(MG_State::pGLContext->GetActiveTextureUnit());
-            const auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& sampler = tu.GetSamplerObject();
-            *params = sampler ? static_cast<GLint>(sampler->GetExternalIndex()) : 0;
             return;
-        }
         case GL_SAMPLES:
             *params = ResolveDrawFramebufferSampleCount();
             return;
@@ -1618,87 +1774,6 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_STEREO:
             *params = 0; // stereo surfaces are not exposed
             return;
-        case GL_TEXTURE_BINDING_1D: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::Texture1D);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_1D_ARRAY: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::Texture1DArray);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_2D: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::Texture2D);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            MGLOG_D("Get GL_TEXTURE_BINDING_2D: %d", *params);
-            return;
-        }
-        case GL_TEXTURE_BINDING_2D_ARRAY: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::Texture2DArray);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_2D_MULTISAMPLE: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::Texture2DMultisample);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_2D_MULTISAMPLE_ARRAY: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::Texture2DMultisampleArray);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_3D: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::Texture3D);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_BUFFER: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::TextureBuffer);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_CUBE_MAP: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::TextureCubeMap);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
-        case GL_TEXTURE_BINDING_RECTANGLE: {
-            Int unit = MG_State::pGLContext->GetActiveTextureUnit();
-            auto& tu = MG_State::pGLContext->GetTextureUnitObject(unit);
-            const auto& slot = tu.GetBindingSlot(TextureTarget::TextureRectangle);
-            const auto& obj = slot.GetBoundObject();
-            *params = obj ? static_cast<GLint>(obj->GetExternalIndex()) : 0;
-            return;
-        }
         case GL_TEXTURE_COMPRESSION_HINT:
             *params = static_cast<GLint>(MG_State::pGLContext->GetHint(pname));
             return;
@@ -1971,6 +2046,18 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS:
             *params = static_cast<GLint>(GetIndexedBufferQueryPointCount(BufferTarget::ShaderStorage));
             break;
+        case GL_MAX_SHADER_STORAGE_BLOCK_SIZE:
+            // 64-bit state (see GetInteger64v); the 32-bit query saturates, per the GL
+            // state-query conversion rules.
+            *params = static_cast<GLint>(std::min<Uint64>(dynamicParameters.MaxShaderStorageBlockSize,
+                                                          static_cast<Uint64>(INT32_MAX)));
+            break;
+        case GL_MAX_ATOMIC_COUNTER_BUFFER_BINDINGS:
+            *params = static_cast<GLint>(GetIndexedBufferQueryPointCount(BufferTarget::AtomicCounter));
+            break;
+        case GL_MAX_ATOMIC_COUNTER_BUFFER_SIZE:
+            *params = kFrontendMaxAtomicCounterBufferSize;
+            break;
         case GL_MAX_TEXTURE_BUFFER_SIZE:
             *params = dynamicParameters.MaxTextureBufferSize;
             break;

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.h

@@ -19,6 +19,8 @@ namespace MobileGL::MG_Impl::GLImpl {
     void GetIntegerv(GLenum pname, GLint* params);
     void GetInteger64v(GLenum pname, GLint64* params);
     void GetIntegeri_v(GLenum target, GLuint index, GLint* data);
+    void GetFloati_v(GLenum target, GLuint index, GLfloat* data);
+    void GetDoublei_v(GLenum target, GLuint index, GLdouble* data);
     void GetInteger64i_v(GLenum target, GLuint index, GLint64* data);
     GLenum GetError();
     GLenum GetGraphicsResetStatus();

MobileGL/MG_Impl/GLImpl/Program/GL_Program.cpp

@@ -7,6 +7,7 @@
 // End of Source File Header
 
 #include "GL_Program.h"
+#include "ProgramInterface.h"
 #include "Config.h"
 #include <cmath>
 #include <limits>
@@ -16,6 +17,7 @@
 #include <MG_Util/Converters/GLToMG/ProgramEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/ProgramEnumConverter.h>
 #include <MG_Util/Converters/SPIRVCrossToGL/SpvcTypeConverter.h>
+#include <MG_Util/Async/ShaderCompilePool.h>
 #include <MG_Backend/BackendObjects.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
@@ -108,32 +110,37 @@ namespace MobileGL::MG_Impl::GLImpl {
         return programObject;
     }
 
-    static bool IsProgramInterfaceEnum(GLenum programInterface) {
+    // The four non-location interface queries validate the NAME only: GL 4.6 imposes the
-        switch (programInterface) {
+    // successful-link requirement on GetProgramResourceLocation/LocationIndex alone, and
-        case GL_UNIFORM:
+    // requires the others to report a program that has never linked as one with zero active
-        case GL_UNIFORM_BLOCK:
+    // resources. Being stricter leaves a stray GL_INVALID_OPERATION behind that aborts the
-        case GL_PROGRAM_INPUT:
+    // caller's next subcase.
-        case GL_PROGRAM_OUTPUT:
+    static const SharedPtr<MG_State::GLState::ProgramObject>& TryToGetProgramForInterfaceQuery(GLuint program,
-        case GL_BUFFER_VARIABLE:
+                                                                                               const char* caller) {
-        case GL_SHADER_STORAGE_BLOCK:
+        static const SharedPtr<MG_State::GLState::ProgramObject> nullProgramObject = nullptr;
-        case GL_ATOMIC_COUNTER_BUFFER:
+        if (!MG_State::pGLContext->ValidateProgramName(program)) {
-        case GL_TRANSFORM_FEEDBACK_VARYING:
+            const ErrorCode error = MG_State::pGLContext->ValidateShaderName(program)
-        case GL_VERTEX_SUBROUTINE:
+                ? ErrorCode::InvalidOperation
-        case GL_TESS_CONTROL_SUBROUTINE:
+                : ErrorCode::InvalidValue;
-        case GL_TESS_EVALUATION_SUBROUTINE:
+            MG_State::pGLContext->RecordError(
-        case GL_GEOMETRY_SUBROUTINE:
+                error,
-        case GL_FRAGMENT_SUBROUTINE:
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
-        case GL_COMPUTE_SUBROUTINE:
+                                             std::to_string(program) + " is not a program object."));
-        case GL_VERTEX_SUBROUTINE_UNIFORM:
+            return nullProgramObject;
-        case GL_TESS_CONTROL_SUBROUTINE_UNIFORM:
-        case GL_TESS_EVALUATION_SUBROUTINE_UNIFORM:
-        case GL_GEOMETRY_SUBROUTINE_UNIFORM:
-        case GL_FRAGMENT_SUBROUTINE_UNIFORM:
-        case GL_COMPUTE_SUBROUTINE_UNIFORM:
-            return true;
-        default:
-            return false;
         }
+        auto& programObject = MG_State::pGLContext->GetProgramObject(program);
+        if (!programObject) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             std::to_string(program) + " is not a program object."));
+            return nullProgramObject;
+        }
+        return programObject;
+    }
+
+    static bool IsProgramInterfaceEnum(GLenum programInterface) {
+        return ProgramInterface::IsInterfaceEnum(programInterface);
     }
 
     static bool IsSubroutineUniformInterface(GLenum programInterface) {
@@ -156,11 +163,16 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_ACTIVE_RESOURCES:
             break;
         case GL_MAX_NAME_LENGTH:
-            valid = valid && programInterface != GL_ATOMIC_COUNTER_BUFFER;
+            // Neither buffer interface has resource names. GL_TRANSFORM_FEEDBACK_BUFFER only
+            // became reachable here when IsInterfaceEnum grew the GL 4.4 interfaces, so it
+            // needs the same exclusion GL_ATOMIC_COUNTER_BUFFER already had.
+            valid = valid && programInterface != GL_ATOMIC_COUNTER_BUFFER &&
+                    programInterface != GL_TRANSFORM_FEEDBACK_BUFFER;
             break;
         case GL_MAX_NUM_ACTIVE_VARIABLES:
             valid = programInterface == GL_UNIFORM_BLOCK || programInterface == GL_ATOMIC_COUNTER_BUFFER ||
-                    programInterface == GL_SHADER_STORAGE_BLOCK;
+                    programInterface == GL_SHADER_STORAGE_BLOCK ||
+                    programInterface == GL_TRANSFORM_FEEDBACK_BUFFER;
             break;
         case GL_MAX_NUM_COMPATIBLE_SUBROUTINES:
             valid = IsSubroutineUniformInterface(programInterface);
@@ -179,7 +191,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     static bool ValidateNamedProgramResourceInterface(GLenum programInterface, const char* caller) {
-        if (!IsProgramInterfaceEnum(programInterface) || programInterface == GL_ATOMIC_COUNTER_BUFFER) {
+        if (!ProgramInterface::IsNamedInterface(programInterface)) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
@@ -189,66 +201,6 @@ namespace MobileGL::MG_Impl::GLImpl {
         return true;
     }
 
-    static Int GetKnownProgramResourceCount(const SharedPtr<MG_State::GLState::ProgramObject>& programObject,
-                                            GLenum programInterface) {
-        switch (programInterface) {
-        case GL_UNIFORM:
-            return programObject->GetUniformCount();
-        case GL_UNIFORM_BLOCK:
-            return programObject->GetActiveUniformBlocksCount();
-        case GL_PROGRAM_INPUT:
-            return programObject->GetActiveAttributesCount();
-        case GL_PROGRAM_OUTPUT:
-            return programObject->GetActiveFragmentOutputCount();
-        default:
-            return -1;
-        }
-    }
-
-    // The GL_UNIFORM interface and glGetActiveUniform(s)iv are the same query in two
-    // spellings, so they answer from the same place - the frontend reflection. The backend
-    // program is not that place: it does not exist at all for a program whose types its
-    // shading language cannot express (a double-precision uniform has no ESSL form), and
-    // the interface queries would then describe a program with no uniforms.
-    //
-    // Writes the GL_UNIFORM value of `prop` for active uniform `index`; false for a prop
-    // the reflection does not model, which the caller forwards to the backend instead.
-    Bool GetUniformResourceProp(const SharedPtr<MG_State::GLState::ProgramObject>& programObject, Uint index,
-                                GLenum prop, GLint* out) {
-        switch (prop) {
-        case GL_TYPE:
-            *out = static_cast<GLint>(programObject->GetActiveUniformType(index));
-            return true;
-        case GL_ARRAY_SIZE:
-            *out = programObject->GetActiveUniformArraySize(index);
-            return true;
-        case GL_NAME_LENGTH:
-            *out = static_cast<GLint>(programObject->GetActiveUniformName(index).length() + 1);
-            return true;
-        case GL_BLOCK_INDEX:
-            *out = programObject->GetActiveUniformBlockIndex(index);
-            return true;
-        case GL_OFFSET:
-            *out = programObject->GetActiveUniformOffset(index);
-            return true;
-        case GL_ARRAY_STRIDE:
-            *out = programObject->GetActiveUniformArrayStride(index);
-            return true;
-        case GL_MATRIX_STRIDE:
-            *out = programObject->GetActiveUniformMatrixStride(index);
-            return true;
-        case GL_IS_ROW_MAJOR:
-            *out = programObject->GetActiveUniformIsRowMajor(index);
-            return true;
-        case GL_LOCATION:
-            // A block member has no location; GetUniformLocation already reports -1 for one.
-            *out = programObject->GetUniformLocation(programObject->GetActiveUniformName(index));
-            return true;
-        default:
-            return false;
-        }
-    }
-
     void CopyStr(GLsizei bufSize, GLsizei* length, GLchar* dst, const char* src, GLsizei srcLength) {
         if (bufSize <= 0) {
             if (length) *length = 0;
@@ -355,19 +307,72 @@ namespace MobileGL::MG_Impl::GLImpl {
         shaderObject->Compile();
     }
 
+    // glMaxShaderCompilerThreadsKHR / glMaxShaderCompilerThreadsARB - one implementation,
+    // because GL_KHR_parallel_shader_compile and GL_ARB_parallel_shader_compile define the
+    // same entry point with the same semantics and GetProcAddress.cpp maps both spellings.
+    //
+    // The three cases the extension defines, and what each means here:
+    //
+    //   count == 0            "no compiler threads": compilation must happen on the
+    //                         application's thread. Everything already in flight is joined
+    //                         first, so that after this call returns NOTHING is outstanding
+    //                         and every GL_COMPLETION_STATUS_KHR reads GL_TRUE - which is
+    //                         the observable the extension actually specifies. The pool
+    //                         keeps its worker threads (this is not teardown); what changes
+    //                         is that AsyncShaderCompileActive() now says no, so
+    //                         glCompileShader/glLinkProgram run their bodies inline.
+    //   count == 0xFFFFFFFF   "implementation maximum": the pool's full thread count.
+    //   otherwise             a concurrency budget, clamped to the thread count - asking for
+    //                         more threads than exist cannot conjure any.
+    //
+    // A nonzero count is also what LIFTS a previous zero: the suspension lasts exactly until
+    // the application asks for threads again, and nothing else re-arms it (no implicit
+    // restore at eglInitialize, at a context switch or at a join). An application that turned
+    // compiler threads off keeps them off until it says otherwise.
+    //
+    // Legal - and a no-op beyond bookkeeping - while MOBILEGL_ASYNC_SHADER_COMPILE is off:
+    // compilation is already inline, and the call must not fail just because MobileGL had
+    // nothing to suspend.
+    void MaxShaderCompilerThreadsKHR_State(GLuint count) {
+        namespace Async = MG_Util::Async;
+        if (count == 0) {
+            MGLOG_D("%s: count = 0; joining all pending shader work and compiling inline", __func__);
+            Async::SetAsyncShaderCompileSuspended(true);
+            // Suspend BEFORE joining, not after. The post-condition this call owes the
+            // application is "nothing is in flight when I return", and only this order
+            // guarantees it: with the latch already set, anything the join itself causes to
+            // be compiled runs inline and is therefore already settled when the join ends.
+            // Joining first would leave a window in which a fresh enqueue is still legal.
+            if (MG_State::pGLContext) MG_State::pGLContext->JoinAllPendingShaderWork();
+            return;
+        }
+
+        Async::ShaderCompilePool& pool = Async::ShaderCompilePool::Get();
+        const Uint threadCount = pool.GetThreadCount();
+        const Uint requested = count == 0xFFFFFFFFu ? threadCount : std::min<Uint>(count, threadCount);
+        pool.SetMaxConcurrency(requested);
+        Async::SetAsyncShaderCompileSuspended(false);
+        MGLOG_D("%s: count = %u; concurrency = %u of %u threads", __func__, count, requested, threadCount);
+    }
+
     GLuint CreateProgram_State() {
         return MG_State::pGLContext->CreateProgram();
     }
 
     GLuint CreateShader_State(GLenum type) {
-        auto shaderId = MG_State::pGLContext->CreateShader(MG_Util::ConvertGLEnumToShaderStage(type));
+        // GL 4.6 core 7.1: shaderType is an enum, so an unrecognised one is INVALID_ENUM (it
-        if (shaderId == 0) {
+        // used to be documented as INVALID_VALUE). The check has to happen HERE: the state
+        // layer hands out a name for ShaderStage::Unknown just as happily as for a real
+        // stage, so the old "shaderId == 0 means bad type" test could never fire and an
+        // unknown shaderType silently produced a usable shader name and no error at all.
+        const ShaderStage stage = MG_Util::ConvertGLEnumToShaderStage(type);
+        if (stage == ShaderStage::Unknown) {
             MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidValue,
+                ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "`shaderType` is not an accepted value."));
             return 0;
         }
-        return shaderId;
+        return MG_State::pGLContext->CreateShader(stage);
     }
 
     void DeleteProgram_State(GLuint program) {
@@ -693,6 +698,19 @@ namespace MobileGL::MG_Impl::GLImpl {
             break;
         }
 
+        // GL_KHR_parallel_shader_compile. THIS CASE MUST NOT JOIN - it is the one program
+        // query whose entire purpose is to answer without waiting, and routing it through
+        // any of ProgramObject's Artifacts() accessors (the join gate, invariant I5) would
+        // block the caller and make the extension a lie: an application polling it would
+        // serialize itself on the very link it is trying to overlap. IsLinkComplete() is the
+        // node-direct reader that exists for exactly this.
+        //
+        // No link at all reads GL_TRUE, which is what the extension requires: the query
+        // means "is anything still outstanding", not "has this program ever been linked".
+        case GL_COMPLETION_STATUS_KHR:
+            *params = programObject->IsLinkComplete() ? GL_TRUE : GL_FALSE;
+            break;
+
         case GL_PROGRAM_BINARY_LENGTH:
             // No program binary format is exposed, so a program never has a retrievable
             // binary and its length is zero (ARB_get_program_binary).
@@ -746,6 +764,13 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_SHADER_SOURCE_LENGTH:
             *params = shaderObject->GetShaderSource().empty() ? 0 : (GLint)shaderObject->GetShaderSource().length() + 1;
             break;
+        // GL_KHR_parallel_shader_compile. THIS CASE MUST NOT JOIN - see the identical case in
+        // GetProgramiv_State. GL_COMPILE_STATUS two cases up deliberately DOES join (it has
+        // to: it reports the outcome); this one reports whether there is an outcome yet, and
+        // reading it through Compiled() would defeat the whole extension.
+        case GL_COMPLETION_STATUS_KHR:
+            *params = shaderObject->IsCompileComplete() ? GL_TRUE : GL_FALSE;
+            break;
         default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
@@ -786,6 +811,31 @@ namespace MobileGL::MG_Impl::GLImpl {
         return loc;
     }
 
+    // A float matrix lives in the global UBO under std140 rules - one 16-byte-aligned column
+    // vector per column - while the value glGetUniform* must return is tightly packed
+    // columns * rows floats. Only mat4 is the same either way; every other shape needs the
+    // padding undone, and the readback has to undo exactly what UniformMatrixfv_Object put
+    // there. Returns false when `ttype` is not a float matrix (nothing to unpack).
+    Bool TryGatherFloatMatrixColumns(const glslang::TType* ttype, const char* pBase, void* params) {
+        if (ttype == nullptr || !ttype->isMatrix() || ttype->getBasicType() == glslang::EbtDouble) return false;
+        const Int columns = ttype->getMatrixCols();
+        const Int rows = ttype->getMatrixRows();
+        for (Int column = 0; column < columns; ++column) {
+            Memcpy(static_cast<char*>(params) + static_cast<SizeT>(column) * rows * sizeof(GLfloat),
+                   pBase + static_cast<SizeT>(column) * 4 * sizeof(GLfloat), rows * sizeof(GLfloat));
+        }
+        return true;
+    }
+
+    // Bytes a uniform actually occupies in the global UBO. It is the tight GL type size for
+    // everything except a float matrix, whose padded columns make it wider.
+    SizeT UniformStorageSpanInBytes(const glslang::TType* ttype, SizeT tightSize) {
+        if (ttype != nullptr && ttype->isMatrix() && ttype->getBasicType() != glslang::EbtDouble) {
+            return static_cast<SizeT>(ttype->getMatrixCols()) * 4 * sizeof(GLfloat);
+        }
+        return tightSize;
+    }
+
     void GetUniform_State(GLuint program, GLint location, void* params) {
         auto& programObject = TryToGetProgramObject(program);
         if (!programObject) return;
@@ -816,23 +866,16 @@ namespace MobileGL::MG_Impl::GLImpl {
             auto size = programObject->GetUniformSizesInBytes(location);
             char* pUBO = (char*)programObject->MapUBO();
             auto* ttype = programObject->GetUniformTType(location);
+            const SizeT span = UniformStorageSpanInBytes(ttype, size);
             if (pUBO == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
-                offset + size > programObject->GetUBOSize()) {
+                offset + span > programObject->GetUBOSize()) {
                 MGLOG_E("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
                         program, location);
                 return;
             }
 
-            if (!ttype->isMatrix() || ttype->getMatrixCols() != 3)
+            if (!TryGatherFloatMatrixColumns(ttype, pUBO + offset, params)) {
                 Memcpy(params, pUBO + offset, size);
-            else {
-                // TODO: we only deal with mat3 yet, deal with other types later
-                // assuming float here, which may not be the case
-                auto* pBase = pUBO + offset;
-                for (int i = 0; i < ttype->getMatrixRows(); i++) {
-                    Memcpy((char*)params + ttype->getMatrixCols() * sizeof(float) * i, pBase + 4 * sizeof(float) * i,
-                           ttype->getMatrixCols() * sizeof(float));
-                }
             }
         }
         // TODO: handle 1i variant as texture unit
@@ -871,23 +914,16 @@ namespace MobileGL::MG_Impl::GLImpl {
         auto size = programObject->GetUniformSizesInBytes(location);
         char* pUBO = static_cast<char*>(programObject->MapUBO());
         auto* ttype = programObject->GetUniformTType(location);
+        const SizeT span = UniformStorageSpanInBytes(ttype, size);
         if (pUBO == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
-            offset + size > programObject->GetUBOSize()) {
+            offset + span > programObject->GetUBOSize()) {
             MGLOG_E("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
                     program, location);
             return;
         }
 
         if constexpr (std::is_same_v<T, GLfloat>) {
-            if (ttype->getBasicType() != glslang::EbtDouble && ttype->isMatrix() &&
+            if (TryGatherFloatMatrixColumns(ttype, pUBO + offset, params)) return;
-                ttype->getMatrixCols() == 3) {
-                auto* pBase = pUBO + offset;
-                for (int i = 0; i < ttype->getMatrixRows(); i++) {
-                    Memcpy(reinterpret_cast<char*>(params) + ttype->getMatrixCols() * sizeof(GLfloat) * i,
-                           pBase + 4 * sizeof(GLfloat) * i, ttype->getMatrixCols() * sizeof(GLfloat));
-                }
-                return;
-            }
         }
 
         // A double-precision uniform is the one case where the stored component type can
@@ -1188,6 +1224,52 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
+    // glUniformMatrix*fv / glProgramUniformMatrix*fv, every shape (square and non-square).
+    // A float matrix sits in the global UBO under std140 rules: each of its `columns`
+    // column vectors starts on its own 16-byte boundary no matter how many rows it has, so
+    // the only shape that may be written as one contiguous block is mat4. Writing a matNxM
+    // as N*M packed floats puts every column after the first at the wrong byte offset.
+    template <typename Program>
+    void UniformMatrixfv_Object(Program& programObject, const char* caller, GLint location, GLsizei count,
+                                GLboolean transpose, const GLfloat* value, Int columns, Int rows,
+                                const String& ownerDescription) {
+        // std140: a column vector of a float matrix is padded out to a vec4.
+        constexpr SizeT kColumnStride = 4 * sizeof(GLfloat);
+        const SizeT componentCount = static_cast<SizeT>(columns) * static_cast<SizeT>(rows);
+        GLfloat column[4] = {};
+        for (GLint matrix = 0; matrix < count; ++matrix) {
+            if (matrix > 0 && !programObject.UniformLocationsAliasSameUniform(location, location + matrix)) {
+                // GL 3.3 2.11.4: values for elements beyond the end of the uniform array
+                // are ignored. Never step onto a neighboring uniform's location.
+                break;
+            }
+            if (!programObject.IsValidUniformLocation(location + matrix)) {
+                RecordInvalidUniformLocationError(caller, location + matrix, ownerDescription);
+                return;
+            }
+            if (programObject.IsUniformOpaqueAtLocation(location + matrix)) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                                 "Opaque uniforms cannot be set with matrix Uniform calls."));
+                return;
+            }
+            if (value == nullptr) return;
+            const GLfloat* source = value + static_cast<SizeT>(matrix) * componentCount;
+            for (Int c = 0; c < columns; ++c) {
+                for (Int r = 0; r < rows; ++r) {
+                    column[r] = transpose == GL_TRUE ? source[r * columns + c] : source[c * rows + r];
+                }
+                const SizeT byteOffset = static_cast<SizeT>(c) * kColumnStride;
+                switch (rows) {
+                case 2: Uniform_State<2>(programObject, location + matrix, column, byteOffset); break;
+                case 3: Uniform_State<3>(programObject, location + matrix, column, byteOffset); break;
+                default: Uniform_State<4>(programObject, location + matrix, column, byteOffset); break;
+                }
+            }
+        }
+    }
+
     // Helper function to transpose a 2x2 matrix
     void TransposeMatrix2x2(const GLfloat* input, GLfloat* output) {
         // Input matrix is in column-major order (OpenGL default)
@@ -1293,8 +1375,8 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void UniformMatrix2fv_State(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
-        // For 2x2 matrices, we have 4 elements per matrix
+        // A mat2 is NOT four contiguous floats in the global UBO: std140 pads each column
-        // If transpose is GL_TRUE, we need to transpose the matrix data
+        // vector out to 16 bytes, so column 1 starts at byte 16, not byte 8.
         if (location == -1) return;
 
         auto& programObject = MG_State::pGLContext->GetProgramForUniform();
@@ -1305,26 +1387,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        // For matrix uniforms, we handle each matrix individually
+        UniformMatrixfv_Object(*programObject, __func__, location, count, transpose, value, 2, 2,
-        for (GLint i = 0; i < count; i++) {
+                               "the current program object");
-            if (i > 0 && !programObject->UniformLocationsAliasSameUniform(location, location + i)) {
-                // Values for elements beyond the end of the uniform array are ignored.
-                break;
-            }
-            if (!programObject->IsValidUniformLocation(location + i)) {
-                RecordInvalidUniformLocationError(__func__, location + i, "the current program object");
-                return;
-            }
-            if (transpose == GL_TRUE) {
-                // Transpose the matrix before uploading
-                GLfloat transposedMatrix[4];
-                TransposeMatrix2x2(value + i * 4, transposedMatrix);
-                Uniform_State<4>(*programObject, location + i, transposedMatrix);
-            } else {
-                // No transpose needed, directly copy the matrix data
-                Uniform_State<4>(*programObject, location + i, value + i * 4);
-            }
-        }
     }
 
     void UniformMatrix3fv_State(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
@@ -1402,7 +1466,8 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
-    void UniformMatrixNonSquarefv_State(const char* caller, GLint location, GLsizei count) {
+    void UniformMatrixNonSquarefv_State(const char* caller, GLint location, GLsizei count, GLboolean transpose,
+                                        const GLfloat* value, Int columns, Int rows) {
         if (location == -1) return;
 
         auto& programObject = MG_State::pGLContext->GetProgramForUniform();
@@ -1413,21 +1478,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        for (GLint i = 0; i < count; i++) {
+        UniformMatrixfv_Object(*programObject, caller, location, count, transpose, value, columns, rows,
-            if (!programObject->IsValidUniformLocation(location + i)) {
+                               "the current program object");
-                RecordInvalidUniformLocationError(caller, location + i, "the current program object");
-                return;
-            }
-            if (programObject->IsUniformOpaqueAtLocation(location + i)) {
-                MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidOperation,
-                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
-                                                 "Opaque uniforms cannot be set with matrix Uniform calls."));
-                return;
-            }
-        }
-
-        // TODO: Implement non-square matrix uniform uploads for non-opaque uniforms.
     }
 
     void ProgramUniformMatrix2fv_State(GLuint program, GLint location, GLsizei count, GLboolean transpose,
@@ -1445,23 +1497,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        for (GLint i = 0; i < count; i++) {
+        UniformMatrixfv_Object(*programObject, __func__, location, count, transpose, value, 2, 2,
-            if (i > 0 && !programObject->UniformLocationsAliasSameUniform(location, location + i)) {
+                               "program " + std::to_string(program));
-                // Values for elements beyond the end of the uniform array are ignored.
-                break;
-            }
-            if (!programObject->IsValidUniformLocation(location + i)) {
-                RecordInvalidUniformLocationError(__func__, location + i, "program " + std::to_string(program));
-                return;
-            }
-            if (transpose == GL_TRUE) {
-                GLfloat transposedMatrix[4];
-                TransposeMatrix2x2(value + i * 4, transposedMatrix);
-                Uniform_State<4>(*programObject, location + i, transposedMatrix);
-            } else {
-                Uniform_State<4>(*programObject, location + i, value + i * 4);
-            }
-        }
     }
 
     void ProgramUniformMatrix3fv_State(GLuint program, GLint location, GLsizei count, GLboolean transpose,
@@ -1536,7 +1573,8 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
-    void ProgramUniformMatrixNonSquarefv_State(const char* caller, GLuint program, GLint location, GLsizei count) {
+    void ProgramUniformMatrixNonSquarefv_State(const char* caller, GLuint program, GLint location, GLsizei count,
+                                               GLboolean transpose, const GLfloat* value, Int columns, Int rows) {
         if (location == -1) return;
 
         auto& programObject = TryToGetProgramObject(program);
@@ -1550,21 +1588,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        for (GLint i = 0; i < count; i++) {
+        UniformMatrixfv_Object(*programObject, caller, location, count, transpose, value, columns, rows,
-            if (!programObject->IsValidUniformLocation(location + i)) {
+                               "program " + std::to_string(program));
-                RecordInvalidUniformLocationError(caller, location + i, "program " + std::to_string(program));
-                return;
-            }
-            if (programObject->IsUniformOpaqueAtLocation(location + i)) {
-                MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidOperation,
-                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
-                                                 "Opaque uniforms cannot be set with matrix Uniform calls."));
-                return;
-            }
-        }
-
-        // TODO: Implement non-square matrix uniform uploads for non-opaque uniforms.
     }
 
     GLuint GetUniformBlockIndex_State(GLuint program, const GLchar* uniformBlockName) {
@@ -1832,6 +1857,16 @@ namespace MobileGL::MG_Impl::GLImpl {
     void CompileShader(GLuint shader) {
         CompileShader_State(shader);
     }
+
+    void MaxShaderCompilerThreadsKHR(GLuint count) {
+        MaxShaderCompilerThreadsKHR_State(count);
+    }
+
+    // GL_ARB_parallel_shader_compile's spelling of the same entry point.
+    void MaxShaderCompilerThreadsARB(GLuint count) {
+        MaxShaderCompilerThreadsKHR_State(count);
+    }
+
     GLuint CreateProgram(void) {
         return CreateProgram_State();
     }
@@ -2363,27 +2398,27 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void UniformMatrix2x3fv(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
-        UniformMatrixNonSquarefv_State(__func__, location, count);
+        UniformMatrixNonSquarefv_State(__func__, location, count, transpose, value, 2, 3);
     }
 
     void UniformMatrix3x2fv(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
-        UniformMatrixNonSquarefv_State(__func__, location, count);
+        UniformMatrixNonSquarefv_State(__func__, location, count, transpose, value, 3, 2);
     }
 
     void UniformMatrix2x4fv(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
-        UniformMatrixNonSquarefv_State(__func__, location, count);
+        UniformMatrixNonSquarefv_State(__func__, location, count, transpose, value, 2, 4);
     }
 
     void UniformMatrix4x2fv(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
-        UniformMatrixNonSquarefv_State(__func__, location, count);
+        UniformMatrixNonSquarefv_State(__func__, location, count, transpose, value, 4, 2);
     }
 
     void UniformMatrix3x4fv(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
-        UniformMatrixNonSquarefv_State(__func__, location, count);
+        UniformMatrixNonSquarefv_State(__func__, location, count, transpose, value, 3, 4);
     }
 
     void UniformMatrix4x3fv(GLint location, GLsizei count, GLboolean transpose, const GLfloat* value) {
-        UniformMatrixNonSquarefv_State(__func__, location, count);
+        UniformMatrixNonSquarefv_State(__func__, location, count, transpose, value, 4, 3);
     }
 
     void ProgramUniform1f(GLuint program, GLint location, GLfloat v0) {
@@ -2508,32 +2543,32 @@ namespace MobileGL::MG_Impl::GLImpl {
 
     void ProgramUniformMatrix2x3fv(GLuint program, GLint location, GLsizei count, GLboolean transpose,
                                    const GLfloat* value) {
-        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count);
+        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count, transpose, value, 2, 3);
     }
 
     void ProgramUniformMatrix3x2fv(GLuint program, GLint location, GLsizei count, GLboolean transpose,
                                    const GLfloat* value) {
-        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count);
+        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count, transpose, value, 3, 2);
     }
 
     void ProgramUniformMatrix2x4fv(GLuint program, GLint location, GLsizei count, GLboolean transpose,
                                    const GLfloat* value) {
-        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count);
+        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count, transpose, value, 2, 4);
     }
 
     void ProgramUniformMatrix4x2fv(GLuint program, GLint location, GLsizei count, GLboolean transpose,
                                    const GLfloat* value) {
-        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count);
+        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count, transpose, value, 4, 2);
     }
 
     void ProgramUniformMatrix3x4fv(GLuint program, GLint location, GLsizei count, GLboolean transpose,
                                    const GLfloat* value) {
-        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count);
+        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count, transpose, value, 3, 4);
     }
 
     void ProgramUniformMatrix4x3fv(GLuint program, GLint location, GLsizei count, GLboolean transpose,
                                    const GLfloat* value) {
-        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count);
+        ProgramUniformMatrixNonSquarefv_State(__func__, program, location, count, transpose, value, 4, 3);
     }
 
     GLuint GetUniformBlockIndex(GLuint program, const GLchar* uniformBlockName) {
@@ -2570,171 +2605,160 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void GetProgramInterfaceiv(GLuint program, GLenum programInterface, GLenum pname, GLint* params) {
-        auto& programObject = TryToGetLinkedProgramForInterfaceQuery(program, __func__);
+        auto& programObject = TryToGetProgramForInterfaceQuery(program, __func__);
         if (!programObject) return;
         if (!ValidateProgramInterfaceivQuery(programInterface, pname)) return;
-        auto getProgramInterfaceiv = MG_Backend::gBackendFunctionsTable.GL.GetProgramInterfaceiv;
+        if (!params) return;
-        if (!getProgramInterfaceiv) {
+        switch (pname) {
-            MG_State::pGLContext->RecordError(
+        case GL_ACTIVE_RESOURCES:
-                ErrorCode::InvalidOperation,
+            *params = ProgramInterface::GetActiveResourceCount(*programObject, programInterface);
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+            return;
-                                             "Backend does not support program interface queries."));
+        case GL_MAX_NAME_LENGTH:
+            *params = ProgramInterface::GetMaxNameLength(*programObject, programInterface);
+            return;
+        case GL_MAX_NUM_ACTIVE_VARIABLES:
+            *params = ProgramInterface::GetMaxNumActiveVariables(*programObject, programInterface);
+            return;
+        default:
+            // GL_MAX_NUM_COMPATIBLE_SUBROUTINES: the subroutine interfaces are always empty
+            // here (glslang refuses `subroutine` when generating SPIR-V), so zero it is.
+            *params = 0;
             return;
         }
-        if (programInterface == GL_UNIFORM) {
-            if (pname == GL_ACTIVE_RESOURCES) {
-                *params = static_cast<GLint>(programObject->GetUniformCount());
-                return;
-            }
-            if (pname == GL_MAX_NAME_LENGTH) {
-                // Stored as the bare length; GL_MAX_NAME_LENGTH counts the terminator.
-                *params = programObject->GetUniformMaxLength() + 1;
-                return;
-            }
-        }
-        getProgramInterfaceiv(program, programInterface, pname, params);
     }
 
     GLuint GetProgramResourceIndex(GLuint program, GLenum programInterface, const GLchar* name) {
-        auto& programObject = TryToGetLinkedProgramForInterfaceQuery(program, __func__);
+        auto& programObject = TryToGetProgramForInterfaceQuery(program, __func__);
         if (!programObject) return GL_INVALID_INDEX;
         if (!ValidateNamedProgramResourceInterface(programInterface, __func__)) return GL_INVALID_INDEX;
         if (!name) return GL_INVALID_INDEX;
-        if (programInterface == GL_UNIFORM) {
+        return ProgramInterface::GetResourceIndex(*programObject, programInterface, name);
-            const Int uniformIndex = programObject->GetActiveUniformIndex(name);
-            return uniformIndex < 0 ? GL_INVALID_INDEX : static_cast<GLuint>(uniformIndex);
-        }
-        auto getProgramResourceIndex = MG_Backend::gBackendFunctionsTable.GL.GetProgramResourceIndex;
-        if (!getProgramResourceIndex) {
-            MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidOperation,
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                             "Backend does not support program interface queries."));
-            return GL_INVALID_INDEX;
-        }
-        GLuint index = getProgramResourceIndex(program, programInterface, name);
-        const String resourceName = name;
-        if (index == GL_INVALID_INDEX && resourceName.length() > 3 &&
-            resourceName.compare(resourceName.length() - 3, 3, "[0]") == 0) {
-            index = getProgramResourceIndex(program, programInterface,
-                                            resourceName.substr(0, resourceName.length() - 3).c_str());
-        }
-        return index;
     }
 
     void GetProgramResourceName(GLuint program, GLenum programInterface, GLuint index, GLsizei bufSize, GLsizei* length,
                                 GLchar* name) {
-        auto& programObject = TryToGetLinkedProgramForInterfaceQuery(program, __func__);
+        auto& programObject = TryToGetProgramForInterfaceQuery(program, __func__);
         if (!programObject) return;
         if (!ValidateNamedProgramResourceInterface(programInterface, __func__)) return;
-        const Int resourceCount = GetKnownProgramResourceCount(programObject, programInterface);
-        if (resourceCount >= 0 && index >= static_cast<GLuint>(resourceCount)) {
-            MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidValue,
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "index is out of range."));
-            return;
-        }
         if (bufSize < 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "bufSize must be non-negative."));
             return;
         }
-        if (programInterface == GL_UNIFORM) {
+        String resourceName;
-            // Same index space GetProgramResourceIndex answers in, and the range check above
+        if (!ProgramInterface::GetResourceName(*programObject, programInterface, index, resourceName)) {
-            // already used it.
-            const String& uniformName = programObject->GetActiveUniformName(index);
-            CopyStr(bufSize, length, name, uniformName.c_str(), static_cast<GLsizei>(uniformName.length()));
-            return;
-        }
-        auto getProgramResourceName = MG_Backend::gBackendFunctionsTable.GL.GetProgramResourceName;
-        if (!getProgramResourceName) {
             MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidOperation,
+                ErrorCode::InvalidValue,
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "index is out of range."));
-                                             "Backend does not support program interface queries."));
             return;
         }
-        getProgramResourceName(program, programInterface, index, bufSize, length, name);
+        CopyStr(bufSize, length, name, resourceName.c_str(), static_cast<GLsizei>(resourceName.length()));
     }
 
     void GetProgramResourceiv(GLuint program, GLenum programInterface, GLuint index, GLsizei propCount,
                               const GLenum* props, GLsizei bufSize, GLsizei* length, GLint* params) {
-        auto& programObject = TryToGetLinkedProgramForInterfaceQuery(program, __func__);
+        auto& programObject = TryToGetProgramForInterfaceQuery(program, __func__);
         if (!programObject) return;
-        if (propCount < 0 || bufSize < 0) {
+        if (!ProgramInterface::IsInterfaceEnum(programInterface)) {
             MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidValue, MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                ErrorCode::InvalidEnum,
-                                                                      "propCount and bufSize must be non-negative."));
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Unsupported program interface."));
             return;
         }
-        if (programInterface == GL_UNIFORM) {
+        if (propCount <= 0 || bufSize < 0) {
-            if (index >= programObject->GetUniformCount()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue, MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                                                      "propCount must be positive and bufSize "
+                                                                      "non-negative."));
+            return;
+        }
+        if (props == nullptr) return;
+        // Both prop checks run BEFORE any value is produced: a property this command does
+        // not know at all is INVALID_ENUM, one it knows but the interface does not carry is
+        // INVALID_OPERATION (GL 4.6 Table 7.2). The two are deliberately different errors.
+        for (GLsizei i = 0; i < propCount; ++i) {
+            if (!ProgramInterface::IsResourceProp(props[i])) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidEnum,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "prop is not a valid property name."));
+                return;
+            }
+            if (!ProgramInterface::InterfaceSupportsProp(programInterface, props[i])) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                                 "prop is not supported for this program interface."));
+                return;
+            }
+        }
+
+        Vector<GLint> values;
+        for (GLsizei i = 0; i < propCount; ++i) {
+            if (!ProgramInterface::GetResourceProp(*programObject, programInterface, index, props[i], values)) {
                 MG_State::pGLContext->RecordError(
                     ErrorCode::InvalidValue,
                     MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "index is out of range."));
                 return;
             }
-            if (props == nullptr || params == nullptr) return;
-            GLsizei written = 0;
-            for (GLsizei i = 0; i < propCount && written < bufSize; ++i) {
-                GLint value = 0;
-                if (!GetUniformResourceProp(programObject, index, props[i], &value)) {
-                    // GL_ATOMIC_COUNTER_BUFFER_INDEX and the GL_REFERENCED_BY_* stage props are
-                    // not modelled here; ask the backend, which indexes resources by name.
-                    auto backendGetIndex = MG_Backend::gBackendFunctionsTable.GL.GetProgramResourceIndex;
-                    auto backendGetiv = MG_Backend::gBackendFunctionsTable.GL.GetProgramResourceiv;
-                    if (backendGetIndex && backendGetiv) {
-                        const GLuint backendIndex = backendGetIndex(program, GL_UNIFORM,
-                                                                   programObject->GetActiveUniformName(index).c_str());
-                        if (backendIndex != GL_INVALID_INDEX) {
-                            GLsizei one = 0;
-                            backendGetiv(program, GL_UNIFORM, backendIndex, 1, &props[i], 1, &one, &value);
-                        }
-                    }
-                }
-                params[written++] = value;
-            }
-            if (length) *length = written;
-            return;
         }
-        auto getProgramResourceiv = MG_Backend::gBackendFunctionsTable.GL.GetProgramResourceiv;
+        if (params == nullptr) return;
-        if (!getProgramResourceiv) {
+        const GLsizei written = static_cast<GLsizei>(std::min<SizeT>(values.size(), static_cast<SizeT>(bufSize)));
-            MG_State::pGLContext->RecordError(
+        for (GLsizei i = 0; i < written; ++i) params[i] = values[i];
-                ErrorCode::InvalidOperation,
+        if (length) *length = written;
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                             "Backend does not support program interface queries."));
-            return;
-        }
-        getProgramResourceiv(program, programInterface, index, propCount, props, bufSize, length, params);
     }
 
     GLint GetProgramResourceLocation(GLuint program, GLenum programInterface, const GLchar* name) {
+        // Unlike the four queries above, this one and GetProgramResourceLocationIndex really
+        // do require a successful link (GL 4.6 §7.3.1.3).
         auto& programObject = TryToGetLinkedProgramForInterfaceQuery(program, __func__);
         if (!programObject) return -1;
-        auto getProgramResourceLocation = MG_Backend::gBackendFunctionsTable.GL.GetProgramResourceLocation;
+        if (!ProgramInterface::InterfaceHasLocations(programInterface)) {
-        if (!getProgramResourceLocation) {
             MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidOperation,
+                ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                             "Backend does not support program interface queries."));
+                                             "Program interface has no locations."));
             return -1;
         }
-        return getProgramResourceLocation(program, programInterface, name);
+        return ProgramInterface::GetResourceLocation(*programObject, programInterface, name);
     }
 
     GLint GetProgramResourceLocationIndex(GLuint program, GLenum programInterface, const GLchar* name) {
         auto& programObject = TryToGetLinkedProgramForInterfaceQuery(program, __func__);
         if (!programObject) return -1;
-        auto getProgramResourceLocationIndex = MG_Backend::gBackendFunctionsTable.GL.GetProgramResourceLocationIndex;
+        if (programInterface != GL_PROGRAM_OUTPUT) {
-        if (!getProgramResourceLocationIndex) return -1;
+            MG_State::pGLContext->RecordError(
-        return getProgramResourceLocationIndex(program, programInterface, name);
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "GetProgramResourceLocationIndex only accepts GL_PROGRAM_OUTPUT."));
+            return -1;
+        }
+        return ProgramInterface::GetResourceLocationIndex(*programObject, programInterface, name);
     }
 
+    // GL 4.6 §7.6.2: <storageBlockIndex> is an active shader storage block index of <program>
+    // - that is, exactly what glGetProgramResourceIndex(GL_SHADER_STORAGE_BLOCK) returned.
+    // Since wave 2 that index is the interface-query layer's, so this is where the one index
+    // space the application sees gets turned into whatever the backend's is; the backends are
+    // handed the block NAME and do their own lookup. Getting this wrong is silent: the call
+    // succeeds and rebinds a DIFFERENT buffer.
     void ShaderStorageBlockBinding(GLuint program, GLuint storageBlockIndex, GLuint storageBlockBinding) {
         auto& programObject = TryToGetProgramObject(program);
         if (!programObject || !programObject->GetLinkStatus()) return;
         if (!ValidateShaderStorageBlockBinding(storageBlockBinding)) return;
+        String blockName;
+        if (!ProgramInterface::GetResourceName(*programObject, GL_SHADER_STORAGE_BLOCK, storageBlockIndex,
+                                               blockName)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "storageBlockIndex is not an active shader storage block index."));
+            return;
+        }
+        // Recorded before the backend call, and independently of whether a backend is even
+        // present: this is the state GL_BUFFER_BINDING reports, and it is also what reseeds a
+        // backend's own reflection cache after any rebuild.
+        programObject->SetShaderStorageBlockBinding(blockName, storageBlockBinding);
         auto shaderStorageBlockBinding = MG_Backend::gBackendFunctionsTable.GL.ShaderStorageBlockBinding;
         if (!shaderStorageBlockBinding) {
             MG_State::pGLContext->RecordError(
@@ -2743,7 +2767,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                              "Backend does not support shader storage block binding."));
             return;
         }
-        shaderStorageBlockBinding(program, storageBlockIndex, storageBlockBinding);
+        shaderStorageBlockBinding(program, blockName.c_str(), storageBlockBinding);
     }
 
     void ValidateProgram(GLuint program) {
@@ -2780,6 +2804,16 @@ namespace MobileGL::MG_Impl::GLImpl {
     // is written as that sequence rather than as a private shortcut - every error it can
     // raise is one of theirs, raised at the point they would raise it.
     GLuint CreateShaderProgramv(GLenum type, GLsizei count, const GLchar* const* strings) {
+        // GL 4.6 core 7.3: a negative count is INVALID_VALUE and is checked before anything
+        // is created, so a bad count never leaks a shader name. An unrecognised type is
+        // INVALID_ENUM, which CreateShader_State raises below.
+        if (count < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "count must be non-negative."));
+            return 0;
+        }
+
         const GLuint shader = CreateShader_State(type);
         if (shader == 0) return 0;
 

MobileGL/MG_Impl/GLImpl/Program/GL_Program.h

@@ -42,6 +42,10 @@ namespace MobileGL::MG_Impl::GLImpl {
     GLboolean IsProgram(GLuint program);
     GLboolean IsShader(GLuint shader);
     void LinkProgram(GLuint program);
+    // GL_KHR_parallel_shader_compile / GL_ARB_parallel_shader_compile. Both names are the
+    // same entry point; see MaxShaderCompilerThreadsKHR_State for the semantics of count.
+    void MaxShaderCompilerThreadsKHR(GLuint count);
+    void MaxShaderCompilerThreadsARB(GLuint count);
     void ShaderSource(GLuint shader, GLsizei count, const GLchar* const* string, const GLint* length);
     void UseProgram(GLuint program);
     void Uniform1f(GLint location, GLfloat v0);

MobileGL/MG_Impl/GLImpl/Program/ProgramInterface.cpp

@@ -0,0 +1,841 @@
+// MobileGL - MobileGL/MG_Impl/GLImpl/Program/ProgramInterface.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "ProgramInterface.h"
+
+#include <MG_State/GLState/ProgramState/ProgramObject.h>
+#include <MG_Util/ShaderTranspiler/Types.h>
+
+#include <cstring>
+
+namespace MobileGL::MG_Impl::GLImpl::ProgramInterface {
+    namespace {
+        // glslang folds atomic counters into synthesized blocks named
+        // "<getAtomicCounterBlockName()>_<binding>" (ParseContextBase.cpp), one per GL
+        // atomic-counter binding point. That block IS the GL_ATOMIC_COUNTER_BUFFER resource
+        // and its trailing number IS GL_BUFFER_BINDING; its members stay GL_UNIFORMs.
+        constexpr const char* kAtomicCounterBlockPrefix = "gl_AtomicCounterBlock";
+
+        enum class BlockKind {
+            Uniform,       // a real GL uniform block
+            GlobalUbo,     // the synthesized MGL_GLOBAL_UBO: GL sees its members as default-block
+            AtomicCounter, // gl_AtomicCounterBlock_<binding>
+            Storage,       // a shader storage block
+        };
+
+        // One row of any interface. Fields a given interface does not have keep the
+        // spec-mandated "not applicable" value, so a prop read never has to special-case
+        // the interface a second time.
+        struct Resource {
+            String name;
+            GLenum type = GL_NONE;
+            GLint arraySize = 1;
+            GLint location = -1;
+            GLint locationIndex = -1;
+            GLint blockIndex = -1;
+            GLint offset = -1;
+            GLint arrayStride = -1;
+            GLint matrixStride = -1;
+            GLint isRowMajor = 0;
+            GLint atomicCounterBufferIndex = -1;
+            GLint topLevelArraySize = 0;
+            GLint topLevelArrayStride = 0;
+            GLint bufferBinding = 0;
+            GLint bufferDataSize = 0;
+            GLint isPerPatch = 0;
+            GLint xfbBufferIndex = 0;
+            Uint32 stages = 0; // EShLanguageMask
+            Vector<GLuint> activeVariables;
+        };
+
+        using ResourceList = Vector<Resource>;
+
+        struct Model {
+            ResourceList uniforms;
+            ResourceList uniformBlocks;
+            ResourceList atomicCounterBuffers;
+            ResourceList bufferVariables;
+            ResourceList storageBlocks;
+            ResourceList programInputs;
+            ResourceList programOutputs;
+            ResourceList xfbVaryings;
+            Bool valid = false;
+        };
+
+        const ResourceList& EmptyList() {
+            static const ResourceList empty;
+            return empty;
+        }
+
+        // ---- name spelling (cluster 6) -------------------------------------------------
+
+        Bool EndsWithZeroSubscript(const String& name) {
+            return name.length() >= 3 && name.compare(name.length() - 3, 3, "[0]") == 0;
+        }
+
+        // The enumerated spelling of an array resource is "name[0]". glslang already applies
+        // that to uniforms and buffer variables (EShReflectionBasicArraySuffix), but never to
+        // stage inputs/outputs, so those get it here.
+        String WithArraySuffix(const String& name, const glslang::TType* type) {
+            if (type == nullptr || !type->isArray() || EndsWithZeroSubscript(name)) return name;
+            return name + "[0]";
+        }
+
+        // GL_ARRAY_SIZE: element count for a sized array, 0 for a runtime-sized one
+        // (a shader storage block's unsized trailing member), 1 for a non-array.
+        GLint ArraySizeOf(const glslang::TType* type, GLint reflectedSize) {
+            if (type != nullptr && type->isArray()) {
+                if (!type->isSizedArray()) return 0;
+                return type->getOuterArraySize();
+            }
+            return reflectedSize < 1 ? 1 : reflectedSize;
+        }
+
+        // Two spellings name the same resource when they are equal, or differ only by the
+        // "[0]" the enumeration appends to an array.
+        Bool NamesMatch(const String& resourceName, const String& query) {
+            if (resourceName == query) return true;
+            if (EndsWithZeroSubscript(resourceName) &&
+                resourceName.compare(0, resourceName.length() - 3, query) == 0) {
+                return true;
+            }
+            return EndsWithZeroSubscript(query) && query.compare(0, query.length() - 3, resourceName) == 0;
+        }
+
+        // Splits "base[k]" into ("base", k). GL 4.6 §7.3.1.1 requires the subscript to be a
+        // decimal integer with no white space and no leading zeros, which is exactly what
+        // separates array-names' "a[1]" (resolves) from "a[01]", "a[0 + 0]" and "a[ 0]" (do
+        // not). Returns false when there is no trailing subscript at all; sets `malformed`
+        // when there is one but it is not a strict decimal.
+        Bool SplitTrailingSubscript(const String& name, String& outBase, Uint& outElement, Bool& outMalformed) {
+            outMalformed = false;
+            if (name.empty() || name.back() != ']') return false;
+            const SizeT bracket = name.rfind('[');
+            if (bracket == String::npos) return false;
+            const SizeT first = bracket + 1;
+            const SizeT last = name.length() - 1; // one past the digits
+            if (first >= last) {
+                outMalformed = true;
+                return false;
+            }
+            // No leading zeros: "0" is the only spelling that may start with '0'.
+            if (name[first] == '0' && last - first > 1) {
+                outMalformed = true;
+                return false;
+            }
+            Uint element = 0;
+            for (SizeT i = first; i < last; ++i) {
+                if (name[i] < '0' || name[i] > '9') {
+                    outMalformed = true;
+                    return false;
+                }
+                element = element * 10 + static_cast<Uint>(name[i] - '0');
+                if (element > 0x0FFFFFFFu) {
+                    outMalformed = true;
+                    return false;
+                }
+            }
+            outBase = name.substr(0, bracket);
+            outElement = element;
+            return true;
+        }
+
+        // ---- block classification ------------------------------------------------------
+
+        Bool IsAtomicCounterBlockName(const String& name) {
+            return name.compare(0, std::strlen(kAtomicCounterBlockPrefix), kAtomicCounterBlockPrefix) == 0;
+        }
+
+        // "gl_AtomicCounterBlock_5" -> 5. The suffix is the GL binding the counters were
+        // declared with, which glslang does NOT keep in the block's own layout qualifier
+        // (that one is remapped to a plain buffer binding).
+        GLint AtomicCounterBlockBinding(const String& name) {
+            const SizeT underscore = name.rfind('_');
+            if (underscore == String::npos || underscore + 1 >= name.length()) return 0;
+            GLint binding = 0;
+            for (SizeT i = underscore + 1; i < name.length(); ++i) {
+                if (name[i] < '0' || name[i] > '9') return 0;
+                binding = binding * 10 + (name[i] - '0');
+            }
+            return binding;
+        }
+
+        // Element index of an arrayed block instance ("TrickyBuffer[1]" -> 1).
+        GLint BlockArrayElement(const String& name) {
+            String base;
+            Uint element = 0;
+            Bool malformed = false;
+            if (!SplitTrailingSubscript(name, base, element, malformed)) return 0;
+            return static_cast<GLint>(element);
+        }
+
+        BlockKind ClassifyBlock(const glslang::TObjectReflection& block) {
+            if (std::strstr(block.name.c_str(), MG_Util::ShaderTranspiler::GLOBAL_UBO_NAME) != nullptr) {
+                return BlockKind::GlobalUbo;
+            }
+            if (IsAtomicCounterBlockName(block.name)) return BlockKind::AtomicCounter;
+            const glslang::TType* type = block.getType();
+            if (type != nullptr && type->getQualifier().storage == glslang::EvqBuffer) return BlockKind::Storage;
+            return BlockKind::Uniform;
+        }
+
+        // std140/std430 column stride, the same vec4-rounded rule ProgramObject applies to
+        // uniform matrices. 0 for a non-matrix.
+        GLint MatrixStrideOf(const glslang::TType* type) {
+            if (type == nullptr || !type->isMatrix()) return 0;
+            const bool rowMajor = type->getQualifier().layoutMatrix == glslang::ElmRowMajor;
+            const int strideVectorComponents = rowMajor ? type->getMatrixCols() : type->getMatrixRows();
+            constexpr int scalarSize = 4;
+            const int vectorAlignment = (strideVectorComponents <= 1)    ? scalarSize
+                                        : (strideVectorComponents == 2) ? 2 * scalarSize
+                                                                        : 4 * scalarSize;
+            return (vectorAlignment + 15) & ~15;
+        }
+
+        GLint IsRowMajorOf(const glslang::TType* type) {
+            if (type == nullptr || !type->isMatrix()) return 0;
+            return type->getQualifier().layoutMatrix == glslang::ElmRowMajor ? 1 : 0;
+        }
+
+        GLint MappedLocation(Int rawLocation) {
+            // glslang parks "no location" at layoutLocationEnd; GL spells it -1.
+            if (rawLocation < 0 || rawLocation >= static_cast<Int>(glslang::TQualifier::layoutLocationEnd)) return -1;
+            return rawLocation;
+        }
+
+        // ---- model construction --------------------------------------------------------
+
+        void BuildBlocks(ProgramObject& program, const glslang::TProgram& reflection, Model& model,
+                         Vector<BlockKind>& blockKind, Vector<Int>& blockInterfaceIndex) {
+            const Int blockCount = const_cast<glslang::TProgram&>(reflection).getNumUniformBlocks();
+            blockKind.assign(blockCount, BlockKind::Uniform);
+            blockInterfaceIndex.assign(blockCount, -1);
+
+            for (Int tIndex = 0; tIndex < blockCount; ++tIndex) {
+                const auto& block = const_cast<glslang::TProgram&>(reflection).getUniformBlock(tIndex);
+                const BlockKind kind = ClassifyBlock(block);
+                blockKind[tIndex] = kind;
+                if (kind == BlockKind::AtomicCounter) {
+                    Resource resource;
+                    // GL_ATOMIC_COUNTER_BUFFER resources have no name (and GetProgramResource
+                    // Index/Name reject the interface outright, which is why this stays empty).
+                    resource.bufferBinding = AtomicCounterBlockBinding(block.name);
+                    resource.bufferDataSize = block.size;
+                    resource.stages = static_cast<Uint32>(block.stages);
+                    blockInterfaceIndex[tIndex] = static_cast<Int>(model.atomicCounterBuffers.size());
+                    model.atomicCounterBuffers.push_back(Move(resource));
+                } else if (kind == BlockKind::Storage) {
+                    Resource resource;
+                    resource.name = block.name;
+                    // glslang reports the DECLARED binding for every instance of an arrayed
+                    // block; GL gives element k the binding base + k. That is only the initial
+                    // value: GL_BUFFER_BINDING must report the CURRENT binding, so a later
+                    // glShaderStorageBlockBinding wins over the declaration (GL 4.6 §7.6.2 -
+                    // exactly the same rule GL_UNIFORM_BLOCK follows through
+                    // GetUniformBlockBinding below).
+                    const GLint declared = block.getBinding();
+                    resource.bufferBinding = declared < 0 ? 0 : declared + BlockArrayElement(block.name);
+                    const Int rebound = program.GetShaderStorageBlockBindingOverride(block.name);
+                    if (rebound >= 0) resource.bufferBinding = static_cast<GLint>(rebound);
+                    resource.bufferDataSize = block.size;
+                    resource.stages = static_cast<Uint32>(block.stages);
+                    blockInterfaceIndex[tIndex] = static_cast<Int>(model.storageBlocks.size());
+                    model.storageBlocks.push_back(Move(resource));
+                }
+            }
+
+            // GL_UNIFORM_BLOCK keeps the index space glUniformBlockBinding and
+            // glGetActiveUniformBlockiv already use, so an index handed out here is usable
+            // with them (which is exactly what the CTS does).
+            const Int glBlockCount = program.GetActiveUniformBlocksCount();
+            for (Int glIndex = 0; glIndex < glBlockCount; ++glIndex) {
+                Resource resource;
+                resource.name = program.GetUniformBlockName(glIndex);
+                resource.bufferBinding = static_cast<GLint>(program.GetUniformBlockBinding(glIndex));
+                resource.bufferDataSize = static_cast<GLint>(program.GetUBOSizeAt(glIndex));
+                const Int tIndex = program.TProgramBlockIndex(static_cast<Uint>(glIndex));
+                if (tIndex >= 0 && tIndex < blockCount) {
+                    resource.stages =
+                        static_cast<Uint32>(const_cast<glslang::TProgram&>(reflection).getUniformBlock(tIndex).stages);
+                }
+                model.uniformBlocks.push_back(Move(resource));
+            }
+        }
+
+        void BuildUniformsAndBufferVariables(ProgramObject& program, const glslang::TProgram& reflection, Model& model,
+                                             const Vector<BlockKind>& blockKind,
+                                             const Vector<Int>& blockInterfaceIndex) {
+            const Uint uniformCount = program.GetUniformCount();
+            for (Uint glIndex = 0; glIndex < uniformCount; ++glIndex) {
+                const Int tIndex = program.TProgramUniformIndex(glIndex);
+                const auto& refl = const_cast<glslang::TProgram&>(reflection).getUniform(tIndex);
+                const glslang::TType* type = refl.getType();
+                const Int owner = refl.index;
+                const BlockKind kind = (owner >= 0 && owner < static_cast<Int>(blockKind.size()))
+                                           ? blockKind[owner]
+                                           : BlockKind::GlobalUbo;
+
+                Resource resource;
+                resource.name = refl.name;
+                resource.type = static_cast<GLenum>(refl.glDefineType);
+                resource.arraySize = ArraySizeOf(type, refl.size);
+                resource.stages = static_cast<Uint32>(refl.stages);
+
+                if (kind == BlockKind::Storage) {
+                    resource.blockIndex = blockInterfaceIndex[owner];
+                    resource.offset = refl.offset;
+                    resource.arrayStride = refl.arrayStride;
+                    resource.matrixStride = MatrixStrideOf(type);
+                    resource.isRowMajor = IsRowMajorOf(type);
+                    // GL requires 1 for a member that is not inside a top-level array (and for
+                    // the top-level array itself); glslang leaves 0/-1 there.
+                    resource.topLevelArraySize = refl.topLevelArraySize > 0 ? refl.topLevelArraySize : 1;
+                    resource.topLevelArrayStride = refl.topLevelArrayStride;
+                    model.bufferVariables.push_back(Move(resource));
+                    continue;
+                }
+
+                if (kind == BlockKind::AtomicCounter) {
+                    // An atomic counter is a default-block uniform with no location and no
+                    // owning uniform block; what it does have is a buffer to point at.
+                    resource.type = GL_UNSIGNED_INT_ATOMIC_COUNTER;
+                    resource.blockIndex = -1;
+                    resource.offset = refl.offset;
+                    resource.arrayStride = refl.arrayStride;
+                    resource.matrixStride = 0;
+                    resource.atomicCounterBufferIndex = blockInterfaceIndex[owner];
+                    resource.location = -1;
+                } else {
+                    resource.blockIndex = program.GetActiveUniformBlockIndex(glIndex);
+                    resource.offset = program.GetActiveUniformOffset(glIndex);
+                    resource.arrayStride = program.GetActiveUniformArrayStride(glIndex);
+                    resource.matrixStride = program.GetActiveUniformMatrixStride(glIndex);
+                    resource.isRowMajor = program.GetActiveUniformIsRowMajor(glIndex);
+                    // A member of a named uniform block has no location, whatever the
+                    // frontend's own location table says (it hands one out to every uniform
+                    // so glUniform* can address block members through the global UBO).
+                    resource.location =
+                        resource.blockIndex >= 0 ? -1 : program.GetUniformLocation(refl.name);
+                }
+                model.uniforms.push_back(Move(resource));
+            }
+
+            // GL_ACTIVE_VARIABLES, both directions.
+            for (SizeT i = 0; i < model.uniforms.size(); ++i) {
+                const Resource& uniform = model.uniforms[i];
+                if (uniform.atomicCounterBufferIndex >= 0 &&
+                    uniform.atomicCounterBufferIndex < static_cast<GLint>(model.atomicCounterBuffers.size())) {
+                    model.atomicCounterBuffers[uniform.atomicCounterBufferIndex].activeVariables.push_back(
+                        static_cast<GLuint>(i));
+                }
+            }
+            for (SizeT blockIndex = 0; blockIndex < model.uniformBlocks.size(); ++blockIndex) {
+                // Members of an arrayed block are reflected once, against instance [0].
+                const Int owner = static_cast<Int>(program.GetUniformBlockMemberOwnerIndex(static_cast<Uint>(blockIndex)));
+                for (SizeT i = 0; i < model.uniforms.size(); ++i) {
+                    if (model.uniforms[i].blockIndex == owner) {
+                        model.uniformBlocks[blockIndex].activeVariables.push_back(static_cast<GLuint>(i));
+                    }
+                }
+            }
+            for (SizeT blockIndex = 0; blockIndex < model.storageBlocks.size(); ++blockIndex) {
+                for (SizeT i = 0; i < model.bufferVariables.size(); ++i) {
+                    if (model.bufferVariables[i].blockIndex == static_cast<GLint>(blockIndex)) {
+                        model.storageBlocks[blockIndex].activeVariables.push_back(static_cast<GLuint>(i));
+                    }
+                }
+            }
+        }
+
+        void BuildStageIO(ProgramObject& program, const glslang::TProgram& reflection, Model& model) {
+            auto& mutableReflection = const_cast<glslang::TProgram&>(reflection);
+
+            const Int inputCount = mutableReflection.getNumPipeInputs();
+            for (Int index = 0; index < inputCount; ++index) {
+                const auto& refl = mutableReflection.getPipeInput(index);
+                const glslang::TType* type = refl.getType();
+                Resource resource;
+                // The Vulkan-semantics parse reflects the vertex builtins under their SPIR-V
+                // names; GL enumerates the GL spellings.
+                const String& glName = ProgramObject::NormalizeBuiltinPipeInputName(refl.name);
+                resource.name = WithArraySuffix(glName, type);
+                resource.type = static_cast<GLenum>(refl.glDefineType);
+                resource.arraySize = ArraySizeOf(type, refl.size);
+                resource.location = program.GetAttributeLocation(refl.name);
+                if (resource.location < 0) resource.location = MappedLocation(static_cast<Int>(refl.layoutLocation()));
+                resource.isPerPatch = (type != nullptr && type->getQualifier().patch) ? 1 : 0;
+                resource.stages = static_cast<Uint32>(refl.stages);
+                model.programInputs.push_back(Move(resource));
+            }
+
+            const Int outputCount = mutableReflection.getNumPipeOutputs();
+            for (Int index = 0; index < outputCount; ++index) {
+                const auto& refl = mutableReflection.getPipeOutput(index);
+                const glslang::TType* type = refl.getType();
+                Resource resource;
+                resource.name = WithArraySuffix(refl.name, type);
+                resource.type = static_cast<GLenum>(refl.glDefineType);
+                resource.arraySize = ArraySizeOf(type, refl.size);
+                resource.location = MappedLocation(program.GetFragmentDataLocation(refl.name.c_str()));
+                if (resource.location < 0) {
+                    // A built-in output (gl_FragDepth, gl_SampleMask) and a non-fragment stage
+                    // output both have no location, and therefore no color index either.
+                    resource.locationIndex = -1;
+                } else {
+                    resource.locationIndex = program.GetFragmentDataIndex(refl.name.c_str());
+                    // glBindFragDataLocationIndexed wins; otherwise the shader's
+                    // layout(index = N), which the frag-data maps never saw.
+                    if (resource.locationIndex == 0 && type != nullptr && type->getQualifier().hasIndex()) {
+                        resource.locationIndex = static_cast<GLint>(type->getQualifier().layoutIndex);
+                    }
+                }
+                resource.isPerPatch = (type != nullptr && type->getQualifier().patch) ? 1 : 0;
+                resource.stages = static_cast<Uint32>(refl.stages);
+                model.programOutputs.push_back(Move(resource));
+            }
+        }
+
+        void BuildXfb(ProgramObject& program, Model& model) {
+            const auto& requested = program.GetTransformFeedbackInterfaceNames();
+            const auto& captured = program.GetTransformFeedbackVaryings();
+            for (const String& name : requested) {
+                Resource resource;
+                resource.name = name;
+                // ARB_transform_feedback3's layout controls are enumerated as resources of
+                // type NONE: gl_NextBuffer with array size 0, gl_SkipComponentsN with N.
+                if (name == "gl_NextBuffer") {
+                    resource.type = GL_NONE;
+                    resource.arraySize = 0;
+                } else if (name.size() == 18 && name.compare(0, 17, "gl_SkipComponents") == 0 && name[17] >= '1' &&
+                           name[17] <= '4') {
+                    resource.type = GL_NONE;
+                    resource.arraySize = name[17] - '0';
+                } else {
+                    resource.type = GL_NONE;
+                    resource.arraySize = 1;
+                    for (const auto& varying : captured) {
+                        if (varying.name != name) continue;
+                        resource.type = varying.type;
+                        resource.arraySize = varying.size < 1 ? 1 : varying.size;
+                        resource.offset = static_cast<GLint>(varying.offsetBytes);
+                        resource.xfbBufferIndex = static_cast<GLint>(varying.bufferIndex);
+                        break;
+                    }
+                }
+                model.xfbVaryings.push_back(Move(resource));
+            }
+        }
+
+        Model BuildModel(ProgramObject& program) {
+            Model model;
+            if (!program.GetLinkStatus()) return model;
+            const glslang::TProgram* reflection = program.GetReflection();
+            if (reflection == nullptr) return model;
+            model.valid = true;
+
+            Vector<BlockKind> blockKind;
+            Vector<Int> blockInterfaceIndex;
+            BuildBlocks(program, *reflection, model, blockKind, blockInterfaceIndex);
+            BuildUniformsAndBufferVariables(program, *reflection, model, blockKind, blockInterfaceIndex);
+            BuildStageIO(program, *reflection, model);
+            BuildXfb(program, model);
+            return model;
+        }
+
+        const ResourceList& Select(const Model& model, GLenum programInterface) {
+            switch (programInterface) {
+            case GL_UNIFORM:
+                return model.uniforms;
+            case GL_UNIFORM_BLOCK:
+                return model.uniformBlocks;
+            case GL_ATOMIC_COUNTER_BUFFER:
+                return model.atomicCounterBuffers;
+            case GL_BUFFER_VARIABLE:
+                return model.bufferVariables;
+            case GL_SHADER_STORAGE_BLOCK:
+                return model.storageBlocks;
+            case GL_PROGRAM_INPUT:
+                return model.programInputs;
+            case GL_PROGRAM_OUTPUT:
+                return model.programOutputs;
+            case GL_TRANSFORM_FEEDBACK_VARYING:
+                return model.xfbVaryings;
+            default:
+                // The subroutine interfaces are accepted by the API but nothing can populate
+                // them: glslang refuses `subroutine` when generating SPIR-V, so a program
+                // using one never links. Zero active resources is the honest answer.
+                return EmptyList();
+            }
+        }
+    } // namespace
+
+    Bool IsInterfaceEnum(GLenum programInterface) {
+        switch (programInterface) {
+        case GL_UNIFORM:
+        case GL_UNIFORM_BLOCK:
+        case GL_PROGRAM_INPUT:
+        case GL_PROGRAM_OUTPUT:
+        case GL_BUFFER_VARIABLE:
+        case GL_SHADER_STORAGE_BLOCK:
+        case GL_ATOMIC_COUNTER_BUFFER:
+        case GL_TRANSFORM_FEEDBACK_VARYING:
+        case GL_TRANSFORM_FEEDBACK_BUFFER:
+        case GL_VERTEX_SUBROUTINE:
+        case GL_TESS_CONTROL_SUBROUTINE:
+        case GL_TESS_EVALUATION_SUBROUTINE:
+        case GL_GEOMETRY_SUBROUTINE:
+        case GL_FRAGMENT_SUBROUTINE:
+        case GL_COMPUTE_SUBROUTINE:
+        case GL_VERTEX_SUBROUTINE_UNIFORM:
+        case GL_TESS_CONTROL_SUBROUTINE_UNIFORM:
+        case GL_TESS_EVALUATION_SUBROUTINE_UNIFORM:
+        case GL_GEOMETRY_SUBROUTINE_UNIFORM:
+        case GL_FRAGMENT_SUBROUTINE_UNIFORM:
+        case GL_COMPUTE_SUBROUTINE_UNIFORM:
+            return true;
+        default:
+            return false;
+        }
+    }
+
+    Bool IsNamedInterface(GLenum programInterface) {
+        // GL 4.6 §7.3.1.2: the two buffer interfaces have no resource names, and asking for
+        // one is INVALID_ENUM (deliberately asymmetric with GetProgramInterfaceiv, which
+        // does count them).
+        return IsInterfaceEnum(programInterface) && programInterface != GL_ATOMIC_COUNTER_BUFFER &&
+               programInterface != GL_TRANSFORM_FEEDBACK_BUFFER;
+    }
+
+    Bool InterfaceHasLocations(GLenum programInterface) {
+        switch (programInterface) {
+        case GL_UNIFORM:
+        case GL_PROGRAM_INPUT:
+        case GL_PROGRAM_OUTPUT:
+        case GL_VERTEX_SUBROUTINE_UNIFORM:
+        case GL_TESS_CONTROL_SUBROUTINE_UNIFORM:
+        case GL_TESS_EVALUATION_SUBROUTINE_UNIFORM:
+        case GL_GEOMETRY_SUBROUTINE_UNIFORM:
+        case GL_FRAGMENT_SUBROUTINE_UNIFORM:
+        case GL_COMPUTE_SUBROUTINE_UNIFORM:
+            return true;
+        default:
+            return false;
+        }
+    }
+
+    Bool IsResourceProp(GLenum prop) {
+        switch (prop) {
+        case GL_NAME_LENGTH:
+        case GL_TYPE:
+        case GL_ARRAY_SIZE:
+        case GL_OFFSET:
+        case GL_BLOCK_INDEX:
+        case GL_ARRAY_STRIDE:
+        case GL_MATRIX_STRIDE:
+        case GL_IS_ROW_MAJOR:
+        case GL_ATOMIC_COUNTER_BUFFER_INDEX:
+        case GL_BUFFER_BINDING:
+        case GL_BUFFER_DATA_SIZE:
+        case GL_NUM_ACTIVE_VARIABLES:
+        case GL_ACTIVE_VARIABLES:
+        case GL_REFERENCED_BY_VERTEX_SHADER:
+        case GL_REFERENCED_BY_TESS_CONTROL_SHADER:
+        case GL_REFERENCED_BY_TESS_EVALUATION_SHADER:
+        case GL_REFERENCED_BY_GEOMETRY_SHADER:
+        case GL_REFERENCED_BY_FRAGMENT_SHADER:
+        case GL_REFERENCED_BY_COMPUTE_SHADER:
+        case GL_TOP_LEVEL_ARRAY_SIZE:
+        case GL_TOP_LEVEL_ARRAY_STRIDE:
+        case GL_LOCATION:
+        case GL_LOCATION_INDEX:
+        case GL_IS_PER_PATCH:
+        case GL_LOCATION_COMPONENT:
+        case GL_TRANSFORM_FEEDBACK_BUFFER_INDEX:
+        case GL_TRANSFORM_FEEDBACK_BUFFER_STRIDE:
+        case GL_NUM_COMPATIBLE_SUBROUTINES:
+        case GL_COMPATIBLE_SUBROUTINES:
+            return true;
+        default:
+            return false;
+        }
+    }
+
+    // GL 4.6 Table 7.2, transcribed row by row: which interfaces each property applies to.
+    // Too tight a table turns a currently-answered prop into a fresh INVALID_OPERATION, so
+    // the rows below are deliberately no narrower than the spec's.
+    Bool InterfaceSupportsProp(GLenum programInterface, GLenum prop) {
+        const Bool isSubroutine =
+            programInterface == GL_VERTEX_SUBROUTINE || programInterface == GL_TESS_CONTROL_SUBROUTINE ||
+            programInterface == GL_TESS_EVALUATION_SUBROUTINE || programInterface == GL_GEOMETRY_SUBROUTINE ||
+            programInterface == GL_FRAGMENT_SUBROUTINE || programInterface == GL_COMPUTE_SUBROUTINE;
+        const Bool isSubroutineUniform =
+            programInterface == GL_VERTEX_SUBROUTINE_UNIFORM ||
+            programInterface == GL_TESS_CONTROL_SUBROUTINE_UNIFORM ||
+            programInterface == GL_TESS_EVALUATION_SUBROUTINE_UNIFORM ||
+            programInterface == GL_GEOMETRY_SUBROUTINE_UNIFORM ||
+            programInterface == GL_FRAGMENT_SUBROUTINE_UNIFORM || programInterface == GL_COMPUTE_SUBROUTINE_UNIFORM;
+
+        switch (prop) {
+        case GL_NAME_LENGTH:
+            return programInterface != GL_ATOMIC_COUNTER_BUFFER && programInterface != GL_TRANSFORM_FEEDBACK_BUFFER;
+        case GL_TYPE:
+        case GL_ARRAY_SIZE:
+            return programInterface == GL_UNIFORM || programInterface == GL_PROGRAM_INPUT ||
+                   programInterface == GL_PROGRAM_OUTPUT || programInterface == GL_BUFFER_VARIABLE ||
+                   programInterface == GL_TRANSFORM_FEEDBACK_VARYING ||
+                   (prop == GL_ARRAY_SIZE && isSubroutineUniform);
+        case GL_OFFSET:
+            return programInterface == GL_UNIFORM || programInterface == GL_BUFFER_VARIABLE ||
+                   programInterface == GL_TRANSFORM_FEEDBACK_VARYING;
+        case GL_BLOCK_INDEX:
+        case GL_ARRAY_STRIDE:
+        case GL_MATRIX_STRIDE:
+        case GL_IS_ROW_MAJOR:
+            return programInterface == GL_UNIFORM || programInterface == GL_BUFFER_VARIABLE;
+        case GL_ATOMIC_COUNTER_BUFFER_INDEX:
+            return programInterface == GL_UNIFORM;
+        case GL_BUFFER_BINDING:
+        case GL_NUM_ACTIVE_VARIABLES:
+        case GL_ACTIVE_VARIABLES:
+            // Table 7.2 lists GL_TRANSFORM_FEEDBACK_BUFFER on these three rows too. This
+            // implementation enumerates no resources on that interface, so the query still
+            // ends in an error - but INVALID_VALUE for the out-of-range index, not the
+            // INVALID_OPERATION a narrower table would invent.
+            return programInterface == GL_UNIFORM_BLOCK || programInterface == GL_ATOMIC_COUNTER_BUFFER ||
+                   programInterface == GL_SHADER_STORAGE_BLOCK ||
+                   programInterface == GL_TRANSFORM_FEEDBACK_BUFFER;
+        case GL_BUFFER_DATA_SIZE:
+            return programInterface == GL_UNIFORM_BLOCK || programInterface == GL_ATOMIC_COUNTER_BUFFER ||
+                   programInterface == GL_SHADER_STORAGE_BLOCK;
+        case GL_REFERENCED_BY_VERTEX_SHADER:
+        case GL_REFERENCED_BY_TESS_CONTROL_SHADER:
+        case GL_REFERENCED_BY_TESS_EVALUATION_SHADER:
+        case GL_REFERENCED_BY_GEOMETRY_SHADER:
+        case GL_REFERENCED_BY_FRAGMENT_SHADER:
+        case GL_REFERENCED_BY_COMPUTE_SHADER:
+            return programInterface == GL_UNIFORM || programInterface == GL_UNIFORM_BLOCK ||
+                   programInterface == GL_ATOMIC_COUNTER_BUFFER || programInterface == GL_BUFFER_VARIABLE ||
+                   programInterface == GL_SHADER_STORAGE_BLOCK || programInterface == GL_PROGRAM_INPUT ||
+                   programInterface == GL_PROGRAM_OUTPUT || isSubroutineUniform;
+        case GL_TOP_LEVEL_ARRAY_SIZE:
+        case GL_TOP_LEVEL_ARRAY_STRIDE:
+            return programInterface == GL_BUFFER_VARIABLE;
+        case GL_LOCATION:
+            return InterfaceHasLocations(programInterface);
+        case GL_LOCATION_INDEX:
+            return programInterface == GL_PROGRAM_OUTPUT;
+        case GL_IS_PER_PATCH:
+        case GL_LOCATION_COMPONENT:
+            return programInterface == GL_PROGRAM_INPUT || programInterface == GL_PROGRAM_OUTPUT;
+        case GL_TRANSFORM_FEEDBACK_BUFFER_INDEX:
+            return programInterface == GL_TRANSFORM_FEEDBACK_VARYING;
+        case GL_TRANSFORM_FEEDBACK_BUFFER_STRIDE:
+            return programInterface == GL_TRANSFORM_FEEDBACK_BUFFER;
+        case GL_NUM_COMPATIBLE_SUBROUTINES:
+        case GL_COMPATIBLE_SUBROUTINES:
+            return isSubroutineUniform;
+        default:
+            (void)isSubroutine;
+            return false;
+        }
+    }
+
+    Int GetActiveResourceCount(ProgramObject& program, GLenum programInterface) {
+        const Model model = BuildModel(program);
+        return static_cast<Int>(Select(model, programInterface).size());
+    }
+
+    Int GetMaxNameLength(ProgramObject& program, GLenum programInterface) {
+        if (!IsNamedInterface(programInterface)) return 0;
+        const Model model = BuildModel(program);
+        SizeT longest = 0;
+        for (const Resource& resource : Select(model, programInterface)) {
+            longest = std::max(longest, resource.name.length() + 1);
+        }
+        return static_cast<Int>(longest);
+    }
+
+    Int GetMaxNumActiveVariables(ProgramObject& program, GLenum programInterface) {
+        const Model model = BuildModel(program);
+        SizeT longest = 0;
+        for (const Resource& resource : Select(model, programInterface)) {
+            longest = std::max(longest, resource.activeVariables.size());
+        }
+        return static_cast<Int>(longest);
+    }
+
+    GLuint GetResourceIndex(ProgramObject& program, GLenum programInterface, const char* name) {
+        if (name == nullptr || name[0] == '\0') return GL_INVALID_INDEX;
+        const Model model = BuildModel(program);
+        const ResourceList& resources = Select(model, programInterface);
+        const String query = name;
+        // The layout controls of an interleaved capture are enumerable but not addressable
+        // by name (GL 4.6 §7.3.1.1).
+        if (programInterface == GL_TRANSFORM_FEEDBACK_VARYING &&
+            (query == "gl_NextBuffer" ||
+             (query.size() == 18 && query.compare(0, 17, "gl_SkipComponents") == 0))) {
+            return GL_INVALID_INDEX;
+        }
+        for (SizeT i = 0; i < resources.size(); ++i) {
+            if (NamesMatch(resources[i].name, query)) return static_cast<GLuint>(i);
+        }
+        return GL_INVALID_INDEX;
+    }
+
+    Bool GetResourceName(ProgramObject& program, GLenum programInterface, GLuint index, String& outName) {
+        const Model model = BuildModel(program);
+        const ResourceList& resources = Select(model, programInterface);
+        if (index >= resources.size()) return false;
+        outName = resources[index].name;
+        return true;
+    }
+
+    Bool GetResourceProp(ProgramObject& program, GLenum programInterface, GLuint index, GLenum prop,
+                         Vector<GLint>& outValues) {
+        const Model model = BuildModel(program);
+        const ResourceList& resources = Select(model, programInterface);
+        if (index >= resources.size()) return false;
+        const Resource& resource = resources[index];
+
+        const auto referencedBy = [&resource](EShLanguage stage) {
+            return (resource.stages & static_cast<Uint32>(1u << stage)) != 0 ? GL_TRUE : GL_FALSE;
+        };
+
+        switch (prop) {
+        case GL_NAME_LENGTH:
+            outValues.push_back(static_cast<GLint>(resource.name.length() + 1));
+            break;
+        case GL_TYPE:
+            outValues.push_back(static_cast<GLint>(resource.type));
+            break;
+        case GL_ARRAY_SIZE:
+            outValues.push_back(resource.arraySize);
+            break;
+        case GL_OFFSET:
+            outValues.push_back(resource.offset);
+            break;
+        case GL_BLOCK_INDEX:
+            outValues.push_back(resource.blockIndex);
+            break;
+        case GL_ARRAY_STRIDE:
+            outValues.push_back(resource.arrayStride);
+            break;
+        case GL_MATRIX_STRIDE:
+            outValues.push_back(resource.matrixStride);
+            break;
+        case GL_IS_ROW_MAJOR:
+            outValues.push_back(resource.isRowMajor);
+            break;
+        case GL_ATOMIC_COUNTER_BUFFER_INDEX:
+            outValues.push_back(resource.atomicCounterBufferIndex);
+            break;
+        case GL_BUFFER_BINDING:
+            outValues.push_back(resource.bufferBinding);
+            break;
+        case GL_BUFFER_DATA_SIZE:
+            outValues.push_back(resource.bufferDataSize);
+            break;
+        case GL_NUM_ACTIVE_VARIABLES:
+            outValues.push_back(static_cast<GLint>(resource.activeVariables.size()));
+            break;
+        case GL_ACTIVE_VARIABLES:
+            for (const GLuint variable : resource.activeVariables) outValues.push_back(static_cast<GLint>(variable));
+            break;
+        case GL_REFERENCED_BY_VERTEX_SHADER:
+            outValues.push_back(referencedBy(EShLangVertex));
+            break;
+        case GL_REFERENCED_BY_TESS_CONTROL_SHADER:
+            outValues.push_back(referencedBy(EShLangTessControl));
+            break;
+        case GL_REFERENCED_BY_TESS_EVALUATION_SHADER:
+            outValues.push_back(referencedBy(EShLangTessEvaluation));
+            break;
+        case GL_REFERENCED_BY_GEOMETRY_SHADER:
+            outValues.push_back(referencedBy(EShLangGeometry));
+            break;
+        case GL_REFERENCED_BY_FRAGMENT_SHADER:
+            outValues.push_back(referencedBy(EShLangFragment));
+            break;
+        case GL_REFERENCED_BY_COMPUTE_SHADER:
+            outValues.push_back(referencedBy(EShLangCompute));
+            break;
+        case GL_TOP_LEVEL_ARRAY_SIZE:
+            outValues.push_back(resource.topLevelArraySize);
+            break;
+        case GL_TOP_LEVEL_ARRAY_STRIDE:
+            outValues.push_back(resource.topLevelArrayStride);
+            break;
+        case GL_LOCATION:
+            outValues.push_back(resource.location);
+            break;
+        case GL_LOCATION_INDEX:
+            outValues.push_back(resource.locationIndex);
+            break;
+        case GL_IS_PER_PATCH:
+            outValues.push_back(resource.isPerPatch);
+            break;
+        case GL_LOCATION_COMPONENT:
+            outValues.push_back(0);
+            break;
+        case GL_TRANSFORM_FEEDBACK_BUFFER_INDEX:
+            outValues.push_back(resource.xfbBufferIndex);
+            break;
+        default:
+            outValues.push_back(0);
+            break;
+        }
+        return true;
+    }
+
+    GLint GetResourceLocation(ProgramObject& program, GLenum programInterface, const char* name) {
+        if (name == nullptr || name[0] == '\0') return -1;
+        const String query = name;
+
+        String base;
+        Uint element = 0;
+        Bool malformed = false;
+        const Bool subscripted = SplitTrailingSubscript(query, base, element, malformed);
+        if (malformed) return -1;
+
+        const Model model = BuildModel(program);
+        const ResourceList& resources = Select(model, programInterface);
+        for (const Resource& resource : resources) {
+            if (NamesMatch(resource.name, query)) return resource.location;
+        }
+        if (!subscripted || element == 0) return -1;
+        // "d[1]" addresses the second element of an array resource enumerated as "d[0]".
+        for (const Resource& resource : resources) {
+            if (!NamesMatch(resource.name, base)) continue;
+            if (resource.location < 0 || static_cast<GLint>(element) >= resource.arraySize) return -1;
+            return resource.location + static_cast<GLint>(element);
+        }
+        return -1;
+    }
+
+    GLint GetResourceLocationIndex(ProgramObject& program, GLenum programInterface, const char* name) {
+        if (programInterface != GL_PROGRAM_OUTPUT || name == nullptr || name[0] == '\0') return -1;
+        const String query = name;
+        String base;
+        Uint element = 0;
+        Bool malformed = false;
+        const Bool subscripted = SplitTrailingSubscript(query, base, element, malformed);
+        if (malformed) return -1;
+
+        const Model model = BuildModel(program);
+        for (const Resource& resource : model.programOutputs) {
+            if (NamesMatch(resource.name, query)) return resource.locationIndex;
+        }
+        if (!subscripted) return -1;
+        for (const Resource& resource : model.programOutputs) {
+            if (!NamesMatch(resource.name, base)) continue;
+            if (resource.location < 0 || static_cast<GLint>(element) >= resource.arraySize) return -1;
+            return resource.locationIndex;
+        }
+        return -1;
+    }
+} // namespace MobileGL::MG_Impl::GLImpl::ProgramInterface

MobileGL/MG_Impl/GLImpl/Program/ProgramInterface.h

@@ -0,0 +1,66 @@
+// MobileGL - MobileGL/MG_Impl/GLImpl/Program/ProgramInterface.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+
+namespace MobileGL::MG_State::GLState {
+    class ProgramObject;
+}
+
+// The GL program interface (ARB_program_interface_query / GL 4.3 §7.3.1) as a frontend
+// resource model.
+//
+// WHY IT IS HERE AND NOT IN A BACKEND. glGetProgramResource* describes the program the
+// APPLICATION wrote, in the application's namespace. Neither backend program is in that
+// namespace: DirectGLES compiles SPIRV-Cross-generated ESSL where default-block uniforms
+// live inside the synthesized MGL_GLOBAL_UBO (so a GL_UNIFORM location query against it is
+// structurally -1) and stage in/out names are rewritten; DirectVulkan has no GL-level
+// reflection at all and can only re-derive a partial, diverging copy. The one authoritative
+// source is the frontend glslang reflection a link already produced, which is the same
+// place glGetActiveUniform answers from. This layer generalizes that rule to every
+// interface, so the six entry points never consult gBackendFunctionsTable.
+//
+// NAMING RULES LIVE HERE, NOT IN ProgramObject. The interface query spells resources
+// differently from glGetActiveUniform / glGetActiveAttrib (an array is "name[0]", a lookup
+// accepts both "name" and "name[0]", a subscript must be a strict decimal). Those two
+// getters are what GL30-33 exercises and they must not move, so every normalization is
+// applied on the way in and out of THIS file.
+namespace MobileGL::MG_Impl::GLImpl::ProgramInterface {
+    using ProgramObject = MG_State::GLState::ProgramObject;
+
+    // <programInterface> is one of the GL 4.6 Table 7.1 interfaces.
+    Bool IsInterfaceEnum(GLenum programInterface);
+    // Interfaces whose resources have names (everything except GL_ATOMIC_COUNTER_BUFFER).
+    Bool IsNamedInterface(GLenum programInterface);
+    // <prop> is a property token GetProgramResourceiv knows at all (else GL_INVALID_ENUM).
+    Bool IsResourceProp(GLenum prop);
+    // <prop> applies to <programInterface> (else GL_INVALID_OPERATION).
+    Bool InterfaceSupportsProp(GLenum programInterface, GLenum prop);
+    // Interfaces GetProgramResourceLocation accepts (else GL_INVALID_ENUM).
+    Bool InterfaceHasLocations(GLenum programInterface);
+
+    // GL_ACTIVE_RESOURCES / GL_MAX_NAME_LENGTH / GL_MAX_NUM_ACTIVE_VARIABLES. All three
+    // report zero for an interface this implementation cannot enumerate and for a program
+    // that has not linked successfully - which is what the spec requires of a program with
+    // no active resources.
+    Int GetActiveResourceCount(ProgramObject& program, GLenum programInterface);
+    Int GetMaxNameLength(ProgramObject& program, GLenum programInterface);
+    Int GetMaxNumActiveVariables(ProgramObject& program, GLenum programInterface);
+
+    // GL_INVALID_INDEX when <name> names no active resource of the interface.
+    GLuint GetResourceIndex(ProgramObject& program, GLenum programInterface, const char* name);
+    // False when <index> is out of range for the interface (the caller raises INVALID_VALUE).
+    Bool GetResourceName(ProgramObject& program, GLenum programInterface, GLuint index, String& outName);
+    // Appends the value(s) of <prop> for the resource; GL_ACTIVE_VARIABLES appends several.
+    // False when <index> is out of range.
+    Bool GetResourceProp(ProgramObject& program, GLenum programInterface, GLuint index, GLenum prop,
+                         Vector<GLint>& outValues);
+    GLint GetResourceLocation(ProgramObject& program, GLenum programInterface, const char* name);
+    GLint GetResourceLocationIndex(ProgramObject& program, GLenum programInterface, const char* name);
+} // namespace MobileGL::MG_Impl::GLImpl::ProgramInterface

MobileGL/MG_Impl/GLImpl/Sampler/GL_Sampler.cpp

@@ -8,6 +8,7 @@
 
 #include "GL_Sampler.h"
 #include "Validators.h"
+#include "../Getter/GL_Getter.h"
 #include <MG_State/GLState/Core.h>
 #include <MG_Util/Converters/GLToMG/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
@@ -268,9 +269,20 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
+    // The number of texture units a sampler may be bound to. GL 3.3 core 3.8.2 names
+    // GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, which is what the backend advertises; the frontend's
+    // MAX_TEXTURE_IMAGE_UNITS is only the capacity of the unit array, so it is a clamp on the
+    // answer and never the answer itself - gating on it alone accepts every unit up to 192 no
+    // matter what the driver reports.
+    static GLint GetSamplerBindableTextureUnitCount() {
+        GLint maxTextureUnits = 0;
+        GetIntegerv(GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, &maxTextureUnits);
+        return std::min<GLint>(std::max(maxTextureUnits, 0), MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS);
+    }
+
     void BindSampler_State(GLuint unit, GLuint sampler) {
         MGLOG_D("BindSampler_State: unit = %u, sampler = %u", unit, sampler);
-        if (unit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
+        if (static_cast<Uint64>(unit) >= static_cast<Uint64>(GetSamplerBindableTextureUnitCount())) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "BindSampler", "texture unit out of range"));
@@ -309,6 +321,20 @@ namespace MobileGL::MG_Impl::GLImpl {
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "BindSamplers", "count must be non-negative"));
             return;
         }
+        // ARB_multi_bind: the whole [first, first + count) range is checked up front and a
+        // range that runs past the last texture unit is INVALID_OPERATION - not the
+        // INVALID_VALUE the single-bind BindSampler_State reports per element, and nothing is
+        // bound when it fails. Both gates read the same limit (see
+        // GetSamplerBindableTextureUnitCount), so an out-of-range multi-bind can no longer slip
+        // past this check and be caught one element at a time with the wrong error class.
+        const GLint maxTextureUnits = GetSamplerBindableTextureUnitCount();
+        if (static_cast<Uint64>(first) + static_cast<Uint64>(count) > static_cast<Uint64>(maxTextureUnits)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "BindSamplers",
+                                             "first + count exceeds the number of texture units."));
+            return;
+        }
 
         for (GLsizei i = 0; i < count; ++i) {
             BindSampler_State(first + i, samplers ? samplers[i] : 0);

MobileGL/MG_Impl/GLImpl/Sampler/Validators.cpp

@@ -75,7 +75,11 @@ namespace MobileGL::MG_Impl::GLImpl::SamplerImpl {
             break;
 
         case GL_TEXTURE_COMPARE_FUNC:
-            if (param < GL_LEQUAL || param > GL_ALWAYS) {
+            // The eight depth-compare functions are contiguous from GL_NEVER (0x0200) to
+            // GL_ALWAYS (0x0207); GL_LEQUAL sits in the middle of that block, so starting
+            // the range there rejected NEVER/LESS/EQUAL and let GREATER/NOTEQUAL/GEQUAL
+            // through only by accident of them being above LEQUAL.
+            if (param < GL_NEVER || param > GL_ALWAYS) {
                 MG_State::pGLContext->RecordError(ErrorCode::InvalidEnum,
                                                   MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "ValidateSamplerParam",
                                                                                "Invalid compare function parameter"));

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.cpp

@@ -86,17 +86,65 @@ namespace MobileGL::MG_Impl::GLImpl {
             return false;
         }
 
+        // DSA emulation: the by-name entry points are implemented by putting the named texture
+        // on the active unit's slot for their target, running the classic bound-texture code,
+        // then putting the previous binding back.
+        //
+        // Both of those binds are REAL changes to "which texture is bound at this unit" for as
+        // long as `fn` runs, so both have to move the texture bind generation. Backends memoise
+        // per-unit work keyed on that generation and BORROW the binding slot (they hold a
+        // pointer to the slot's shared_ptr, not a copy); a slot swap the generation never saw
+        // let such a memo replay texture A's backend twin against texture B now sitting in the
+        // slot - which re-specified A's backend storage with B's shape and silently destroyed
+        // A's GPU-rendered contents (Minecraft's lightmap, blanked by a by-name upload to an
+        // Iris shadow map, which then discarded every glyph).
+        //
+        // The generation is bumped directly rather than through NoteTextureUnitTouched because
+        // the touched-unit HIGH-WATER MARK must NOT move: glActiveTexture does not advance it,
+        // so a DSA-only app would otherwise have every later draw walk up to the highest unit it
+        // ever aimed a by-name call at. Not advancing it is also sufficient - a unit above the
+        // mark is outside every memo's coverage and outside the epoch walk, so nothing can
+        // observe the transient swap there; at or below it, the bump is exactly what makes the
+        // epoch re-derive. Bumping only on a real change keeps the very common redundant case (a
+        // by-name call on the texture already bound to the active unit) free.
+        //
+        // The restore is a scope guard because `fn` can throw (the unsupported-state paths use
+        // THROW_EXCEPTION): leaking the temporary binding would leave the wrong texture bound to
+        // a live unit for the rest of the context's life.
         template <typename Fn>
         void WithTemporarilyBoundNamedTexture(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
                                              Fn&& fn) {
             if (!textureObject) return;
 
-            auto& activeUnit = MG_State::pGLContext->GetTextureUnitObject(MG_State::pGLContext->GetActiveTextureUnit());
+            const Int activeUnitIndex = MG_State::pGLContext->GetActiveTextureUnit();
+            auto& activeUnit = MG_State::pGLContext->GetTextureUnitObject(activeUnitIndex);
             auto& bindingSlot = activeUnit.GetBindingSlot(textureObject->GetTarget());
             const auto previousBinding = bindingSlot.GetBoundObject();
-            bindingSlot.Bind(textureObject);
+
+            using SlotType = std::remove_reference_t<decltype(bindingSlot)>;
+            class ScopedSlotRestore {
+            public:
+                ScopedSlotRestore(SlotType& slot, SharedPtr<MG_State::GLState::ITextureObject> previous)
+                    : m_slot(slot), m_previous(Move(previous)) {}
+                ~ScopedSlotRestore() {
+                    if (m_slot.Bind(m_previous)) {
+                        MG_State::pGLContext->BumpTextureBindGeneration();
+                    }
+                }
+                ScopedSlotRestore(const ScopedSlotRestore&) = delete;
+                ScopedSlotRestore& operator=(const ScopedSlotRestore&) = delete;
+
+            private:
+                SlotType& m_slot;
+                SharedPtr<MG_State::GLState::ITextureObject> m_previous;
+            };
+
+            if (bindingSlot.Bind(textureObject)) {
+                MG_State::pGLContext->BumpTextureBindGeneration();
+            }
+            ScopedSlotRestore restore(bindingSlot, previousBinding);
+
             fn(MG_Util::ConvertTextureTargetToGLEnum(textureObject->GetTarget()));
-            bindingSlot.Bind(previousBinding);
         }
 
         SizeT ComputeTextureStorageByteSize(TextureInternalFormat textureInternalFormat, GLsizei width, GLsizei height,

MobileGL/MG_Impl/GLImpl/VertexArray/GL_VertexArray.cpp

@@ -106,6 +106,21 @@ namespace MobileGL::MG_Impl::GLImpl {
             return pname == GL_CURRENT_VERTEX_ATTRIB;
         }
 
+        // The two ARB_vertex_attrib_binding per-attribute queries. They do not live on the
+        // resolved VertexAttribute (which is the flat, already-combined view) but on the VAO's
+        // binding-point mapping, so they need the object, not the attribute.
+        static bool TryGetVertexAttribBindingQuery(GLuint index, GLenum pname, GLint& out) {
+            if (pname != GL_VERTEX_ATTRIB_BINDING && pname != GL_VERTEX_ATTRIB_RELATIVE_OFFSET) return false;
+            const auto& vao = MG_State::pGLContext->GetBoundVertexArray();
+            if (!vao) {
+                out = 0;
+                return true;
+            }
+            out = pname == GL_VERTEX_ATTRIB_BINDING ? static_cast<GLint>(vao->GetAttributeBindingIndex(index))
+                                                    : static_cast<GLint>(vao->GetAttributeRelativeOffset(index));
+            return true;
+        }
+
         // The stride a pointer-style call gives its binding point: the argument when it is non-zero,
         // otherwise the tightly packed element size (GL 4.6 core 10.3.2). A packed 2_10_10_10 or
         // 10F_11F_11F attribute is one 32-bit word regardless of its component count.
@@ -179,6 +194,11 @@ namespace MobileGL::MG_Impl::GLImpl {
             case GL_VERTEX_ATTRIB_ARRAY_LONG:
             case GL_VERTEX_ATTRIB_ARRAY_DIVISOR:
             case GL_VERTEX_ATTRIB_ARRAY_POINTER:
+            // ARB_vertex_attrib_binding (core since GL 4.3). The binding-point view is real
+            // state on the VAO (GetAttributeBindingIndex / GetAttributeRelativeOffset), so
+            // both of its per-attribute queries are answerable.
+            case GL_VERTEX_ATTRIB_BINDING:
+            case GL_VERTEX_ATTRIB_RELATIVE_OFFSET:
                 return true;
             default:
                 MG_State::pGLContext->RecordError(
@@ -944,6 +964,13 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_VERTEX_ATTRIB_ARRAY_DIVISOR:
             params[0] = static_cast<GLfloat>(attr->Divisor);
             return;
+        case GL_VERTEX_ATTRIB_BINDING:
+        case GL_VERTEX_ATTRIB_RELATIVE_OFFSET: {
+            GLint value = 0;
+            TryGetVertexAttribBindingQuery(index, pname, value);
+            params[0] = static_cast<GLfloat>(value);
+            return;
+        }
         default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
@@ -1007,6 +1034,13 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_VERTEX_ATTRIB_ARRAY_DIVISOR:
             params[0] = static_cast<GLdouble>(attr->Divisor);
             return;
+        case GL_VERTEX_ATTRIB_BINDING:
+        case GL_VERTEX_ATTRIB_RELATIVE_OFFSET: {
+            GLint value = 0;
+            TryGetVertexAttribBindingQuery(index, pname, value);
+            params[0] = static_cast<GLdouble>(value);
+            return;
+        }
         default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
@@ -1066,6 +1100,10 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_VERTEX_ATTRIB_ARRAY_DIVISOR:
             params[0] = static_cast<GLint>(attr->Divisor);
             return;
+        case GL_VERTEX_ATTRIB_BINDING:
+        case GL_VERTEX_ATTRIB_RELATIVE_OFFSET:
+            TryGetVertexAttribBindingQuery(index, pname, params[0]);
+            return;
         default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
@@ -1204,6 +1242,9 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_VERTEX_ATTRIB_RELATIVE_OFFSET:
             *param = static_cast<GLint>(vao->GetAttributeRelativeOffset(index));
             return;
+        case GL_VERTEX_ATTRIB_BINDING:
+            *param = static_cast<GLint>(vao->GetAttributeBindingIndex(index));
+            return;
         default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,

MobileGL/MG_Impl/GLImpl/VertexArray/Validators.cpp

@@ -147,6 +147,30 @@ namespace MobileGL::MG_Impl::GLImpl::VertexArrayImpl {
             return false;
         }
 
+        // The integer path takes exactly the six signed/unsigned integer types (GL 4.6
+        // core 10.3.2): BYTE, UNSIGNED_BYTE, SHORT, UNSIGNED_SHORT, INT, UNSIGNED_INT.
+        // A blacklist could not express that: GL_FLOAT, GL_HALF_FLOAT,
+        // GL_DOUBLE and GL_FIXED all convert to a perfectly valid DataType, so they slipped
+        // through and were recorded as integer attributes.
+        if (integerPath) {
+            switch (type) {
+            case DataType::Int8:
+            case DataType::Uint8:
+            case DataType::Int16:
+            case DataType::Uint16:
+            case DataType::Int32:
+            case DataType::Uint32:
+                break;
+            default:
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidEnum,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", fn,
+                        std::format("Type is not an integer vertex attribute type (attribute {}).", index)));
+                return false;
+            }
+        }
+
         if (sizeRaw == static_cast<GLint>(GL_BGRA)) {
             // GL_BGRA is a float-path-only size: it needs GL_UNSIGNED_BYTE or a 2_10_10_10 type and
             // normalized == GL_TRUE. On the integer path it is simply an out-of-range size.

MobileGL/MG_IntegrationTest/CMakeLists.txt

@@ -0,0 +1,270 @@
+cmake_minimum_required(VERSION 3.24)
+
+# MobileGL headless GPU integration tests.
+#
+# These are not unit tests: each scenario brings up a real EGL context on a
+# pbuffer, renders real frames through a real backend and asserts on
+# glReadPixels output. They need a GPU, so the module is OFF by default
+# (MOBILEGL_BUILD_INTEGRATION_TEST) and every scenario skips cleanly - never
+# fails, never hangs - on a machine without one. "Cleanly" is not a hope: the
+# harness runs the whole bring-up in a forked child first, because MobileGL
+# ABORTS rather than returning an error on an unusable platform (HeadlessGL.cpp).
+#
+# A clean skip is also indistinguishable from a pass, so set
+# MOBILEGL_ITEST_REQUIRE_GPU wherever the machine is supposed to have a GPU.
+#
+# Backend selection is latched at initialization from MOBILEGL_BACKEND_TYPE, so
+# one process is one backend: the same binary is registered twice, once per
+# backend, under the `integration-gpu` label.
+
+message(STATUS "Generating build files for MobileGL Integration Test...")
+
+set(CMAKE_CXX_STANDARD 23)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+set(MGL_ITEST_ROOT ${CMAKE_CURRENT_LIST_DIR}/../..)
+
+# Only meaningful where MobileGL_s exists (i.e. not Android).
+if (NOT TARGET MobileGL_s)
+    message(STATUS "MobileGL_s is not available; skipping the integration test module")
+    return()
+endif()
+
+# MG_Test already pulls googletest in when MOBILEGL_BUILD_TEST is ON. Stand on
+# our own feet when it is not, so this module can be built by itself.
+if (NOT TARGET GTest::gtest)
+    include(FetchContent)
+    FetchContent_Declare(
+        googletest
+        GIT_REPOSITORY https://github.com/google/googletest.git
+        GIT_TAG        v1.17.0
+    )
+    set(gtest_force_shared_crt ON CACHE BOOL "" FORCE)
+    FetchContent_MakeAvailable(googletest)
+endif()
+
+add_executable(MobileGLIntegrationTest
+    Main.cpp
+    Harness/HeadlessGL.cpp
+    Scenarios/OrientationScenario.cpp
+    Scenarios/CrossFrameBufferScenario.cpp
+    Scenarios/ResidentIndexScenario.cpp
+    Scenarios/MultiDrawScenario.cpp
+    Scenarios/AsyncCompileScenario.cpp
+    Scenarios/XfbAfterClipDistanceScenario.cpp
+    Scenarios/ThreeChannelAttachmentScenario.cpp
+)
+
+target_include_directories(MobileGLIntegrationTest PRIVATE
+    ${MGL_ITEST_ROOT}/include
+    ${MGL_ITEST_ROOT}/MobileGL
+)
+
+# gtest, not gtest_main: Main.cpp installs the harness banner itself.
+target_link_libraries(MobileGLIntegrationTest PRIVATE
+    GTest::gtest
+    MobileGL_s
+)
+
+if (MSVC)
+    # Same reason as MG_Test/Backend/DirectVulkan: the GLES headers declare gl*
+    # as dllimport on Windows, so the in-library GL entry-point definitions only
+    # resolve if the whole static library is part of the link.
+    target_link_options(MobileGLIntegrationTest PRIVATE /WHOLEARCHIVE:MobileGL_s)
+endif()
+target_compile_definitions(MobileGLIntegrationTest PRIVATE -DNOMINMAX)
+
+# --- ctest wiring --------------------------------------------------------
+# A bare libEGL on a glvnd box resolves to whatever vendor comes first, which is
+# usually Mesa/llvmpipe - a software rasteriser silently replacing the GPU under
+# a GPU test. Pin the vendor/ICD json the same way MG_Benchmark's
+# run_driver_bench.sh does.
+#
+# Leaving these empty is not a neutral default, it is the failure mode: an
+# unpinned libEGL lands on llvmpipe and the suite goes green having tested a
+# software rasteriser. So they are DETECTED here rather than defaulted to empty,
+# and an empty result is a loud warning.
+#
+# mgl_itest_find_driver_json(<outVar> <description> <glob> [<glob>...])
+#   Picks the first json a real hardware vendor owns, in preference order, and
+#   never picks a software rasteriser (llvmpipe / lavapipe / swrast) - landing on
+#   one of those silently is the exact accident this pinning exists to prevent.
+function(mgl_itest_find_driver_json outVar)
+    set(candidates "")
+    foreach(pattern IN LISTS ARGN)
+        file(GLOB matches "${pattern}")
+        list(APPEND candidates ${matches})
+    endforeach()
+    list(SORT candidates)
+    # Vendors ship an i686 json beside the x86_64 one and it sorts first. Pinning
+    # the wrong word size is worse than not pinning at all - the loader finds no
+    # driver and the whole suite skips - so drop the mismatched ones outright.
+    if (CMAKE_SIZEOF_VOID_P EQUAL 8)
+        list(FILTER candidates EXCLUDE REGEX "i686|i386")
+    else()
+        list(FILTER candidates EXCLUDE REGEX "x86_64|aarch64")
+    endif()
+    set(software "")
+    foreach(vendor IN ITEMS nvidia amdgpu amd radeon intel_hasvk intel broadcom freedreno panfrost)
+        foreach(candidate IN LISTS candidates)
+            get_filename_component(leaf "${candidate}" NAME)
+            string(TOLOWER "${leaf}" leaf)
+            if (leaf MATCHES "${vendor}")
+                set(${outVar} "${candidate}" PARENT_SCOPE)
+                return()
+            endif()
+        endforeach()
+    endforeach()
+    # Nothing recognised as hardware. Report the first non-software entry if there
+    # is one; otherwise report nothing, so the warning below fires.
+    foreach(candidate IN LISTS candidates)
+        get_filename_component(leaf "${candidate}" NAME)
+        string(TOLOWER "${leaf}" leaf)
+        if (NOT leaf MATCHES "lvp|llvmpipe|lavapipe|swrast|softpipe")
+            set(${outVar} "${candidate}" PARENT_SCOPE)
+            return()
+        endif()
+        set(software "${candidate}")
+    endforeach()
+    set(${outVar} "" PARENT_SCOPE)
+endfunction()
+
+set(MGL_ITEST_DETECTED_EGL_VENDOR "")
+set(MGL_ITEST_DETECTED_VK_ICD "")
+if (UNIX AND NOT APPLE AND NOT ANDROID)
+    mgl_itest_find_driver_json(MGL_ITEST_DETECTED_EGL_VENDOR
+        "/usr/share/glvnd/egl_vendor.d/*.json"
+        "/etc/glvnd/egl_vendor.d/*.json")
+    mgl_itest_find_driver_json(MGL_ITEST_DETECTED_VK_ICD
+        "/usr/share/vulkan/icd.d/*.json"
+        "/etc/vulkan/icd.d/*.json")
+endif()
+
+set(MOBILEGL_ITEST_EGL_VENDOR "${MGL_ITEST_DETECTED_EGL_VENDOR}" CACHE FILEPATH
+    "glvnd EGL vendor json to pin for the integration tests (empty: leave the loader alone)")
+set(MOBILEGL_ITEST_VK_ICD "${MGL_ITEST_DETECTED_VK_ICD}" CACHE FILEPATH
+    "Vulkan ICD json to pin for the DirectVulkan integration tests (empty: leave the loader alone)")
+
+if (MOBILEGL_ITEST_EGL_VENDOR)
+    message(STATUS "Integration tests: pinning EGL vendor ${MOBILEGL_ITEST_EGL_VENDOR}")
+else()
+    message(WARNING
+        "Integration tests: no EGL vendor json found or configured (MOBILEGL_ITEST_EGL_VENDOR is empty). "
+        "An unpinned libEGL on a glvnd system resolves to whichever vendor comes first, which is usually "
+        "Mesa/llvmpipe - the scenarios would then go green against a software rasteriser instead of the GPU. "
+        "Set -DMOBILEGL_ITEST_EGL_VENDOR=/usr/share/glvnd/egl_vendor.d/<vendor>.json.")
+endif()
+if (MOBILEGL_ITEST_VK_ICD)
+    message(STATUS "Integration tests: pinning Vulkan ICD ${MOBILEGL_ITEST_VK_ICD}")
+else()
+    message(WARNING
+        "Integration tests: no Vulkan ICD json found or configured (MOBILEGL_ITEST_VK_ICD is empty). "
+        "DirectVulkan would then load whichever ICD the loader enumerates first, quite possibly lavapipe. "
+        "Set -DMOBILEGL_ITEST_VK_ICD=/usr/share/vulkan/icd.d/<vendor>.json.")
+endif()
+
+# Turns "no usable GPU" from a clean skip into a failure - see ScenarioFixture.h.
+# Without it the integration-gpu label is unfalsifiable: a run that skipped every
+# scenario and a run that passed every scenario are the same green in ctest.
+option(MOBILEGL_ITEST_REQUIRE_GPU
+    "Fail (rather than skip) the integration scenarios when the headless harness is unusable" OFF)
+
+# DirectGLES asks the system EGL for a pbuffer config, and on Mesa the default
+# platform is not X11 unless it is said out loud (run_driver_bench.sh sets the
+# same variable). Wrong platform here is not a soft failure: eglCreatePbuffer
+# fails and every scenario skips.
+if (UNIX AND NOT APPLE AND NOT ANDROID)
+    set(MOBILEGL_ITEST_EGL_PLATFORM "x11" CACHE STRING
+        "EGL_PLATFORM for the integration tests (empty: leave the loader alone)")
+else()
+    set(MOBILEGL_ITEST_EGL_PLATFORM "" CACHE STRING
+        "EGL_PLATFORM for the integration tests (empty: leave the loader alone)")
+endif()
+
+set(MGL_ITEST_COMMON_ENV "")
+if (MOBILEGL_ITEST_EGL_VENDOR)
+    list(APPEND MGL_ITEST_COMMON_ENV "__EGL_VENDOR_LIBRARY_FILENAMES=${MOBILEGL_ITEST_EGL_VENDOR}")
+endif()
+if (MOBILEGL_ITEST_EGL_PLATFORM)
+    list(APPEND MGL_ITEST_COMMON_ENV "EGL_PLATFORM=${MOBILEGL_ITEST_EGL_PLATFORM}")
+endif()
+if (MOBILEGL_ITEST_REQUIRE_GPU)
+    list(APPEND MGL_ITEST_COMMON_ENV "MOBILEGL_ITEST_REQUIRE_GPU=1")
+endif()
+
+set(MGL_ITEST_VULKAN_ENV ${MGL_ITEST_COMMON_ENV})
+if (MOBILEGL_ITEST_VK_ICD)
+    list(APPEND MGL_ITEST_VULKAN_ENV "VK_ICD_FILENAMES=${MOBILEGL_ITEST_VK_ICD}")
+endif()
+
+# The ENVIRONMENT test property is itself a `;`-list, and gtest_discover_tests
+# forwards PROPERTIES as a flat list - so a plain `;`-joined value arrives as
+# four separate arguments and everything after the first is silently read as
+# another property name. Escaping the separators keeps the whole thing one list
+# element until set_tests_properties expands it back. Without this only
+# MOBILEGL_BACKEND_TYPE reaches the test and the vendor/ICD pinning is lost.
+function(mgl_itest_join_environment outVar)
+    set(joined "")
+    foreach(entry IN LISTS ARGN)
+        if (joined)
+            string(APPEND joined "\\;${entry}")
+        else()
+            set(joined "${entry}")
+        endif()
+    endforeach()
+    set(${outVar} "${joined}" PARENT_SCOPE)
+endfunction()
+
+mgl_itest_join_environment(MGL_ITEST_GLES_ENVIRONMENT
+    "MOBILEGL_BACKEND_TYPE=DirectGLES" ${MGL_ITEST_COMMON_ENV})
+mgl_itest_join_environment(MGL_ITEST_VULKAN_ENVIRONMENT
+    "MOBILEGL_BACKEND_TYPE=DirectVulkan" ${MGL_ITEST_VULKAN_ENV})
+mgl_itest_join_environment(MGL_ITEST_VULKAN_ASYNC_ENVIRONMENT
+    "MOBILEGL_BACKEND_TYPE=DirectVulkan" "MOBILEGL_ASYNC_SHADER_COMPILE=1" ${MGL_ITEST_VULKAN_ENV})
+
+# TIMEOUT on every entry: a GPU test that wedges must fail the run, not hang it.
+set(MGL_ITEST_TIMEOUT 120)
+
+include(GoogleTest)
+
+# Discovery runs `--gtest_list_tests`, which does not construct the harness and
+# so needs no GPU. One registration per backend; TEST_PREFIX keeps the two sets
+# of ctest names apart.
+gtest_discover_tests(MobileGLIntegrationTest
+    TEST_PREFIX "DirectGLES."
+    DISCOVERY_TIMEOUT 30
+    PROPERTIES
+        LABELS integration-gpu
+        TIMEOUT ${MGL_ITEST_TIMEOUT}
+        ENVIRONMENT "${MGL_ITEST_GLES_ENVIRONMENT}"
+)
+
+gtest_discover_tests(MobileGLIntegrationTest
+    TEST_PREFIX "DirectVulkan."
+    DISCOVERY_TIMEOUT 30
+    PROPERTIES
+        LABELS integration-gpu
+        TIMEOUT ${MGL_ITEST_TIMEOUT}
+        ENVIRONMENT "${MGL_ITEST_VULKAN_ENVIRONMENT}"
+)
+
+# A third registration, of ONE scenario, with asynchronous shader compilation
+# pinned on. Not a second code path in the renderer: a second ALLOCATION pattern.
+# The async pipeline's job objects change which of the freed blocks the capture
+# phase is handed, and that is what decides whether the destroyed-VAO address is
+# reached at all - on the ablated (pre-fix) tree async=1 reproduced 3 runs out of
+# 3 where the ambient default reproduced 2 of 3. Pinning it here means the
+# high-signal configuration runs whatever the shipped default becomes, instead of
+# the suite quietly weakening the day that default flips. It must be process-wide
+# (the ENVIRONMENT property), not an in-process scope: the compile pool and its
+# threads are stood up at initialization, and their allocations are half the
+# point. DirectVulkan only - the memo this pins is DirectVulkan's.
+gtest_discover_tests(MobileGLIntegrationTest
+    TEST_PREFIX "DirectVulkan.AsyncCompile."
+    TEST_FILTER "XfbAfterClipDistanceScenario.*"
+    DISCOVERY_TIMEOUT 30
+    PROPERTIES
+        LABELS integration-gpu
+        TIMEOUT ${MGL_ITEST_TIMEOUT}
+        ENVIRONMENT "${MGL_ITEST_VULKAN_ASYNC_ENVIRONMENT}"
+)

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.cpp

@@ -0,0 +1,587 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Harness/HeadlessGL.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "HeadlessGL.h"
+
+#include <algorithm>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ostream>
+#include <sstream>
+
+// MobileGL's own headers, in the order MobileGL/Includes.h uses them: GL/gl.h
+// first, then glcorearb.h for the 3.x+ entry points. This binary links
+// MobileGL_s, so every gl*/egl* below binds to MobileGL's implementation, not
+// to a system loader.
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#include <EGL/egl.h>
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+// The pre-flight below runs the whole EGL bring-up in a forked child, which is
+// the only construction that is actually predictive here: MobileGL ABORTS
+// (MOBILEGL_ASSERT -> SIGTRAP) rather than returning an error on an unusable
+// platform, so nothing the parent can call in-process is allowed to be wrong.
+#if !defined(_WIN32) && !defined(__APPLE__) && __has_include(<sys/wait.h>)
+#define MGITEST_HAVE_FORK_PREFLIGHT 1
+#include <csignal>
+#include <ctime>
+#include <sys/resource.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#else
+#define MGITEST_HAVE_FORK_PREFLIGHT 0
+#endif
+
+namespace MGITest {
+
+    namespace {
+        // Small enough that a readback is cheap, big enough that "top third" and
+        // "bottom third" are unambiguous. Non-square on purpose: a transposing
+        // bug cannot hide behind a square.
+        constexpr int kSurfaceWidth = 128;
+        constexpr int kSurfaceHeight = 96;
+
+        std::string EnvOr(const char* name, const char* fallback) {
+            const char* value = std::getenv(name);
+            return (value != nullptr && value[0] != '\0') ? std::string(value) : std::string(fallback);
+        }
+
+        // A skip reason is only useful if it says which call failed AND why, so
+        // every bring-up step reports the EGL error it left behind.
+        std::string WithEglError(const char* what) {
+            std::ostringstream out;
+            out << what << " (eglGetError=0x" << std::hex << eglGetError() << ")";
+            return out.str();
+        }
+
+        // The EGL objects one bring-up produces.
+        struct EglBringUp {
+            void* display = nullptr;
+            void* surface = nullptr;
+            void* context = nullptr;
+            std::string renderer;
+        };
+
+        // THE bring-up, in one function so the pre-flight child and the parent run
+        // literally the same sequence - a pre-flight that tests something narrower
+        // than what the parent will do is exactly the kind of "predictive" check
+        // that is not.
+        //
+        // Returns 0 on success, or the 1-based index of the step that failed, and
+        // fills outReason either way.
+        int RunEglBringUp(EglBringUp& out, std::string& outReason) {
+            EGLDisplay display = eglGetDisplay(EGL_DEFAULT_DISPLAY);
+            if (display == EGL_NO_DISPLAY) {
+                outReason = WithEglError("eglGetDisplay(EGL_DEFAULT_DISPLAY) returned EGL_NO_DISPLAY");
+                return 1;
+            }
+            EGLint major = 0, minor = 0;
+            if (eglInitialize(display, &major, &minor) != EGL_TRUE) {
+                outReason = WithEglError("eglInitialize failed: no usable display/driver on this machine");
+                return 2;
+            }
+            if (eglBindAPI(EGL_OPENGL_API) != EGL_TRUE) {
+                outReason = WithEglError("eglBindAPI(EGL_OPENGL_API) failed");
+                return 3;
+            }
+
+            const EGLint configAttribs[] = {EGL_SURFACE_TYPE,
+                                            EGL_PBUFFER_BIT,
+                                            EGL_RED_SIZE,
+                                            8,
+                                            EGL_GREEN_SIZE,
+                                            8,
+                                            EGL_BLUE_SIZE,
+                                            8,
+                                            EGL_ALPHA_SIZE,
+                                            8,
+                                            EGL_DEPTH_SIZE,
+                                            24,
+                                            EGL_RENDERABLE_TYPE,
+                                            EGL_OPENGL_BIT,
+                                            EGL_NONE};
+            EGLConfig config = nullptr;
+            EGLint configCount = 0;
+            if (eglChooseConfig(display, configAttribs, &config, 1, &configCount) != EGL_TRUE || configCount < 1) {
+                outReason = WithEglError("eglChooseConfig found no pbuffer-capable RGBA8/D24 config");
+                return 4;
+            }
+
+            const EGLint contextAttribs[] = {EGL_CONTEXT_MAJOR_VERSION, 3, EGL_CONTEXT_MINOR_VERSION, 3, EGL_NONE};
+            EGLContext context = eglCreateContext(display, config, EGL_NO_CONTEXT, contextAttribs);
+            if (context == EGL_NO_CONTEXT) {
+                context = eglCreateContext(display, config, EGL_NO_CONTEXT, nullptr);
+            }
+            if (context == EGL_NO_CONTEXT) {
+                outReason = WithEglError("eglCreateContext failed: no desktop-GL context available");
+                return 5;
+            }
+
+            const EGLint pbufferAttribs[] = {EGL_WIDTH, kSurfaceWidth, EGL_HEIGHT, kSurfaceHeight, EGL_NONE};
+            EGLSurface surface = eglCreatePbufferSurface(display, config, pbufferAttribs);
+            if (surface == EGL_NO_SURFACE) {
+                outReason = WithEglError("eglCreatePbufferSurface failed");
+                return 6;
+            }
+            // The step that brings the whole backend up (DirectVulkan creates its
+            // instance, device and surface in here) and therefore the step that
+            // aborts instead of returning an error on an unusable platform.
+            if (eglMakeCurrent(display, surface, surface, context) != EGL_TRUE) {
+                outReason = WithEglError("eglMakeCurrent failed");
+                return 7;
+            }
+
+            const GLubyte* renderer = glGetString(GL_RENDERER);
+            if (renderer == nullptr) {
+                outReason = "glGetString(GL_RENDERER) returned null after eglMakeCurrent";
+                return 8;
+            }
+
+            out.display = display;
+            out.surface = surface;
+            out.context = context;
+            out.renderer = reinterpret_cast<const char*>(renderer);
+            outReason.clear();
+            return 0;
+        }
+
+        // Platform pre-flight, and the reason this module can claim to skip
+        // cleanly rather than merely hope to.
+        //
+        // MobileGL does not return errors when the platform is unusable - it
+        // ABORTS. MOBILEGL_ASSERT raises SIGTRAP, and the DirectVulkan bring-up
+        // asserts its way through instance, physical-device and surface creation
+        // inside eglMakeCurrent. So there is no in-process question the harness
+        // can ask that is guaranteed to be survivable, and the old form (dlopen
+        // the Vulkan loader, count physical devices, look for
+        // VK_EXT_headless_surface) was a guess at the abort conditions rather
+        // than a test of them: it named three of the ways bring-up can die and
+        // was silent about every other one, including every DirectGLES one.
+        //
+        // What is actually predictive is to run the bring-up itself somewhere a
+        // SIGTRAP is a datum instead of a crash. fork() gives exactly that: the
+        // child performs the identical sequence and _exit(0)s on success, and
+        // ANY non-zero exit or ANY signal in the parent's waitpid() means "this
+        // platform is unusable" - whatever the reason, including reasons nobody
+        // has thought of. Only then does the parent do the real bring-up.
+        //
+        // Returns an empty string when the platform survived a full bring-up.
+        std::string PreflightBringUp() {
+#if !MGITEST_HAVE_FORK_PREFLIGHT
+            // No fork(): let the in-process bring-up speak for itself, which is
+            // what this module did before. Windows/macOS are not CI targets for
+            // the headless scenarios.
+            return {};
+#else
+            int channel[2] = {-1, -1};
+            if (pipe(channel) != 0) {
+                return {}; // cannot pre-flight; fall through to the in-process attempt
+            }
+            // The child inherits our stdio buffers; flush so nothing is printed twice.
+            std::fflush(nullptr);
+            const pid_t child = fork();
+            if (child < 0) {
+                close(channel[0]);
+                close(channel[1]);
+                return {};
+            }
+            if (child == 0) {
+                close(channel[0]);
+                // The child is EXPECTED to die on a signal on an unusable
+                // platform; that is the measurement. Do not let each such
+                // measurement drop a core file next to the test binary.
+                const rlimit noCore{0, 0};
+                setrlimit(RLIMIT_CORE, &noCore);
+                std::fprintf(stderr, "[itest] pre-flight child: attempting a full EGL bring-up\n");
+                EglBringUp local;
+                std::string reason;
+                const int step = RunEglBringUp(local, reason);
+                if (!reason.empty()) {
+                    const std::size_t bytes = std::min<std::size_t>(reason.size(), 480);
+                    const ssize_t written = write(channel[1], reason.data(), bytes);
+                    (void)written;
+                }
+                close(channel[1]);
+                // _exit, never exit(): every atexit handler and static destructor
+                // in this address space belongs to the parent's copy of the world,
+                // and the child is holding a live context it must not tear down.
+                _exit(step);
+            }
+
+            close(channel[1]);
+            // Reap first, read after: the message is bounded well below the pipe
+            // buffer so the child can never block writing it, and polling the exit
+            // status is what lets a wedged child be killed instead of hanging the
+            // parent on a read that will never return.
+            constexpr int kPreflightTimeoutMs = 30000;
+            int status = 0;
+            int waitedMs = 0;
+            for (;;) {
+                const pid_t reaped = waitpid(child, &status, WNOHANG);
+                if (reaped == child) break;
+                if (reaped < 0) {
+                    close(channel[0]);
+                    return "waitpid on the EGL bring-up pre-flight child failed";
+                }
+                if (waitedMs >= kPreflightTimeoutMs) {
+                    kill(child, SIGKILL);
+                    (void)waitpid(child, &status, 0);
+                    close(channel[0]);
+                    std::ostringstream out;
+                    out << "the EGL bring-up wedged: a forked pre-flight child made no progress in "
+                        << kPreflightTimeoutMs / 1000 << "s and was killed";
+                    return out.str();
+                }
+                timespec nap{0, 10 * 1000 * 1000};
+                nanosleep(&nap, nullptr);
+                waitedMs += 10;
+            }
+
+            std::string childSays;
+            char buffer[512];
+            for (;;) {
+                const ssize_t got = read(channel[0], buffer, sizeof(buffer));
+                if (got <= 0) break;
+                childSays.append(buffer, static_cast<std::size_t>(got));
+            }
+            close(channel[0]);
+
+            if (WIFSIGNALED(status)) {
+                const int signalNumber = WTERMSIG(status);
+                const char* signalName = strsignal(signalNumber);
+                std::ostringstream out;
+                out << "the EGL bring-up ABORTS on this platform: a forked pre-flight child died on signal "
+                    << signalNumber << " (" << (signalName != nullptr ? signalName : "?") << ")";
+                if (!childSays.empty()) out << " after: " << childSays;
+                out << ". MobileGL asserts rather than returning an error here, so the scenarios would "
+                       "have taken the whole test binary down with them";
+                return out.str();
+            }
+            if (!WIFEXITED(status)) {
+                return "the EGL bring-up pre-flight child neither exited nor was signalled";
+            }
+            const int exitStatus = WEXITSTATUS(status);
+            if (exitStatus != 0) {
+                std::ostringstream out;
+                out << (childSays.empty() ? "the EGL bring-up failed" : childSays)
+                    << " (forked pre-flight child exit status " << exitStatus << ")";
+                return out.str();
+            }
+            return {};
+#endif
+        }
+    } // namespace
+
+    bool RequireGpu() {
+        const char* value = std::getenv("MOBILEGL_ITEST_REQUIRE_GPU");
+        return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0;
+    }
+
+    std::ostream& operator<<(std::ostream& os, const Rgba8& c) {
+        os << "rgba(" << int(c.r) << "," << int(c.g) << "," << int(c.b) << "," << int(c.a) << ")";
+        return os;
+    }
+
+    Rgba8 Image::At(int x, int y) const {
+        if (x < 0 || y < 0 || x >= m_width || y >= m_height) {
+            return Rgba8{};
+        }
+        const std::size_t index = (static_cast<std::size_t>(y) * m_width + x) * 4;
+        return Rgba8{m_pixels[index], m_pixels[index + 1], m_pixels[index + 2], m_pixels[index + 3]};
+    }
+
+    const char* Image::ColorName(int x, int y) const {
+        const Rgba8 c = At(x, y);
+        const bool r = c.r > 160, g = c.g > 160, b = c.b > 160;
+        const bool nr = c.r < 96, ng = c.g < 96, nb = c.b < 96;
+        if (nr && ng && nb) return "black";
+        if (r && g && b) return "white";
+        if (r && ng && nb) return "red";
+        if (nr && g && nb) return "green";
+        if (nr && ng && b) return "blue";
+        if (r && g && nb) return "yellow";
+        return "other";
+    }
+
+    std::size_t Image::ByteDiffCount(const Image& other) const {
+        if (m_width != other.m_width || m_height != other.m_height) {
+            return std::max(m_pixels.size(), other.m_pixels.size());
+        }
+        std::size_t differing = 0;
+        for (std::size_t i = 0; i < m_pixels.size(); ++i) {
+            if (m_pixels[i] != other.m_pixels[i]) ++differing;
+        }
+        return differing;
+    }
+
+    std::string Image::QuadrantSignature() const {
+        if (m_width < 2 || m_height < 2) return "<empty>";
+        // Quadrant CENTRES, so a one-pixel rounding difference at a quadrant edge
+        // never decides the answer. Order is fixed and load-bearing: bottom-left,
+        // bottom-right, top-left, top-right.
+        const int leftX = m_width / 4;
+        const int rightX = m_width * 3 / 4;
+        const int bottomY = m_height / 4;
+        const int topY = m_height * 3 / 4;
+        std::ostringstream out;
+        out << ColorName(leftX, bottomY) << "," << ColorName(rightX, bottomY) << "," << ColorName(leftX, topY) << ","
+            << ColorName(rightX, topY);
+        return out.str();
+    }
+
+    RegionScan ScanRegion(const Image& image, int x0, int x1, int y0, int y1, const char* expectedColor) {
+        RegionScan scan;
+        x0 = std::max(x0, 0);
+        y0 = std::max(y0, 0);
+        x1 = std::min(x1, image.Width() - 1);
+        y1 = std::min(y1, image.Height() - 1);
+        for (int y = y0; y <= y1; ++y) {
+            for (int x = x0; x <= x1; ++x) {
+                ++scan.total;
+                const char* name = image.ColorName(x, y);
+                if (std::strcmp(name, expectedColor) == 0) continue;
+                ++scan.offenders;
+                if (scan.firstX < 0) {
+                    scan.firstX = x;
+                    scan.firstY = y;
+                    scan.firstColor = image.At(x, y);
+                    scan.firstColorName = name;
+                }
+            }
+        }
+        return scan;
+    }
+
+    ::testing::AssertionResult RegionIsMostly(const Image& image, int x0, int x1, int y0, int y1,
+                                              const char* expectedColor, double tolerance,
+                                              const std::string& when) {
+        const RegionScan scan = ScanRegion(image, x0, x1, y0, y1, expectedColor);
+        if (scan.total == 0) {
+            return ::testing::AssertionFailure()
+                   << when << ": region x[" << x0 << "," << x1 << "] y[" << y0 << "," << y1
+                   << "] is empty against a " << image.Width() << "x" << image.Height() << " readback";
+        }
+        const double offendingFraction = static_cast<double>(scan.offenders) / scan.total;
+        if (offendingFraction <= tolerance) {
+            return ::testing::AssertionSuccess();
+        }
+        return ::testing::AssertionFailure()
+               << when << ": region x[" << x0 << "," << x1 << "] y[" << y0 << "," << y1 << "] should be all "
+               << expectedColor << ", but " << scan.offenders << " of " << scan.total << " pixels ("
+               << static_cast<int>(offendingFraction * 100.0 + 0.5) << "%) are not; first offender at (" << scan.firstX
+               << "," << scan.firstY << ") is " << scan.firstColorName << " " << scan.firstColor;
+    }
+
+    HeadlessGL& HeadlessGL::Get() {
+        static HeadlessGL instance;
+        return instance;
+    }
+
+    HeadlessGL::HeadlessGL() {
+        m_backendName = EnvOr("MOBILEGL_BACKEND_TYPE", "<unset>");
+        m_usable = BringUp();
+    }
+
+    bool HeadlessGL::BringUp() {
+        // Ask a disposable copy of this process first. Only if it survived does
+        // the real one try - see PreflightBringUp for why nothing weaker is
+        // predictive against a stack that aborts instead of returning errors.
+        const std::string preflightProblem = PreflightBringUp();
+        if (!preflightProblem.empty()) {
+            m_skipReason = preflightProblem;
+            return false;
+        }
+
+        // Same shape as DriverBench's boot_egl(), minus the dlopen: the provider
+        // is this binary. A pbuffer needs no window system, but MobileGL's own
+        // loader still has to reach a real driver underneath - and the child
+        // above just proved it can.
+        EglBringUp brought;
+        std::string reason;
+        if (RunEglBringUp(brought, reason) != 0) {
+            // The pre-flight passed and the parent's identical attempt did not.
+            // That is a real result, not a machine without a GPU, so say so: it
+            // means something is different between the two attempts (a leaked
+            // exclusive device, an environment the child did not have).
+            m_skipReason = reason + " - although an identical bring-up in a forked pre-flight child succeeded";
+            return false;
+        }
+
+        m_display = brought.display;
+        m_surface = brought.surface;
+        m_context = brought.context;
+        m_width = kSurfaceWidth;
+        m_height = kSurfaceHeight;
+        m_renderer = std::move(brought.renderer);
+        return true;
+    }
+
+    void HeadlessGL::EndFrame() {
+        if (!m_usable) return;
+        eglSwapBuffers(static_cast<EGLDisplay>(m_display), static_cast<EGLSurface>(m_surface));
+        ++m_frameIndex;
+    }
+
+    void HeadlessGL::ShutDown() {
+        if (!m_usable) return;
+        EGLDisplay display = static_cast<EGLDisplay>(m_display);
+        eglMakeCurrent(display, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
+        if (m_context != nullptr) eglDestroyContext(display, static_cast<EGLContext>(m_context));
+        if (m_surface != nullptr) eglDestroySurface(display, static_cast<EGLSurface>(m_surface));
+        eglTerminate(display);
+        m_context = nullptr;
+        m_surface = nullptr;
+        m_display = nullptr;
+        m_usable = false;
+        m_skipReason = "the headless context has already been torn down";
+    }
+
+    // ---- scenario vocabulary ------------------------------------------------
+
+    namespace {
+        unsigned int CompileStage(GLenum stage, const char* source, std::string* outError) {
+            const GLuint shader = glCreateShader(stage);
+            glShaderSource(shader, 1, &source, nullptr);
+            glCompileShader(shader);
+            GLint compiled = 0;
+            glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+            if (compiled == GL_FALSE) {
+                char log[2048] = {};
+                GLsizei length = 0;
+                glGetShaderInfoLog(shader, sizeof(log) - 1, &length, log);
+                if (outError != nullptr) {
+                    *outError = std::string(stage == GL_VERTEX_SHADER ? "vertex" : "fragment") +
+                                " shader failed to compile: " + log;
+                }
+                glDeleteShader(shader);
+                return 0;
+            }
+            return shader;
+        }
+    } // namespace
+
+    unsigned int CompileProgram(const char* vertexSource, const char* fragmentSource, std::string* outError) {
+        const GLuint vs = CompileStage(GL_VERTEX_SHADER, vertexSource, outError);
+        if (vs == 0) return 0;
+        const GLuint fs = CompileStage(GL_FRAGMENT_SHADER, fragmentSource, outError);
+        if (fs == 0) {
+            glDeleteShader(vs);
+            return 0;
+        }
+        const GLuint program = glCreateProgram();
+        glAttachShader(program, vs);
+        glAttachShader(program, fs);
+        // Pinned rather than queried so the scenarios can set up a VAO without a
+        // round trip, and so a driver that reorders attributes cannot change what
+        // the test means.
+        glBindAttribLocation(program, 0, "aPos");
+        glBindAttribLocation(program, 1, "aColor");
+        glLinkProgram(program);
+        glDeleteShader(vs);
+        glDeleteShader(fs);
+        GLint linked = 0;
+        glGetProgramiv(program, GL_LINK_STATUS, &linked);
+        if (linked == GL_FALSE) {
+            char log[2048] = {};
+            GLsizei length = 0;
+            glGetProgramInfoLog(program, sizeof(log) - 1, &length, log);
+            if (outError != nullptr) *outError = std::string("program failed to link: ") + log;
+            glDeleteProgram(program);
+            return 0;
+        }
+        return program;
+    }
+
+    ColorFbo MakeColorFbo(int width, int height) {
+        ColorFbo target;
+        target.width = width;
+        target.height = height;
+        glGenTextures(1, &target.texture);
+        glBindTexture(GL_TEXTURE_2D, target.texture);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+        glBindTexture(GL_TEXTURE_2D, 0);
+
+        glGenFramebuffers(1, &target.fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, target.texture, 0);
+        const GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        if (status != GL_FRAMEBUFFER_COMPLETE) {
+            DestroyColorFbo(target);
+        }
+        return target;
+    }
+
+    void DestroyColorFbo(ColorFbo& target) {
+        if (target.fbo != 0) glDeleteFramebuffers(1, &target.fbo);
+        if (target.texture != 0) glDeleteTextures(1, &target.texture);
+        target.fbo = 0;
+        target.texture = 0;
+    }
+
+    void BindDefaultFramebuffer() {
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        glViewport(0, 0, HeadlessGL::Get().Width(), HeadlessGL::Get().Height());
+    }
+
+    void BindFbo(const ColorFbo& target) {
+        glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+        glViewport(0, 0, target.width, target.height);
+    }
+
+    void ClearTo(float r, float g, float b, float a) {
+        glClearColor(r, g, b, a);
+        glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
+    }
+
+    Image ReadPixels(int width, int height) {
+        Image image(width, height);
+        glPixelStorei(GL_PACK_ALIGNMENT, 1);
+        glReadPixels(0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, image.Data());
+        return image;
+    }
+
+    unsigned int FirstGLError() {
+        const GLenum first = glGetError();
+        if (first == GL_NO_ERROR) return GL_NO_ERROR;
+        // Drain, bounded: a broken stack must not turn an error check into a hang.
+        for (int i = 0; i < 64 && glGetError() != GL_NO_ERROR; ++i) {}
+        return first;
+    }
+
+    const char* GLErrorName(unsigned int error) {
+        switch (error) {
+        case GL_NO_ERROR:
+            return "GL_NO_ERROR";
+        case GL_INVALID_ENUM:
+            return "GL_INVALID_ENUM";
+        case GL_INVALID_VALUE:
+            return "GL_INVALID_VALUE";
+        case GL_INVALID_OPERATION:
+            return "GL_INVALID_OPERATION";
+        case GL_OUT_OF_MEMORY:
+            return "GL_OUT_OF_MEMORY";
+        case GL_INVALID_FRAMEBUFFER_OPERATION:
+            return "GL_INVALID_FRAMEBUFFER_OPERATION";
+        default:
+            return "GL_<unknown>";
+        }
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.h

@@ -0,0 +1,218 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Harness/HeadlessGL.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// A headless GL context and the small vocabulary the scenarios are written in.
+//
+// The scenarios in this module are end-to-end: they drive MobileGL's own GL and
+// EGL entry points (this binary links MobileGL_s, so gl*/egl* resolve straight
+// into the implementation) and assert on glReadPixels output. Nothing here
+// inspects backend state - both bugs this module pins were invisible to
+// state-level assertions and visible only in pixels.
+//
+// Headless by construction, following MG_Benchmark/Driver/DriverBench.c: an EGL
+// context on a PBUFFER surface. No window, no window manager, no human. Unlike
+// DriverBench the scenarios do draw to the DEFAULT framebuffer (that is where
+// the Y-flip lives) and do call eglSwapBuffers (that is the frame boundary the
+// cross-frame scenarios need to be real).
+//
+// One process is one backend: MOBILEGL_BACKEND_TYPE is latched at
+// initialization, so the CMake wiring runs this binary once per backend rather
+// than trying to switch in-process.
+
+#pragma once
+
+#include <gtest/gtest.h>
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+namespace MGITest {
+
+    // True when MOBILEGL_ITEST_REQUIRE_GPU is set in the environment: the runner
+    // is asserting that this machine HAS a usable GPU, so "no GPU" stops being a
+    // clean skip and becomes a failure. Without it the integration-gpu label is
+    // unfalsifiable - a CI job that ran nothing reports exactly the same green as
+    // a job that ran everything.
+    bool RequireGpu();
+
+    struct Rgba8 {
+        std::uint8_t r = 0, g = 0, b = 0, a = 0;
+
+        bool operator==(const Rgba8& other) const {
+            return r == other.r && g == other.g && b == other.b && a == other.a;
+        }
+        bool operator!=(const Rgba8& other) const { return !(*this == other); }
+    };
+
+    // Prints as "rgba(255,0,0,255)" so a gtest failure names the colour it saw.
+    std::ostream& operator<<(std::ostream& os, const Rgba8& c);
+
+    // An RGBA8 readback. Row 0 is the BOTTOM row: that is GL's convention for
+    // glReadPixels and it is what "correctly oriented" means everywhere below.
+    class Image {
+    public:
+        Image() = default;
+        Image(int width, int height)
+            : m_width(width), m_height(height), m_pixels(static_cast<std::size_t>(width) * height * 4, 0) {}
+
+        int Width() const { return m_width; }
+        int Height() const { return m_height; }
+        bool Empty() const { return m_pixels.empty(); }
+        std::uint8_t* Data() { return m_pixels.data(); }
+        const std::uint8_t* Data() const { return m_pixels.data(); }
+
+        Rgba8 At(int x, int y) const;
+        // Nearest of {black, red, green, blue, white, other} - the scenarios only
+        // ever draw those, so this turns a pixel into something readable.
+        const char* ColorName(int x, int y) const;
+
+        bool operator==(const Image& other) const {
+            return m_width == other.m_width && m_height == other.m_height && m_pixels == other.m_pixels;
+        }
+
+        // Count of differing bytes, for a failure message that says how wrong.
+        std::size_t ByteDiffCount(const Image& other) const;
+
+        // The four quadrant centres, in the fixed order
+        // bottom-left, bottom-right, top-left, top-right.
+        //
+        // This replaces the old VerticalSignature(bandCount), which read three
+        // full-width horizontal stripes down the centre line and was therefore
+        // blind to an X flip, to a transpose, and to a 180 rotation composed with
+        // a Y flip - all of those left the stripe order alone. Four quadrant
+        // colours are asymmetric in BOTH axes, so each of the eight square
+        // symmetries produces a different string (see OrientationScenario, which
+        // spells all eight out).
+        std::string QuadrantSignature() const;
+
+    private:
+        int m_width = 0;
+        int m_height = 0;
+        std::vector<std::uint8_t> m_pixels;
+    };
+
+    // The process-wide headless context. Brought up lazily on the first Get() so
+    // that `--gtest_list_tests` (which CMake runs at build time to discover the
+    // cases) never touches a GPU.
+    class HeadlessGL {
+    public:
+        static HeadlessGL& Get();
+
+        // False on a machine with no usable GPU/display/ICD. SkipReason() then
+        // says which step failed; every fixture turns that into GTEST_SKIP().
+        bool Usable() const { return m_usable; }
+        const std::string& SkipReason() const { return m_skipReason; }
+
+        // Backend actually in use, as reported by MOBILEGL_BACKEND_TYPE.
+        const std::string& BackendName() const { return m_backendName; }
+        const std::string& RendererString() const { return m_renderer; }
+
+        int Width() const { return m_width; }
+        int Height() const { return m_height; }
+
+        // THE frame boundary. eglSwapBuffers is what retires a frame in the
+        // renderer, and the cross-frame scenarios are meaningless without it.
+        void EndFrame();
+
+        // Frames completed so far, for failure messages.
+        int FrameIndex() const { return m_frameIndex; }
+
+        // Releases the context and surface and terminates the display. Called
+        // once, after the last scenario: MobileGL frees its backend objects
+        // through eglTerminate, and letting a process simply exit on top of a
+        // live context leaves those objects to be torn down from a static
+        // destructor with no driver left underneath.
+        void ShutDown();
+
+    private:
+        HeadlessGL();
+        HeadlessGL(const HeadlessGL&) = delete;
+        HeadlessGL& operator=(const HeadlessGL&) = delete;
+
+        bool BringUp();
+
+        bool m_usable = false;
+        std::string m_skipReason;
+        std::string m_backendName;
+        std::string m_renderer;
+        int m_width = 0;
+        int m_height = 0;
+        int m_frameIndex = 0;
+        void* m_display = nullptr;
+        void* m_surface = nullptr;
+        void* m_context = nullptr;
+    };
+
+    // ---- the scenario vocabulary -------------------------------------------
+    // Deliberately tiny. A scenario should read like a story; anything that
+    // needs a comment about GL mechanics belongs here instead.
+
+    // Compiles and links vs+fs, pinning attribute 0 to "aPos" and 1 to "aColor".
+    // Returns 0 and fills outError on failure.
+    unsigned int CompileProgram(const char* vertexSource, const char* fragmentSource, std::string* outError);
+
+    struct ColorFbo {
+        unsigned int fbo = 0;
+        unsigned int texture = 0;
+        int width = 0;
+        int height = 0;
+    };
+
+    // A complete RGBA8 render target. Returns fbo==0 on failure.
+    ColorFbo MakeColorFbo(int width, int height);
+    void DestroyColorFbo(ColorFbo& target);
+
+    // Binds a target and sets the viewport to match. Passing fbo 0 means the
+    // default (presentable) framebuffer.
+    void BindDefaultFramebuffer();
+    void BindFbo(const ColorFbo& target);
+
+    void ClearTo(float r, float g, float b, float a);
+
+    // Reads back the whole currently bound READ framebuffer. width/height must
+    // be the target's full size - DirectVulkan's default-framebuffer readback
+    // only re-orients a full-extent read.
+    Image ReadPixels(int width, int height);
+
+    // Drains any GL error queue and returns the first error, or 0.
+    unsigned int FirstGLError();
+    const char* GLErrorName(unsigned int error);
+
+    // ---- whole-region readback predicates ----------------------------------
+    // The scenarios used to assert on two or three individual pixels, which is
+    // provably too weak: a draw in which 3 of a quad's 4 vertices carry stale
+    // data still paints the sampled centre the expected colour (that exact case
+    // is a standing negative-control test - see CrossFrameBufferScenario). The
+    // readback is already fully in memory, so counting every pixel in a region
+    // costs nothing and turns "the middle looks right" into "all of it is right".
+
+    // Everything a caller needs to say what was wrong and where.
+    struct RegionScan {
+        int total = 0;     // pixels examined
+        int offenders = 0; // pixels whose ColorName() != expected
+        int firstX = -1;   // first offender in bottom-to-top, left-to-right order
+        int firstY = -1;
+        Rgba8 firstColor{};
+        std::string firstColorName;
+    };
+
+    // Inclusive pixel bounds, clamped to the image. Row 0 is the bottom row.
+    RegionScan ScanRegion(const Image& image, int x0, int x1, int y0, int y1, const char* expectedColor);
+
+    // gtest predicate wrapper: EXPECT_TRUE(RegionIsMostly(...)) reports the
+    // offender count, the offender fraction and the FIRST offending pixel's
+    // coordinates and colour. `tolerance` is the fraction of the region allowed
+    // to disagree; pass 0.0 to demand every pixel (which is what the scenarios
+    // do - they inset their regions away from primitive edges so exactness is
+    // achievable).
+    ::testing::AssertionResult RegionIsMostly(const Image& image, int x0, int x1, int y0, int y1,
+                                              const char* expectedColor, double tolerance,
+                                              const std::string& when);
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Harness/ScenarioFixture.h

@@ -0,0 +1,84 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Harness/ScenarioFixture.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// The base fixture every scenario derives from. Its only jobs are to bring the
+// headless context up once per process and to decide what "this machine has no
+// usable GPU" means.
+//
+// By default it means a clean GTEST_SKIP() - never a failure, never a hang -
+// because a developer box or a container without a GPU should not fail a run it
+// was never able to perform. But a skip is indistinguishable from a pass in
+// every CI summary, so the `integration-gpu` label on its own is unfalsifiable:
+// a runner whose driver pinning silently broke reports the same green as one
+// that rendered every frame. MOBILEGL_ITEST_REQUIRE_GPU is the caller saying
+// "this machine HAS a GPU and I am relying on these scenarios actually running";
+// with it set, an unusable harness is a FAILURE carrying the pre-flight's reason.
+
+#pragma once
+
+#include <gtest/gtest.h>
+
+#include "HeadlessGL.h"
+
+namespace MGITest {
+
+    class ScenarioTest : public ::testing::Test {
+    protected:
+        void SetUp() override {
+            m_ready = false;
+            HeadlessGL& gl = HeadlessGL::Get();
+            if (!gl.Usable()) {
+                if (RequireGpu()) {
+                    // FAIL() is a FATAL failure but does NOT mark the test skipped,
+                    // so a derived SetUp that guards on IsSkipped() alone would run
+                    // straight into GL calls with no current context and SIGSEGV -
+                    // that exact crash shipped from the first version of this guard.
+                    // Derived fixtures must gate on Ready() (below), which is false
+                    // on BOTH the skip path and this failure path.
+                    FAIL() << "MOBILEGL_ITEST_REQUIRE_GPU is set, so an unusable harness is a failure, not a skip. "
+                           << "Backend " << gl.BackendName() << " could not be brought up: " << gl.SkipReason();
+                }
+                GTEST_SKIP() << "no usable GPU/display/ICD for backend " << gl.BackendName() << ": " << gl.SkipReason();
+            }
+            if (RequireGpu() && LooksLikeSoftwareRasterizer(gl.RendererString())) {
+                // "Ran on llvmpipe" must not be able to pass as "ran on the GPU":
+                // a misconfigured vendor pin silently lands on the software
+                // rasterizer, and REQUIRE_GPU exists precisely to make that loud.
+                FAIL() << "MOBILEGL_ITEST_REQUIRE_GPU is set but the context landed on a software rasterizer: "
+                       << gl.RendererString();
+            }
+            // A scenario starts from a clean slate but shares the context (and so
+            // the renderer's memos) with every other scenario in this process -
+            // which is exactly the situation both shipped bugs needed.
+            RecordProperty("backend", gl.BackendName());
+            RecordProperty("renderer", gl.RendererString());
+            m_ready = true;
+        }
+
+        // The ONLY gate a derived SetUp/TearDown may use: `if (!Ready()) return;`.
+        // True only when the base SetUp brought the context up and neither skipped
+        // nor failed. IsSkipped() alone is WRONG here (see the comment at FAIL()).
+        bool Ready() const { return m_ready; }
+
+        static HeadlessGL& Gl() { return HeadlessGL::Get(); }
+
+    private:
+        static bool LooksLikeSoftwareRasterizer(const std::string& renderer) {
+            static const char* kNames[] = {"llvmpipe", "lavapipe", "softpipe", "SwiftShader", "swrast"};
+            for (const char* name : kNames) {
+                if (renderer.find(name) != std::string::npos) {
+                    return true;
+                }
+            }
+            return false;
+        }
+
+        bool m_ready = false;
+    };
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Main.cpp

@@ -0,0 +1,53 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Main.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Entry point for the headless GPU integration scenarios.
+//
+// The banner lives in a gtest Environment rather than in main() on purpose:
+// Environment::SetUp does not run for `--gtest_list_tests`, which is what CMake
+// invokes at build time to discover the cases. Discovery therefore never brings
+// up EGL, never needs a GPU and cannot hang.
+
+#include <gtest/gtest.h>
+#include <cstdio>
+
+#include "Harness/HeadlessGL.h"
+
+namespace {
+
+    class HarnessBanner : public ::testing::Environment {
+    public:
+        void SetUp() override {
+            const MGITest::HeadlessGL& gl = MGITest::HeadlessGL::Get();
+            std::fprintf(stderr, "MobileGL integration scenarios: backend=%s\n", gl.BackendName().c_str());
+            if (gl.Usable()) {
+                std::fprintf(stderr, "  renderer: %s\n  surface:  %dx%d pbuffer (headless)\n",
+                             gl.RendererString().c_str(), gl.Width(), gl.Height());
+            } else if (MGITest::RequireGpu()) {
+                std::fprintf(stderr,
+                             "  FAILING every scenario (MOBILEGL_ITEST_REQUIRE_GPU is set): %s\n",
+                             gl.SkipReason().c_str());
+            } else {
+                std::fprintf(stderr,
+                             "  SKIPPING every scenario: %s\n"
+                             "  (set MOBILEGL_ITEST_REQUIRE_GPU=1 to make this a failure instead - a run that\n"
+                             "   skipped everything is otherwise indistinguishable from one that passed)\n",
+                             gl.SkipReason().c_str());
+            }
+        }
+
+        void TearDown() override { MGITest::HeadlessGL::Get().ShutDown(); }
+    };
+
+} // namespace
+
+int main(int argc, char** argv) {
+    ::testing::InitGoogleTest(&argc, argv);
+    ::testing::AddGlobalTestEnvironment(new HarnessBanner());
+    return RUN_ALL_TESTS();
+}

MobileGL/MG_IntegrationTest/Scenarios/AsyncCompileScenario.cpp

@@ -0,0 +1,467 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/AsyncCompileScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario E - asynchronous shader compilation and GL_KHR_parallel_shader_compile
+// on a REAL driver.
+//
+// WHY THIS EXISTS ALONGSIDE THE UNIT SUITES. MG_Test/Program's async suites already
+// drive the same GL entry points, but they stop at the frontend: nothing there ever
+// reaches a driver, so nothing there can catch the failure this scenario is built for
+// - artifacts produced on a worker thread that the BACKEND then rejects, mis-binds or
+// renders differently from the ones the GL thread produced. The frontend cannot tell
+// the two apart; a pixel can.
+//
+// The five things it pins, in order:
+//
+//   (a) 64 heavy compiles are enqueued and polled through GL_COMPLETION_STATUS_KHR.
+//       At least one must be observed GL_FALSE - i.e. the query really answers while
+//       work is outstanding rather than silently joining. Skipped, never failed, when
+//       the machine drained the whole batch before the first poll: a fast box must not
+//       be able to turn this into a red.
+//   (b) Forcing the join afterwards produces the right answer for every one of them:
+//       GL_COMPILE_STATUS true, an empty info log, and a program that links.
+//   (c) The extension string matches the configuration. This is the half a recorded
+//       trace can never cover - Iris and Sodium change their submission schedule the
+//       moment they see the string - so it is asserted against a real backend's real
+//       GL_EXTENSIONS, through both glGetString and glGetStringi.
+//   (d) glMaxShaderCompilerThreadsKHR(0) leaves nothing in flight: every subsequent
+//       GL_COMPLETION_STATUS_KHR reads GL_TRUE immediately, and compilation after it
+//       is synchronous. That is what the extension requires of a zero count.
+//   (e) THE ONE THAT NEEDS A GPU: the same frame, drawn with programs compiled and
+//       linked asynchronously and then with programs compiled and linked inline, must
+//       come out byte-identical under glReadPixels. Anything the worker thread got
+//       wrong about the compile environment, the reflection or the SPIR-V shows up
+//       here as a pixel difference and nowhere else.
+//
+// Backend selection is the module's usual one process, one backend (MOBILEGL_BACKEND_TYPE),
+// so this file runs twice per ctest invocation.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#include "Config.h"
+#include "MG_Util/Async/ShaderCompilePool.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+// GL_KHR_parallel_shader_compile. Spelled out rather than relying on the host's
+// glext.h: this module is built against whatever GL headers the machine has, and an
+// older one has neither token. Both are also GL_*_ARB with identical values.
+#ifndef GL_MAX_SHADER_COMPILER_THREADS_KHR
+#define GL_MAX_SHADER_COMPILER_THREADS_KHR 0x91B0
+#endif
+#ifndef GL_COMPLETION_STATUS_KHR
+#define GL_COMPLETION_STATUS_KHR 0x91B1
+#endif
+
+// The entry point under test, resolved by the linker straight into MobileGL_s like
+// every other gl* call in this module. Declared here for the same reason as the
+// tokens above.
+extern "C" void glMaxShaderCompilerThreadsKHR(GLuint count);
+
+namespace MGITest {
+    namespace {
+
+        using MobileGL::MG_Config::QuirkOverride;
+
+        // Same shape as the other scenarios: a two-attribute pass-through, so the only
+        // thing that can differ between the two compilation modes is the compilation.
+        constexpr const char* kVertexSource = R"(#version 330 core
+in vec2 aPos;
+in vec3 aColor;
+out vec3 vColor;
+void main() {
+    vColor = aColor;
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kFragmentSource = R"(#version 330 core
+in vec3 vColor;
+out vec4 oColor;
+void main() {
+    oColor = vec4(vColor, 1.0);
+}
+)";
+
+        // Asymmetric in both axes, so a mode difference that also happens to be a
+        // symmetry of the image cannot hide (the same reason OrientationScenario draws
+        // quadrants rather than stripes).
+        struct Vertex {
+            float x, y;
+            float r, g, b;
+        };
+
+        void AppendQuad(std::vector<Vertex>& out, float x0, float x1, float y0, float y1, float r, float g, float b) {
+            const Vertex bl{x0, y0, r, g, b};
+            const Vertex br{x1, y0, r, g, b};
+            const Vertex tr{x1, y1, r, g, b};
+            const Vertex tl{x0, y1, r, g, b};
+            out.insert(out.end(), {bl, br, tr, bl, tr, tl});
+        }
+
+        std::vector<Vertex> QuadrantGeometry() {
+            std::vector<Vertex> vertices;
+            vertices.reserve(24);
+            AppendQuad(vertices, -1.0f, 0.0f, -1.0f, 0.0f, 0.0f, 0.0f, 1.0f); // bottom-left: blue
+            AppendQuad(vertices, 0.0f, 1.0f, -1.0f, 0.0f, 0.0f, 1.0f, 0.0f);  // bottom-right: green
+            AppendQuad(vertices, -1.0f, 0.0f, 0.0f, 1.0f, 1.0f, 0.0f, 0.0f);  // top-left: red
+            AppendQuad(vertices, 0.0f, 1.0f, 0.0f, 1.0f, 1.0f, 1.0f, 1.0f);   // top-right: white
+            return vertices;
+        }
+
+        // Expensive enough that a compile is not instantaneous, and distinct per index so
+        // the source-hash memo never turns one into a no-op: without both properties the
+        // pool has no backlog and (a) has nothing to observe.
+        std::string BulkyFragmentSource(int index) {
+            std::string source = "#version 330 core\n";
+            source += "in vec3 vColor;\nout vec4 oColor;\n";
+            source += "uniform float uSeed" + std::to_string(index) + ";\n";
+            source += "void main() {\n    float acc = uSeed" + std::to_string(index) + ";\n";
+            for (int i = 0; i < 320; ++i) {
+                source += "    acc = acc * 1.0001 + sin(acc + " + std::to_string(i) + ".0) * cos(acc);\n";
+            }
+            source += "    oColor = vec4(vColor * acc, 1.0);\n}\n";
+            return source;
+        }
+
+        // MOBILEGL_ASYNC_SHADER_COMPILE decides the ambient mode; a scenario that wants
+        // the other one says so here and gets the ambient one back on scope exit. Forcing
+        // it in-process is what lets ONE ctest run compare the two modes against each
+        // other - the whole point of (e).
+        class AsyncModeScope {
+        public:
+            explicit AsyncModeScope(bool async) : m_saved(MobileGL::MG_Config::Features.AsyncShaderCompile) {
+                MobileGL::MG_Config::Features.AsyncShaderCompile =
+                    async ? QuirkOverride::ForceOn : QuirkOverride::ForceOff;
+            }
+            ~AsyncModeScope() { MobileGL::MG_Config::Features.AsyncShaderCompile = m_saved; }
+            AsyncModeScope(const AsyncModeScope&) = delete;
+            AsyncModeScope& operator=(const AsyncModeScope&) = delete;
+
+        private:
+            const QuirkOverride m_saved;
+        };
+
+        // glMaxShaderCompilerThreadsKHR writes process-wide state; a scenario that calls
+        // it has to put the pool back or it changes how every scenario after it compiles.
+        class CompilerThreadScope {
+        public:
+            CompilerThreadScope() = default;
+            ~CompilerThreadScope() {
+                MobileGL::MG_Util::Async::SetAsyncShaderCompileSuspended(false);
+                auto& pool = MobileGL::MG_Util::Async::ShaderCompilePool::Get();
+                pool.SetMaxConcurrency(pool.GetThreadCount());
+            }
+            CompilerThreadScope(const CompilerThreadScope&) = delete;
+            CompilerThreadScope& operator=(const CompilerThreadScope&) = delete;
+        };
+
+        GLint ShaderCompletion(GLuint shader) {
+            GLint status = -1;
+            glGetShaderiv(shader, GL_COMPLETION_STATUS_KHR, &status);
+            return status;
+        }
+
+        GLint ShaderCompileStatus(GLuint shader) {
+            GLint status = GL_FALSE;
+            glGetShaderiv(shader, GL_COMPILE_STATUS, &status);
+            return status;
+        }
+
+        std::string ShaderInfoLog(GLuint shader) {
+            GLint length = 0;
+            glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+            if (length <= 0) return std::string();
+            std::vector<char> buffer(static_cast<std::size_t>(length));
+            GLsizei written = 0;
+            glGetShaderInfoLog(shader, length, &written, buffer.data());
+            return std::string(buffer.data(), static_cast<std::size_t>(written));
+        }
+
+        class AsyncCompileScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                const std::vector<Vertex> vertices = QuadrantGeometry();
+                m_vertexCount = static_cast<int>(vertices.size());
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferData(GL_ARRAY_BUFFER, GLsizeiptr(vertices.size() * sizeof(Vertex)), vertices.data(),
+                             GL_STATIC_DRAW);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(0));
+                glEnableVertexAttribArray(1);
+                glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(8));
+                glBindVertexArray(0);
+
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                if (m_vbo != 0) glDeleteBuffers(1, &m_vbo);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+            }
+
+            // A fresh program every time, compiled and linked in whatever mode is in
+            // force. Reusing one would defeat the comparison: the second mode would just
+            // read the first mode's artifacts back out of the memo.
+            GLuint BuildProgram() {
+                std::string error;
+                const GLuint program = CompileProgram(kVertexSource, kFragmentSource, &error);
+                EXPECT_NE(program, 0u) << error;
+                return program;
+            }
+
+            Image DrawFrameWith(GLuint program) {
+                BindDefaultFramebuffer();
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                glDisable(GL_DEPTH_TEST);
+                glDisable(GL_BLEND);
+                glUseProgram(program);
+                glBindVertexArray(m_vao);
+                glDrawArrays(GL_TRIANGLES, 0, m_vertexCount);
+                glBindVertexArray(0);
+                Image image = ReadPixels(Gl().Width(), Gl().Height());
+                Gl().EndFrame();
+                return image;
+            }
+
+            // Enqueues `count` distinct heavy compiles and returns their names WITHOUT
+            // reading anything back, so the pool is left with a real backlog.
+            std::vector<GLuint> EnqueueBacklog(int count, int seedBase) {
+                std::vector<GLuint> shaders;
+                shaders.reserve(static_cast<std::size_t>(count));
+                m_sources.reserve(m_sources.size() + static_cast<std::size_t>(count));
+                for (int i = 0; i < count; ++i) {
+                    m_sources.push_back(BulkyFragmentSource(seedBase + i));
+                    const char* text = m_sources.back().c_str();
+                    const GLuint shader = glCreateShader(GL_FRAGMENT_SHADER);
+                    glShaderSource(shader, 1, &text, nullptr);
+                    glCompileShader(shader);
+                    shaders.push_back(shader);
+                }
+                return shaders;
+            }
+
+            GLuint m_vao = 0;
+            GLuint m_vbo = 0;
+            int m_vertexCount = 0;
+            // Kept alive for the whole case: glShaderSource copies, but keeping the
+            // strings makes a failure message able to name the source it came from.
+            std::vector<std::string> m_sources;
+        };
+
+        // ---- (a) + (b) ------------------------------------------------------------
+        // A backlog is enqueued, polled without joining, then forced to settle and
+        // checked for correctness. Both halves in one case on purpose: (b) is only
+        // interesting for shaders that (a) proved were genuinely still outstanding.
+        TEST_F(AsyncCompileScenario, CompletionStatusPollingThenForcedJoin) {
+            if (!Ready()) return;
+            const AsyncModeScope async(true);
+            const CompilerThreadScope threads;
+            // One worker, so the queue behind it is what the poll observes.
+            glMaxShaderCompilerThreadsKHR(1);
+
+            const std::vector<GLuint> shaders = EnqueueBacklog(64, 6000);
+
+            int outstanding = 0;
+            for (const GLuint shader : shaders) {
+                const GLint completion = ShaderCompletion(shader);
+                ASSERT_TRUE(completion == GL_TRUE || completion == GL_FALSE)
+                    << "GL_COMPLETION_STATUS_KHR returned " << completion;
+                if (completion == GL_FALSE) ++outstanding;
+            }
+            if (outstanding == 0) {
+                GTEST_SKIP() << "this machine drained 64 heavy compiles before the first poll; "
+                                "nothing was outstanding to observe";
+            }
+
+            // (b) Forced join: every one of them is correct, and usable.
+            for (const GLuint shader : shaders) {
+                EXPECT_EQ(ShaderCompileStatus(shader), GL_TRUE) << ShaderInfoLog(shader);
+                EXPECT_TRUE(ShaderInfoLog(shader).empty());
+                EXPECT_EQ(ShaderCompletion(shader), GL_TRUE) << "GL_COMPILE_STATUS must have joined";
+            }
+
+            // And a link over one of them really produces a usable program on this driver.
+            const GLuint vs = glCreateShader(GL_VERTEX_SHADER);
+            glShaderSource(vs, 1, &kVertexSource, nullptr);
+            glCompileShader(vs);
+            const GLuint program = glCreateProgram();
+            glAttachShader(program, vs);
+            glAttachShader(program, shaders.front());
+            glBindAttribLocation(program, 0, "aPos");
+            glBindAttribLocation(program, 1, "aColor");
+            glLinkProgram(program);
+            GLint linked = GL_FALSE;
+            glGetProgramiv(program, GL_LINK_STATUS, &linked);
+            EXPECT_EQ(linked, GL_TRUE);
+            EXPECT_GE(glGetUniformLocation(program, "uSeed6000"), 0);
+
+            glDeleteProgram(program);
+            glDeleteShader(vs);
+            for (const GLuint shader : shaders) glDeleteShader(shader);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // ---- (c) ------------------------------------------------------------------
+        // The extension string, read from a real backend that really brought a driver
+        // up. No mode forcing here: a backend builds its advertised list once, from the
+        // configuration in force at its first use, so the meaningful assertion is
+        // against the AMBIENT configuration - which is exactly what makes this case
+        // worth running in both of the suite's flag states.
+        TEST_F(AsyncCompileScenario, ExtensionStringMatchesTheConfiguration) {
+            if (!Ready()) return;
+            const bool expected = MobileGL::MG_Util::Async::AsyncShaderCompileEnabled();
+
+            const char* extensions = reinterpret_cast<const char*>(glGetString(GL_EXTENSIONS));
+            ASSERT_NE(extensions, nullptr);
+            const std::string extensionString(extensions);
+            const bool inString = extensionString.find("GL_KHR_parallel_shader_compile") != std::string::npos;
+            EXPECT_EQ(inString, expected)
+                << "backend " << Gl().BackendName() << " GL_EXTENSIONS = " << extensionString;
+
+            // LWJGL builds GLCapabilities from the INDEXED form on a core profile, so the
+            // two spellings disagreeing would be invisible to the check above and fatal
+            // to a real application.
+            GLint count = 0;
+            glGetIntegerv(GL_NUM_EXTENSIONS, &count);
+            ASSERT_GT(count, 0);
+            bool inIndexed = false;
+            for (GLint i = 0; i < count; ++i) {
+                const char* name = reinterpret_cast<const char*>(glGetStringi(GL_EXTENSIONS, GLuint(i)));
+                if (name != nullptr && std::string(name) == "GL_KHR_parallel_shader_compile") inIndexed = true;
+            }
+            EXPECT_EQ(inIndexed, expected);
+
+            // The companion query, which an application reads right after the string.
+            GLint maxThreads = -1;
+            glGetIntegerv(GL_MAX_SHADER_COMPILER_THREADS_KHR, &maxThreads);
+            if (expected) {
+                EXPECT_GE(maxThreads, 1);
+            } else {
+                EXPECT_EQ(maxThreads, 0);
+            }
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // ---- (d) ------------------------------------------------------------------
+        // A zero count must leave nothing in flight and keep it that way.
+        TEST_F(AsyncCompileScenario, ZeroCompilerThreadsSettlesEverythingImmediately) {
+            if (!Ready()) return;
+            const AsyncModeScope async(true);
+            const CompilerThreadScope threads;
+            glMaxShaderCompilerThreadsKHR(1);
+
+            const std::vector<GLuint> backlog = EnqueueBacklog(48, 6200);
+            glMaxShaderCompilerThreadsKHR(0);
+
+            for (const GLuint shader : backlog) {
+                EXPECT_EQ(ShaderCompletion(shader), GL_TRUE)
+                    << "glMaxShaderCompilerThreadsKHR(0) must join everything still in flight";
+                EXPECT_EQ(ShaderCompileStatus(shader), GL_TRUE) << ShaderInfoLog(shader);
+            }
+
+            // Compilation after the zero count is synchronous too.
+            const std::vector<GLuint> serial = EnqueueBacklog(6, 6300);
+            for (const GLuint shader : serial) {
+                EXPECT_EQ(ShaderCompletion(shader), GL_TRUE) << "a compile after a zero count must be synchronous";
+            }
+
+            for (const GLuint shader : backlog) glDeleteShader(shader);
+            for (const GLuint shader : serial) glDeleteShader(shader);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // ---- (e) ------------------------------------------------------------------
+        // The one that needs the GPU. Two programs, identical source, one built with
+        // compilation and linking on worker threads and one built inline; the frames
+        // they draw must be byte-identical.
+        //
+        // Compared through the DEFAULT framebuffer deliberately: that is where the
+        // backend's orientation and present path live, so the comparison covers the
+        // whole pipeline rather than the reflection tables alone.
+        TEST_F(AsyncCompileScenario, AsyncAndSyncProgramsRenderIdenticalFrames) {
+            if (!Ready()) return;
+
+            Image asyncImage;
+            {
+                const AsyncModeScope async(true);
+                const GLuint program = BuildProgram();
+                ASSERT_NE(program, 0u);
+                asyncImage = DrawFrameWith(program);
+                glDeleteProgram(program);
+            }
+
+            Image syncImage;
+            {
+                const AsyncModeScope async(false);
+                const GLuint program = BuildProgram();
+                ASSERT_NE(program, 0u);
+                syncImage = DrawFrameWith(program);
+                glDeleteProgram(program);
+            }
+
+            ASSERT_FALSE(asyncImage.Empty());
+            ASSERT_FALSE(syncImage.Empty());
+            // The frame is the expected one in the first place - two identically WRONG
+            // frames would otherwise pass.
+            EXPECT_EQ(asyncImage.QuadrantSignature(), "blue,green,red,white")
+                << "the asynchronously compiled program did not draw the expected frame";
+            EXPECT_EQ(asyncImage, syncImage)
+                << "asynchronous and synchronous compilation rendered different frames ("
+                << asyncImage.ByteDiffCount(syncImage) << " bytes differ); backend " << Gl().BackendName();
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // The same comparison over a batch, which is the shape a shaderpack load has:
+        // many programs enqueued before any of them is read back, then each one drawn.
+        // A per-worker state leak (glslang's thread-local pools are the obvious
+        // candidate) shows up here and not in the single-program case above.
+        TEST_F(AsyncCompileScenario, ABatchOfAsyncProgramsAllRenderCorrectly) {
+            if (!Ready()) return;
+            constexpr int kPrograms = 12;
+
+            std::vector<GLuint> programs;
+            {
+                const AsyncModeScope async(true);
+                const CompilerThreadScope threads;
+                glMaxShaderCompilerThreadsKHR(1);
+                // Everything enqueued before anything is read: the only shape in which
+                // more than one job is in flight at a time.
+                for (int i = 0; i < kPrograms; ++i) {
+                    programs.push_back(BuildProgram());
+                }
+            }
+
+            for (int i = 0; i < kPrograms; ++i) {
+                ASSERT_NE(programs[static_cast<std::size_t>(i)], 0u) << "program " << i;
+                const Image image = DrawFrameWith(programs[static_cast<std::size_t>(i)]);
+                EXPECT_EQ(image.QuadrantSignature(), "blue,green,red,white") << "program " << i;
+            }
+            for (const GLuint program : programs) glDeleteProgram(program);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/CrossFrameBufferScenario.cpp

@@ -0,0 +1,761 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/CrossFrameBufferScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario B - "the draw rendered last frame's buffer".
+//
+// The shipped bug (DirectVulkan, TryBindResolvedVertexBindings and the EBO
+// memo in UploadAndBindIndexBuffer): both memos revalidated themselves ACROSS a
+// frame boundary by comparing recorded per-buffer slice epochs, and on a match
+// skipped the per-frame buffer acquire. The acquire is the frame's content-sync
+// point; skipping it trusted the BumpSliceEpoch call-site inventory to cover
+// every way a buffer's GPU copy can go stale, and at least one path escaped it.
+// Result: a draw in a later frame renders from a STALE buffer slice - random
+// triangles in Minecraft/Sodium on Adreno, corrupted journeymap and
+// common-mods retraces.
+//
+// What pins it: mutate a buffer AFTER a frame boundary and BEFORE the next
+// draw, then prove the pixels show the NEW content. Every mutation API gets its
+// own test case, so a failure names the culprit rather than saying "buffers".
+// The index buffer is covered too: the EBO memo had exactly the same hole.
+//
+// The scene is deliberately trivial and entirely buffer-driven:
+//
+//     vertices 0..3   left half of the viewport,  RED
+//     vertices 4..7   right half of the viewport, GREEN
+//     indices  A      {0,1,2, 0,2,3}  -> the left, red quad
+//     indices  B      {4,5,6, 4,6,7}  -> the right, green quad
+//
+// A vertex-buffer test rewrites the left quad's colour red -> green and expects
+// the left half to turn green. An index-buffer test rewrites the indices
+// A -> B and expects the picture to jump from a red left half to a green right
+// half. Either way "stale" and "fresh" are different colours in different
+// places; no thresholds, no interpretation.
+//
+// Two families of scenario live here, and they catch different halves of the
+// same rule:
+//
+//   CrossFrameBufferScenario - one case per buffer-mutation API. Every one of
+//     these APIs is supposed to retire the memo; today they all do (each notify
+//     path bumps the slice epoch), so these pass on the buggy revision too.
+//     They are the standing statement of the contract: whatever a future memo
+//     keys on, a write through ANY of these APIs must reach the next frame's
+//     draw. They are also where a coherent persistent write - the one shape
+//     that changes a buffer with no GL call at all - is pinned.
+//
+//   StreamedArenaScenario - the case that actually caught the shipped bug. It
+//     attacks the other half of the rule: a buffer nobody wrote at all, whose
+//     GPU-side bytes moved out from under the memo anyway.
+
+#include <cstdio>
+#include <cstring>
+#include <functional>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVertexSource = R"(#version 330 core
+in vec2 aPos;
+in vec3 aColor;
+out vec3 vColor;
+void main() {
+    vColor = aColor;
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kFragmentSource = R"(#version 330 core
+in vec3 vColor;
+out vec4 oColor;
+void main() {
+    oColor = vec4(vColor, 1.0);
+}
+)";
+
+        struct Vertex {
+            float x, y;
+            float r, g, b;
+        };
+
+        constexpr int kLeftQuadFirstVertex = 0;
+        constexpr int kLeftQuadVertexCount = 4;
+        constexpr int kIndexCount = 6;
+
+        // Enough consecutive frames drawing the same VAO that any per-(VAO, frame)
+        // memo is fully armed before the mutation lands.
+        constexpr int kWarmupFrames = 3;
+
+        std::vector<Vertex> SceneVertices(bool leftQuadIsGreen) {
+            const float lr = leftQuadIsGreen ? 0.0f : 1.0f;
+            const float lg = leftQuadIsGreen ? 1.0f : 0.0f;
+            return {
+                // 0..3: left half
+                {-1.0f, -1.0f, lr, lg, 0.0f},
+                {0.0f, -1.0f, lr, lg, 0.0f},
+                {0.0f, 1.0f, lr, lg, 0.0f},
+                {-1.0f, 1.0f, lr, lg, 0.0f},
+                // 4..7: right half
+                {0.0f, -1.0f, 0.0f, 1.0f, 0.0f},
+                {1.0f, -1.0f, 0.0f, 1.0f, 0.0f},
+                {1.0f, 1.0f, 0.0f, 1.0f, 0.0f},
+                {0.0f, 1.0f, 0.0f, 1.0f, 0.0f},
+            };
+        }
+
+        const GLuint kIndicesLeftQuad[kIndexCount] = {0, 1, 2, 0, 2, 3};
+        const GLuint kIndicesRightQuad[kIndexCount] = {4, 5, 6, 4, 6, 7};
+
+        // How far inside each half the whole-region checks start. The two quads
+        // meet on a pixel boundary, so a couple of pixels of margin makes "every
+        // single pixel in the region" an achievable demand.
+        constexpr int kHalfInset = 2;
+
+        // Asserts the left and right halves of the viewport, with a message that
+        // says what the app had asked GL to draw by then.
+        //
+        // This counts EVERY pixel in each half rather than sampling its centre.
+        // Sampling two pixels was demonstrably too weak: a draw in which three of
+        // the left quad's four vertices still carry stale data paints a centre
+        // pixel of exactly the expected colour and passed the old assertion. That
+        // case is now a standing negative control - see
+        // PartialStalenessIsCaughtByWholeRegionChecks below, which constructs it
+        // deliberately and proves the region scan reports it.
+        void ExpectHalves(const Image& image, const char* expectedLeft, const char* expectedRight,
+                          const std::string& when) {
+            const int w = image.Width();
+            const int h = image.Height();
+            EXPECT_TRUE(RegionIsMostly(image, kHalfInset, w / 2 - kHalfInset, kHalfInset, h - kHalfInset, expectedLeft,
+                                       0.0, when + " [left half]"));
+            EXPECT_TRUE(RegionIsMostly(image, w / 2 + kHalfInset, w - kHalfInset, kHalfInset, h - kHalfInset,
+                                       expectedRight, 0.0, when + " [right half]"));
+        }
+
+        // How the app hands the new bytes to GL. Each is its own test case.
+        enum class Mutation {
+            SubData,            // glBufferSubData
+            MapWriteUnmap,      // glMapBufferRange(WRITE) + glUnmapBuffer
+            PersistentFlush,    // write through a persistent map + glFlushMappedBufferRange
+            PersistentCoherent, // write through a COHERENT persistent map, no GL call at all
+            OrphanReupload,     // glBufferData(NULL) then a full re-upload
+            CopySubData,        // glCopyBufferSubData from a staging buffer
+        };
+
+        bool NeedsImmutableStorage(Mutation mutation) {
+            return mutation == Mutation::PersistentFlush || mutation == Mutation::PersistentCoherent;
+        }
+
+        // The coherent variant is the one shape in which an application changes a
+        // buffer's contents with NO GL call whatsoever - the write lands in the
+        // mapping and that is the end of it. Sodium's chunk streaming is written
+        // this way, and it is the case a per-buffer "has anything changed?" epoch
+        // cannot see on its own.
+        bool NeedsCoherentMapping(Mutation mutation) {
+            return mutation == Mutation::PersistentCoherent;
+        }
+
+        class CrossFrameBufferScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string error;
+                m_program = CompileProgram(kVertexSource, kFragmentSource, &error);
+                ASSERT_NE(m_program, 0u) << error;
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "program setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                ReleaseBuffers();
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            // Builds the VAO/VBO/EBO. `immutable` switches to glBufferStorage plus a
+            // persistent mapping of both buffers, which is the only shape in which the
+            // persistent-write mutation is legal.
+            void BuildScene(bool immutable, bool coherent = false) {
+                const std::vector<Vertex> vertices = SceneVertices(/*leftQuadIsGreen=*/false);
+                m_vertexBytes = GLsizeiptr(vertices.size() * sizeof(Vertex));
+                m_indexBytes = GLsizeiptr(sizeof(kIndicesLeftQuad));
+
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glGenBuffers(1, &m_ebo);
+                glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, m_ebo);
+
+                if (immutable) {
+                    const GLbitfield storageFlags = GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_DYNAMIC_STORAGE_BIT |
+                                                    (coherent ? GL_MAP_COHERENT_BIT : 0);
+                    glBufferStorage(GL_ARRAY_BUFFER, m_vertexBytes, vertices.data(), storageFlags);
+                    glBufferStorage(GL_ELEMENT_ARRAY_BUFFER, m_indexBytes, kIndicesLeftQuad, storageFlags);
+                    const GLenum storageError = FirstGLError();
+                    if (storageError != GL_NO_ERROR) {
+                        m_storageUnsupported = true;
+                        m_storageError = storageError;
+                        return;
+                    }
+                    const GLbitfield mapFlags = GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT |
+                                                (coherent ? GL_MAP_COHERENT_BIT : GL_MAP_FLUSH_EXPLICIT_BIT);
+                    m_vertexMap =
+                        static_cast<unsigned char*>(glMapBufferRange(GL_ARRAY_BUFFER, 0, m_vertexBytes, mapFlags));
+                    m_indexMap = static_cast<unsigned char*>(
+                        glMapBufferRange(GL_ELEMENT_ARRAY_BUFFER, 0, m_indexBytes, mapFlags));
+                    if (m_vertexMap == nullptr || m_indexMap == nullptr) {
+                        m_storageUnsupported = true;
+                        m_storageError = FirstGLError();
+                        return;
+                    }
+                } else {
+                    glBufferData(GL_ARRAY_BUFFER, m_vertexBytes, vertices.data(), GL_STATIC_DRAW);
+                    glBufferData(GL_ELEMENT_ARRAY_BUFFER, m_indexBytes, kIndicesLeftQuad, GL_STATIC_DRAW);
+                }
+
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(0));
+                glEnableVertexAttribArray(1);
+                glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(8));
+                glBindVertexArray(0);
+
+                glGenBuffers(1, &m_staging);
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "scene setup left a GL error behind";
+            }
+
+            void ReleaseBuffers() {
+                if (m_vertexMap != nullptr || m_indexMap != nullptr) {
+                    glBindVertexArray(m_vao);
+                    if (m_vertexMap != nullptr) {
+                        glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                        glUnmapBuffer(GL_ARRAY_BUFFER);
+                    }
+                    if (m_indexMap != nullptr) {
+                        glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, m_ebo);
+                        glUnmapBuffer(GL_ELEMENT_ARRAY_BUFFER);
+                    }
+                    glBindVertexArray(0);
+                    m_vertexMap = nullptr;
+                    m_indexMap = nullptr;
+                }
+                if (m_staging != 0) glDeleteBuffers(1, &m_staging);
+                if (m_ebo != 0) glDeleteBuffers(1, &m_ebo);
+                if (m_vbo != 0) glDeleteBuffers(1, &m_vbo);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                m_staging = m_ebo = m_vbo = m_vao = 0;
+            }
+
+            void DrawScene() {
+                glDisable(GL_DEPTH_TEST);
+                glDisable(GL_BLEND);
+                glUseProgram(m_program);
+                glBindVertexArray(m_vao);
+                glDrawElements(GL_TRIANGLES, kIndexCount, GL_UNSIGNED_INT, nullptr);
+                glBindVertexArray(0);
+            }
+
+            void BeginFrame() {
+                BindDefaultFramebuffer();
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            }
+
+            Image ReadFrame() { return ReadPixels(Gl().Width(), Gl().Height()); }
+
+            // ---- the mutations ---------------------------------------------
+            // Each writes `newBytes` over the first `rangeBytes` of `buffer`;
+            // `wholeBytes`/`wholeSize` are the full contents an orphan+re-upload
+            // needs. `target` is the binding point the buffer normally lives at.
+            void ApplyMutation(Mutation mutation, GLenum target, GLuint buffer, unsigned char* persistentMap,
+                               const void* newBytes, GLsizeiptr rangeBytes, const void* wholeBytes,
+                               GLsizeiptr wholeSize) {
+                // The element-array binding is VAO state, so mutating the EBO happens
+                // with the scene's VAO bound - exactly as an application would.
+                glBindVertexArray(m_vao);
+                switch (mutation) {
+                case Mutation::SubData: {
+                    glBindBuffer(target, buffer);
+                    glBufferSubData(target, 0, rangeBytes, newBytes);
+                    break;
+                }
+                case Mutation::MapWriteUnmap: {
+                    glBindBuffer(target, buffer);
+                    void* mapped =
+                        glMapBufferRange(target, 0, rangeBytes, GL_MAP_WRITE_BIT | GL_MAP_INVALIDATE_RANGE_BIT);
+                    ASSERT_NE(mapped, nullptr) << "glMapBufferRange(WRITE) returned null";
+                    std::memcpy(mapped, newBytes, std::size_t(rangeBytes));
+                    ASSERT_EQ(glUnmapBuffer(target), GLboolean(GL_TRUE)) << "glUnmapBuffer reported data loss";
+                    break;
+                }
+                case Mutation::PersistentFlush: {
+                    ASSERT_NE(persistentMap, nullptr) << "no persistent mapping for this buffer";
+                    std::memcpy(persistentMap, newBytes, std::size_t(rangeBytes));
+                    glBindBuffer(target, buffer);
+                    glFlushMappedBufferRange(target, 0, rangeBytes);
+                    break;
+                }
+                case Mutation::PersistentCoherent: {
+                    // Deliberately no GL call: a coherent persistent mapping is a
+                    // promise that the write alone is enough.
+                    ASSERT_NE(persistentMap, nullptr) << "no persistent mapping for this buffer";
+                    std::memcpy(persistentMap, newBytes, std::size_t(rangeBytes));
+                    break;
+                }
+                case Mutation::OrphanReupload: {
+                    glBindBuffer(target, buffer);
+                    glBufferData(target, wholeSize, nullptr, GL_STATIC_DRAW);
+                    glBufferSubData(target, 0, wholeSize, wholeBytes);
+                    break;
+                }
+                case Mutation::CopySubData: {
+                    glBindBuffer(GL_COPY_READ_BUFFER, m_staging);
+                    glBufferData(GL_COPY_READ_BUFFER, rangeBytes, newBytes, GL_STATIC_DRAW);
+                    glBindBuffer(GL_COPY_WRITE_BUFFER, buffer);
+                    glCopyBufferSubData(GL_COPY_READ_BUFFER, GL_COPY_WRITE_BUFFER, 0, 0, rangeBytes);
+                    glBindBuffer(GL_COPY_WRITE_BUFFER, 0);
+                    glBindBuffer(GL_COPY_READ_BUFFER, 0);
+                    break;
+                }
+                }
+                glBindVertexArray(0);
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "the mutation itself raised a GL error";
+            }
+
+            // ---- the story -------------------------------------------------
+            // Steady state for a few frames, one frame boundary, then the
+            // mutation, then the draw that must show the new content.
+            void RunAcrossFrameBoundary(Mutation mutation, const std::function<void()>& mutate,
+                                        const char* expectedLeftAfter, const char* expectedRightAfter) {
+                ASSERT_NO_FATAL_FAILURE(BuildScene(NeedsImmutableStorage(mutation), NeedsCoherentMapping(mutation)));
+                if (m_storageUnsupported) {
+                    GTEST_SKIP() << "immutable/persistent buffer storage is unavailable on this stack ("
+                                 << GLErrorName(m_storageError) << "); the persistent-map mutation cannot "
+                                 << "be expressed here";
+                }
+
+                for (int frame = 0; frame < kWarmupFrames; ++frame) {
+                    BeginFrame();
+                    DrawScene();
+                    Gl().EndFrame();
+                }
+
+                BeginFrame();
+                DrawScene();
+                const Image before = ReadFrame();
+                ExpectHalves(before, "red", "black", "steady state before the mutation");
+                ASSERT_FALSE(::testing::Test::HasFailure())
+                    << "the scenario never reached its steady state, so nothing after this means anything";
+
+                // >>> a genuine frame boundary. Everything below happens in the NEXT
+                // frame, which is the whole point: a mutation inside one frame proves
+                // nothing about a memo that revalidates itself across frames.
+                Gl().EndFrame();
+
+                BeginFrame();
+                ASSERT_NO_FATAL_FAILURE(mutate());
+                DrawScene();
+                const Image after = ReadFrame();
+                Gl().EndFrame();
+
+                ExpectHalves(after, expectedLeftAfter, expectedRightAfter,
+                             "the draw after the mutation drew STALE buffer content");
+                EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            }
+
+            // The two things a scenario mutates.
+            void MutateVertexColorsToGreen(Mutation mutation) {
+                const std::vector<Vertex> updated = SceneVertices(/*leftQuadIsGreen=*/true);
+                const GLsizeiptr leftQuadBytes = GLsizeiptr(kLeftQuadVertexCount * sizeof(Vertex));
+                ApplyMutation(mutation, GL_ARRAY_BUFFER, m_vbo, m_vertexMap, updated.data() + kLeftQuadFirstVertex,
+                              leftQuadBytes, updated.data(), m_vertexBytes);
+            }
+
+            void MutateIndicesToRightQuad(Mutation mutation) {
+                ApplyMutation(mutation, GL_ELEMENT_ARRAY_BUFFER, m_ebo, m_indexMap, kIndicesRightQuad, m_indexBytes,
+                              kIndicesRightQuad, m_indexBytes);
+            }
+
+            unsigned int m_program = 0;
+            unsigned int m_vao = 0;
+            unsigned int m_vbo = 0;
+            unsigned int m_ebo = 0;
+            unsigned int m_staging = 0;
+            GLsizeiptr m_vertexBytes = 0;
+            GLsizeiptr m_indexBytes = 0;
+            unsigned char* m_vertexMap = nullptr;
+            unsigned char* m_indexMap = nullptr;
+            bool m_storageUnsupported = false;
+            unsigned int m_storageError = 0;
+        };
+
+        // ---- vertex buffer: the left quad must turn green ------------------
+
+        TEST_F(CrossFrameBufferScenario, VertexBufferSubData) {
+            RunAcrossFrameBoundary(
+                Mutation::SubData, [&] { MutateVertexColorsToGreen(Mutation::SubData); }, "green", "black");
+        }
+
+        TEST_F(CrossFrameBufferScenario, VertexMapWriteUnmap) {
+            RunAcrossFrameBoundary(
+                Mutation::MapWriteUnmap, [&] { MutateVertexColorsToGreen(Mutation::MapWriteUnmap); }, "green", "black");
+        }
+
+        TEST_F(CrossFrameBufferScenario, VertexPersistentMapFlush) {
+            RunAcrossFrameBoundary(
+                Mutation::PersistentFlush, [&] { MutateVertexColorsToGreen(Mutation::PersistentFlush); }, "green",
+                "black");
+        }
+
+        TEST_F(CrossFrameBufferScenario, VertexPersistentCoherentWrite) {
+            RunAcrossFrameBoundary(
+                Mutation::PersistentCoherent, [&] { MutateVertexColorsToGreen(Mutation::PersistentCoherent); }, "green",
+                "black");
+        }
+
+        TEST_F(CrossFrameBufferScenario, VertexOrphanAndReupload) {
+            RunAcrossFrameBoundary(
+                Mutation::OrphanReupload, [&] { MutateVertexColorsToGreen(Mutation::OrphanReupload); }, "green",
+                "black");
+        }
+
+        TEST_F(CrossFrameBufferScenario, VertexCopyBufferSubData) {
+            RunAcrossFrameBoundary(
+                Mutation::CopySubData, [&] { MutateVertexColorsToGreen(Mutation::CopySubData); }, "green", "black");
+        }
+
+        // ---- index buffer: the picture must jump to the right, green quad --
+        // The EBO memo had the same cross-frame hole as the vertex one, and no
+        // vertex-only test can see it.
+
+        TEST_F(CrossFrameBufferScenario, IndexBufferSubData) {
+            RunAcrossFrameBoundary(
+                Mutation::SubData, [&] { MutateIndicesToRightQuad(Mutation::SubData); }, "black", "green");
+        }
+
+        TEST_F(CrossFrameBufferScenario, IndexMapWriteUnmap) {
+            RunAcrossFrameBoundary(
+                Mutation::MapWriteUnmap, [&] { MutateIndicesToRightQuad(Mutation::MapWriteUnmap); }, "black", "green");
+        }
+
+        TEST_F(CrossFrameBufferScenario, IndexPersistentMapFlush) {
+            RunAcrossFrameBoundary(
+                Mutation::PersistentFlush, [&] { MutateIndicesToRightQuad(Mutation::PersistentFlush); }, "black",
+                "green");
+        }
+
+        // Kept, with its coverage stated exactly, because it is the one case in
+        // this file that is served a stale slice by the buggy revision and passes
+        // anyway - and a test that reads as coverage without being coverage is
+        // worse than no test.
+        //
+        // COVERS: the coherent-persistent index contract - a write into a coherent
+        // persistent mapping, with no GL call at all, must reach the next frame's
+        // draw. That is a real contract and this is the only case that states it
+        // for indices.
+        //
+        // DOES NOT COVER: the EBO cross-frame memo. Instrumented against the
+        // re-enabled buggy path, it enters the cross-frame branch 4 times and is
+        // served its recorded slice all 4 times - and still passes, because the
+        // backend adopted the persistent map into that very storage
+        // (AcquirePersistentMap succeeded), so the application's writes landed in
+        // the bytes the "stale" slice names. It would only discriminate on a stack
+        // where that adoption is declined and the CPU shadow stays authoritative;
+        // measured over this whole module, 50 of 50 coherent persistent write maps
+        // were adopted. See ResidentIndexScenario.cpp for the full account.
+        TEST_F(CrossFrameBufferScenario, IndexPersistentCoherentWrite) {
+            RunAcrossFrameBoundary(
+                Mutation::PersistentCoherent, [&] { MutateIndicesToRightQuad(Mutation::PersistentCoherent); }, "black",
+                "green");
+        }
+
+        TEST_F(CrossFrameBufferScenario, IndexOrphanAndReupload) {
+            RunAcrossFrameBoundary(
+                Mutation::OrphanReupload, [&] { MutateIndicesToRightQuad(Mutation::OrphanReupload); }, "black",
+                "green");
+        }
+
+        TEST_F(CrossFrameBufferScenario, IndexCopyBufferSubData) {
+            RunAcrossFrameBoundary(
+                Mutation::CopySubData, [&] { MutateIndicesToRightQuad(Mutation::CopySubData); }, "black", "green");
+        }
+
+        // ---- a self-test of the assertions, not of MobileGL ------------------
+        //
+        // Every case above leans on ExpectHalves. ExpectHalves used to sample the
+        // centre pixel of each half - two pixels for a 12288-pixel readback - and
+        // that is measurably too weak to stand behind a claim about buffer
+        // freshness: a quad whose four vertices are only PARTLY updated still
+        // paints a sampled centre the expected colour, because the centre is a
+        // barycentric blend dominated by the vertices that DID update.
+        //
+        // So construct that case on purpose. Update the left quad's colour to
+        // green in the buffer but leave exactly one of its four vertices holding
+        // the old red, once for each vertex, and check two things:
+        //
+        //   - the whole-region scan reports every one of the four (the tightening
+        //     is real, and this test fails the moment someone loosens it back to
+        //     sampling);
+        //   - at least one of the four is invisible to a single centre sample
+        //     (the blind spot was real, and this records which vertices it hid).
+        //
+        // Nothing here calls a memo path; it is the assertion itself under test.
+        TEST_F(CrossFrameBufferScenario, PartialStalenessIsCaughtByWholeRegionChecks) {
+            ASSERT_NO_FATAL_FAILURE(BuildScene(/*immutable=*/false));
+
+            const std::vector<Vertex> allGreen = SceneVertices(/*leftQuadIsGreen=*/true);
+            const std::vector<Vertex> allRed = SceneVertices(/*leftQuadIsGreen=*/false);
+            const GLsizeiptr leftQuadBytes = GLsizeiptr(kLeftQuadVertexCount * sizeof(Vertex));
+
+            int centreSampleMissed = 0;
+            std::string missedVertices;
+            for (int staleVertex = 0; staleVertex < kLeftQuadVertexCount; ++staleVertex) {
+                // Every left-quad vertex turns green except this one.
+                std::vector<Vertex> partial(allGreen.begin(), allGreen.begin() + kLeftQuadVertexCount);
+                partial[std::size_t(staleVertex)] = allRed[std::size_t(staleVertex)];
+
+                glBindVertexArray(m_vao);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferSubData(GL_ARRAY_BUFFER, 0, leftQuadBytes, partial.data());
+                glBindVertexArray(0);
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "the partial update itself raised a GL error";
+
+                BeginFrame();
+                DrawScene();
+                const Image image = ReadFrame();
+                Gl().EndFrame();
+
+                const int w = image.Width();
+                const int h = image.Height();
+                const RegionScan scan =
+                    ScanRegion(image, kHalfInset, w / 2 - kHalfInset, kHalfInset, h - kHalfInset, "green");
+                EXPECT_GT(scan.offenders, 0)
+                    << "vertex " << staleVertex << " of the left quad kept its stale red colour and the "
+                    << "whole-region scan saw nothing wrong across " << scan.total << " pixels - the assertion "
+                    << "is not tight enough to stand behind any freshness claim in this file";
+
+                // What the old two-pixel form of ExpectHalves would have concluded.
+                if (std::strcmp(image.ColorName(w / 4, h / 2), "green") == 0) {
+                    ++centreSampleMissed;
+                    if (!missedVertices.empty()) missedVertices += ",";
+                    missedVertices += std::to_string(staleVertex);
+                }
+            }
+
+            EXPECT_GT(centreSampleMissed, 0)
+                << "no single-vertex staleness was invisible to a centre sample, so this negative control "
+                << "is no longer demonstrating anything - re-derive it before trusting it";
+            if (centreSampleMissed > 0) {
+                RecordProperty("centre_sample_blind_to_stale_vertices", missedVertices);
+                std::fprintf(stderr,
+                             "[itest] whole-region scan caught all %d single-stale-vertex cases; a centre "
+                             "sample alone was blind to %d of them (vertices %s)\n",
+                             kLeftQuadVertexCount, centreSampleMissed, missedVertices.c_str());
+            }
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // ---- the same bug, seen from the other side --------------------------
+        //
+        // The mutation cases above ask "did the new bytes reach the GPU?". This
+        // one asks the question a STREAMED buffer forces: "do the old bytes even
+        // still exist?".
+        //
+        // A GL_STREAM_DRAW / GL_DYNAMIC_DRAW buffer is not given permanent GPU
+        // storage. Every frame its contents are copied into that frame's
+        // transient upload arena, which is a bump allocator reset at the start of
+        // each frame slot - so a slice handed out in frame N names bytes that
+        // frame N+frames-in-flight hands to whoever uploads first. A memo that
+        // revalidates across a frame boundary and skips the acquire never
+        // re-uploads, so it keeps binding an offset the arena has since given
+        // away: the draw reads whatever the next tenant put there. That is the
+        // "random triangles" shape of this bug - the buffer nobody touched is the
+        // one that renders wrong.
+        //
+        // The scene makes the next tenant deterministic instead of arbitrary: a
+        // second streamed object of exactly the same size is uploaded and drawn
+        // FIRST in every frame, so it lands on precisely the bytes the memo still
+        // points at. A draw that renders the decoy's geometry instead of its own
+        // is unmissable.
+
+        class StreamedArenaScenario : public ScenarioTest {
+        protected:
+            static constexpr int kQuietFrames = 2; // frames in which only the subject draws
+            static constexpr int kChurnFrames = 8; // > frames-in-flight, so the ring wraps
+
+            struct StreamedObject {
+                unsigned int vao = 0;
+                unsigned int vbo = 0;
+                unsigned int ebo = 0;
+            };
+
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string error;
+                m_program = CompileProgram(kVertexSource, kFragmentSource, &error);
+                ASSERT_NE(m_program, 0u) << error;
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                for (StreamedObject* object : {&m_subject, &m_decoy}) {
+                    if (object->ebo != 0) glDeleteBuffers(1, &object->ebo);
+                    if (object->vbo != 0) glDeleteBuffers(1, &object->vbo);
+                    if (object->vao != 0) glDeleteVertexArrays(1, &object->vao);
+                    *object = StreamedObject{};
+                }
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            // GL_STREAM_DRAW is what puts a buffer on the transient arena
+            // (ShouldUseTransientVertexIndexBuffer) - and what Minecraft uses for
+            // exactly this kind of geometry.
+            void BuildStreamedObject(StreamedObject& object, const std::vector<Vertex>& vertices,
+                                     const GLuint (&indices)[kIndexCount]) {
+                glGenVertexArrays(1, &object.vao);
+                glBindVertexArray(object.vao);
+                glGenBuffers(1, &object.vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, object.vbo);
+                glBufferData(GL_ARRAY_BUFFER, GLsizeiptr(vertices.size() * sizeof(Vertex)), vertices.data(),
+                             GL_STREAM_DRAW);
+                glGenBuffers(1, &object.ebo);
+                glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, object.ebo);
+                glBufferData(GL_ELEMENT_ARRAY_BUFFER, GLsizeiptr(sizeof(indices)), indices, GL_STREAM_DRAW);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(0));
+                glEnableVertexAttribArray(1);
+                glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(8));
+                glBindVertexArray(0);
+            }
+
+            void Draw(const StreamedObject& object) {
+                glDisable(GL_DEPTH_TEST);
+                glDisable(GL_BLEND);
+                glUseProgram(m_program);
+                glBindVertexArray(object.vao);
+                glDrawElements(GL_TRIANGLES, kIndexCount, GL_UNSIGNED_INT, nullptr);
+                glBindVertexArray(0);
+            }
+
+            // Re-uploading the decoy is what forces it onto a fresh arena slice
+            // this frame - i.e. what makes it the arena's next tenant.
+            void RestreamDecoy(const std::vector<Vertex>& vertices, const GLuint (&indices)[kIndexCount]) {
+                glBindVertexArray(m_decoy.vao);
+                glBindBuffer(GL_ARRAY_BUFFER, m_decoy.vbo);
+                glBufferSubData(GL_ARRAY_BUFFER, 0, GLsizeiptr(vertices.size() * sizeof(Vertex)), vertices.data());
+                glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, m_decoy.ebo);
+                glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, GLsizeiptr(sizeof(indices)), indices);
+                glBindVertexArray(0);
+            }
+
+            unsigned int m_program = 0;
+            StreamedObject m_subject;
+            StreamedObject m_decoy;
+        };
+
+        // Vertex data. Subject and decoy differ in geometry AND colour, so a
+        // subject draw that reads the decoy's arena bytes paints the decoy's quad.
+        TEST_F(StreamedArenaScenario, StreamedVertexDataSurvivesArenaRecycling) {
+            const std::vector<Vertex> full = SceneVertices(/*leftQuadIsGreen=*/false);
+            const std::vector<Vertex> subjectVertices(full.begin(), full.begin() + 4);   // left, red
+            const std::vector<Vertex> decoyVertices(full.begin() + 4, full.begin() + 8); // right, green
+            ASSERT_EQ(subjectVertices.size(), decoyVertices.size());                     // same arena footprint
+
+            BuildStreamedObject(m_subject, subjectVertices, kIndicesLeftQuad);
+            BuildStreamedObject(m_decoy, decoyVertices, kIndicesLeftQuad);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "scene setup left a GL error behind";
+
+            // Quiet frames: the subject is the only thing uploading, so its data
+            // sits at the head of the arena and its memo records that offset.
+            for (int frame = 0; frame < kQuietFrames; ++frame) {
+                BindDefaultFramebuffer();
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                Draw(m_subject);
+                Gl().EndFrame();
+            }
+
+            // Churn frames: the decoy re-streams and draws first every frame. The
+            // subject is never touched again - it must still render itself.
+            for (int frame = 0; frame < kChurnFrames; ++frame) {
+                BindDefaultFramebuffer();
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                RestreamDecoy(decoyVertices, kIndicesLeftQuad);
+                Draw(m_decoy);
+                Draw(m_subject);
+                const Image image = ReadPixels(Gl().Width(), Gl().Height());
+                ExpectHalves(image, "red", "green",
+                             "churn frame " + std::to_string(frame) +
+                                 ": the untouched streamed vertex buffer rendered someone else's arena bytes");
+                Gl().EndFrame();
+            }
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // Index data. Both objects carry the SAME eight vertices, so only the
+        // element buffer can decide which half is drawn - this isolates the EBO
+        // memo, which had its own copy of the cross-frame hole.
+        //
+        // COVERS: that an untouched streamed index buffer still renders its own
+        // geometry after the arena it lives in has been recycled by another
+        // object - the index-side statement of the invariant the vertex case
+        // above actually catches.
+        //
+        // DOES NOT COVER: the EBO cross-frame memo. Instrumented against the
+        // re-enabled buggy path this case reaches that branch ZERO times: the memo
+        // is recorded only on the RESIDENT index path (UploadAndBindIndexBuffer
+        // stores it in the arm after AcquireResidentSlice), and a streamed EBO
+        // never gets there. So it passes on the buggy revision exactly as it does
+        // on the fixed one, and it is not evidence about the fix.
+        //
+        // It stays because it is the tripwire for the change that would make the
+        // EBO memo dangerous: memoise the streamed index path - the obvious next
+        // step for the same optimisation - and the reach stops being zero and this
+        // test fails on the first churn frame. See ResidentIndexScenario.cpp.
+        TEST_F(StreamedArenaScenario, StreamedIndexDataSurvivesArenaRecycling) {
+            const std::vector<Vertex> shared = SceneVertices(/*leftQuadIsGreen=*/false);
+
+            BuildStreamedObject(m_subject, shared, kIndicesLeftQuad); // draws the left, red quad
+            BuildStreamedObject(m_decoy, shared, kIndicesRightQuad);  // draws the right, green quad
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "scene setup left a GL error behind";
+
+            for (int frame = 0; frame < kQuietFrames; ++frame) {
+                BindDefaultFramebuffer();
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                Draw(m_subject);
+                Gl().EndFrame();
+            }
+
+            for (int frame = 0; frame < kChurnFrames; ++frame) {
+                BindDefaultFramebuffer();
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                RestreamDecoy(shared, kIndicesRightQuad);
+                Draw(m_decoy);
+                Draw(m_subject);
+                const Image image = ReadPixels(Gl().Width(), Gl().Height());
+                ExpectHalves(image, "red", "green",
+                             "churn frame " + std::to_string(frame) +
+                                 ": the untouched streamed index buffer rendered someone else's arena bytes");
+                Gl().EndFrame();
+            }
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/MultiDrawScenario.cpp

@@ -0,0 +1,522 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/MultiDrawScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario D - glMultiDrawElements(BaseVertex) against the draws it stands for.
+//
+// Neither entry point exists in OpenGL ES, so DirectGLES emulates both through a
+// ladder of tiers (MG_Backend/DirectGLES/MultiDraw.cpp): a native
+// glMultiDrawElementsBaseVertexEXT, synthesized indirect commands drawn one at a
+// time or in one batch, a per-sub-draw replay, a CPU rewrite of the index stream,
+// and a compute shader that flattens the whole batch into a single draw. They
+// share nothing but their contract, which is the one thing asserted here:
+//
+//     a multi-draw must paint exactly what the unrolled single draws paint.
+//
+// The reference side never enters the emulation - it is a loop of
+// glDrawElementsBaseVertex / glDrawElements - so a tier cannot make itself look
+// right by breaking both sides the same way.
+//
+// The Minecraft retraces already cover the common shape (GL_UNSIGNED_INT indices
+// in a bound element array buffer, small base vertices, GL_TRIANGLES) on every
+// tier. What they contain none of, and what these cases are for, is the set of
+// shapes where a tier has to decline or compensate rather than replay:
+//
+//   * narrow index types, where a rewritten stream has to widen (BYTE/SHORT);
+//   * a base vertex past the index type's range, where folding it into the
+//     indices at the source width silently wraps - GL adds base vertices at full
+//     precision, so `ushort index 10 + baseVertex 70000` is vertex 70010 and not
+//     vertex 4474;
+//   * primitive restart, where a rewritten stream must carry the sentinel across
+//     unrebased or the restart is lost and the strip welds shut;
+//   * client-memory index arrays, which have no buffer for the indirect tiers to
+//     address or for the compute tier to read;
+//   * a strip mode, which the flattening tier must decline outright because
+//     concatenation would weld one sub-draw's last primitive to the next
+//     sub-draw's first.
+//
+// One process is one tier (MOBILEGL_ESPRYT_MULTIDRAW_MODE is read once at
+// startup), so a single run exercises whichever tier this driver resolved to.
+// Running the binary once per mode is what covers the ladder; each run is a
+// complete, self-contained proof for the tier it landed on.
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glext.h>
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVertexSource = R"(#version 330 core
+layout(location = 0) in vec2 aPos;
+layout(location = 1) in vec3 aColor;
+out vec3 vColor;
+void main() {
+    vColor = aColor;
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kFragmentSource = R"(#version 330 core
+in vec3 vColor;
+out vec4 oColor;
+void main() {
+    oColor = vec4(vColor, 1.0);
+}
+)";
+
+        struct Vertex {
+            float x, y;
+            float r, g, b;
+        };
+
+        // Four column quads spanning the viewport left to right, in four colours,
+        // so a sub-draw that lands in the wrong place, draws the wrong vertices or
+        // does not draw at all changes the picture rather than hiding inside it.
+        constexpr int kColumns = 4;
+
+        const Rgba8 kColumnColors[kColumns] = {
+            {255, 0, 0, 255},
+            {0, 255, 0, 255},
+            {0, 0, 255, 255},
+            {255, 255, 255, 255},
+        };
+
+        // `padVertices` leading dummies force every sub-draw to need its own base
+        // vertex: without one applied, a draw reads the padding and paints black.
+        std::vector<Vertex> ColumnVertices(int padVertices) {
+            std::vector<Vertex> vertices(static_cast<std::size_t>(padVertices), Vertex{0.0f, 0.0f, 0.0f, 0.0f, 0.0f});
+            for (int column = 0; column < kColumns; ++column) {
+                const float x0 = -1.0f + 2.0f * static_cast<float>(column) / kColumns;
+                const float x1 = -1.0f + 2.0f * static_cast<float>(column + 1) / kColumns;
+                const Rgba8 color = kColumnColors[column];
+                const float r = color.r / 255.0f;
+                const float g = color.g / 255.0f;
+                const float b = color.b / 255.0f;
+                vertices.push_back({x0, -1.0f, r, g, b});
+                vertices.push_back({x1, -1.0f, r, g, b});
+                vertices.push_back({x1, 1.0f, r, g, b});
+                vertices.push_back({x0, 1.0f, r, g, b});
+            }
+            return vertices;
+        }
+
+        // Every sub-draw uses the SAME six indices, 0..3 relative to its own quad;
+        // only the base vertex tells the columns apart. That makes the base vertex
+        // the load-bearing part of the batch.
+        const std::uint32_t kQuadIndices[6] = {0, 1, 2, 0, 2, 3};
+
+        // One column, as a restart-separated pair of triangle strips. Two strips in
+        // one sub-draw means the sentinel is genuinely interior: drop it and the two
+        // halves weld into a single strip that paints across the gap between them.
+        // Indices are relative to the sub-draw's own quad, like kQuadIndices.
+        template <typename Index>
+        std::vector<Index> RestartStripIndices(Index restartSentinel) {
+            // 3,0,2,1 is the strip winding of the quad; splitting it around the
+            // sentinel gives two degenerate-free halves that redraw the same area.
+            return {Index{3}, Index{0}, Index{2}, restartSentinel, Index{0}, Index{2}, Index{1}};
+        }
+
+        class MultiDrawScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string error;
+                m_program = CompileProgram(kVertexSource, kFragmentSource, &error);
+                ASSERT_NE(m_program, 0u) << error;
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "program setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                ReleaseBuffers();
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            // VAO + VBO, and an EBO only when `indexBytes` is non-null: a null one
+            // leaves GL_ELEMENT_ARRAY_BUFFER unbound so the sub-draws address client
+            // memory, which is the shape that forces the buffer-reading tiers out.
+            void BuildScene(int padVertices, const void* indexBytes, std::size_t indexByteCount) {
+                ReleaseBuffers();
+                const std::vector<Vertex> vertices = ColumnVertices(padVertices);
+
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferData(GL_ARRAY_BUFFER, static_cast<GLsizeiptr>(vertices.size() * sizeof(Vertex)),
+                             vertices.data(), GL_STATIC_DRAW);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<const void*>(0));
+                glEnableVertexAttribArray(1);
+                glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, sizeof(Vertex),
+                                      reinterpret_cast<const void*>(sizeof(float) * 2));
+
+                if (indexBytes != nullptr) {
+                    glGenBuffers(1, &m_ebo);
+                    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, m_ebo);
+                    glBufferData(GL_ELEMENT_ARRAY_BUFFER, static_cast<GLsizeiptr>(indexByteCount), indexBytes,
+                                 GL_STATIC_DRAW);
+                }
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "scene setup left a GL error behind";
+            }
+
+            void ReleaseBuffers() {
+                if (m_ebo != 0) glDeleteBuffers(1, &m_ebo);
+                if (m_vbo != 0) glDeleteBuffers(1, &m_vbo);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                m_ebo = 0;
+                m_vbo = 0;
+                m_vao = 0;
+            }
+
+            GLuint m_program = 0;
+            GLuint m_vao = 0;
+            GLuint m_vbo = 0;
+            GLuint m_ebo = 0;
+        };
+
+        // Runs `draw`, reads the default framebuffer back and returns the image.
+        template <typename DrawFn>
+        Image RenderPass(GLuint program, GLuint vao, DrawFn&& draw) {
+            BindDefaultFramebuffer();
+            glViewport(0, 0, HeadlessGL::Get().Width(), HeadlessGL::Get().Height());
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            glUseProgram(program);
+            glBindVertexArray(vao);
+            draw();
+            return ReadPixels(HeadlessGL::Get().Width(), HeadlessGL::Get().Height());
+        }
+
+        // The whole point of the file: two renderings of the same geometry, one
+        // through the multi-draw emulation and one through the single-draw entry
+        // points it stands for, must be identical to the byte.
+        void ExpectSameImage(const Image& multiDraw, const Image& unrolled, const std::string& what) {
+            ASSERT_FALSE(multiDraw.Empty()) << what << ": multi-draw readback was empty";
+            ASSERT_FALSE(unrolled.Empty()) << what << ": reference readback was empty";
+            EXPECT_EQ(multiDraw, unrolled)
+                << what << ": glMultiDraw* painted something else than the draws it stands for ("
+                << multiDraw.ByteDiffCount(unrolled) << " bytes differ; multi-draw quadrants "
+                << multiDraw.QuadrantSignature() << ", unrolled quadrants " << unrolled.QuadrantSignature() << ")";
+            // A pair of blank frames would satisfy the comparison above and prove
+            // nothing at all - the failure mode a multi-draw path most often has is
+            // drawing NOTHING (see the shipped glMultiDrawElementsBaseVertexEXT stub
+            // that silently dropped every draw). Demand the columns really landed.
+            EXPECT_NE(multiDraw.QuadrantSignature(), "black,black,black,black") << what << ": nothing was drawn at all";
+        }
+
+        // ---- GL_UNSIGNED_INT indices in a buffer, per-sub-draw base vertices ----
+
+        TEST_F(MultiDrawScenario, BaseVertexBatchMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 5; // odd, so nothing lines up by accident
+            BuildScene(kPad, kQuadIndices, sizeof(kQuadIndices));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = 6;
+                offsets[i] = reinterpret_cast<const void*>(0);
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLES, counts, GL_UNSIGNED_INT, offsets, kColumns, baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElementsBaseVertex(GL_TRIANGLES, counts[i], GL_UNSIGNED_INT, offsets[i], baseVertices[i]);
+                }
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "GL_UNSIGNED_INT indices, per-sub-draw base vertices");
+        }
+
+        // ---- glMultiDrawElements: no base vertices, distinct index offsets ----
+
+        TEST_F(MultiDrawScenario, PlainBatchMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            // No padding and no base vertices: each sub-draw reaches its own column
+            // through its index offset instead.
+            std::vector<std::uint32_t> indices;
+            for (int column = 0; column < kColumns; ++column) {
+                for (const std::uint32_t index : kQuadIndices) {
+                    indices.push_back(index + static_cast<std::uint32_t>(column * 4));
+                }
+            }
+            BuildScene(0, indices.data(), indices.size() * sizeof(std::uint32_t));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = 6;
+                offsets[i] = reinterpret_cast<const void*>(static_cast<std::uintptr_t>(i * 6 * sizeof(std::uint32_t)));
+            }
+
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElements(GL_TRIANGLES, counts, GL_UNSIGNED_INT, offsets, kColumns);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElements(GL_TRIANGLES, counts[i], GL_UNSIGNED_INT, offsets[i]);
+                }
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "glMultiDrawElements with no base vertices");
+        }
+
+        // ---- narrow index types ----
+        // A tier that rewrites the stream emits GL_UNSIGNED_INT whatever came in,
+        // so these two say the widening reproduces the original draw exactly.
+
+        TEST_F(MultiDrawScenario, UnsignedShortBatchMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 3;
+            std::uint16_t indices[6];
+            for (int i = 0; i < 6; ++i)
+                indices[i] = static_cast<std::uint16_t>(kQuadIndices[i]);
+            BuildScene(kPad, indices, sizeof(indices));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = 6;
+                offsets[i] = reinterpret_cast<const void*>(0);
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLES, counts, GL_UNSIGNED_SHORT, offsets, kColumns, baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElementsBaseVertex(GL_TRIANGLES, counts[i], GL_UNSIGNED_SHORT, offsets[i], baseVertices[i]);
+                }
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "GL_UNSIGNED_SHORT indices");
+        }
+
+        TEST_F(MultiDrawScenario, UnsignedByteBatchMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 3;
+            std::uint8_t indices[6];
+            for (int i = 0; i < 6; ++i)
+                indices[i] = static_cast<std::uint8_t>(kQuadIndices[i]);
+            // 24 bytes: a word multiple, which the compute tier needs of the source
+            // buffer when the index type is narrower than a word.
+            std::uint8_t padded[24] = {};
+            for (int i = 0; i < 6; ++i)
+                padded[i] = indices[i];
+            BuildScene(kPad, padded, sizeof(padded));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = 6;
+                offsets[i] = reinterpret_cast<const void*>(0);
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLES, counts, GL_UNSIGNED_BYTE, offsets, kColumns, baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElementsBaseVertex(GL_TRIANGLES, counts[i], GL_UNSIGNED_BYTE, offsets[i], baseVertices[i]);
+                }
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "GL_UNSIGNED_BYTE indices");
+        }
+
+        // ---- a base vertex the index type cannot spell ----
+        // GL adds the base vertex at full precision, so folding it into a
+        // GL_UNSIGNED_SHORT index stream at the source width wraps and addresses the
+        // wrong vertex. The columns here start past 65535, which no ushort index can
+        // reach on its own.
+
+        TEST_F(MultiDrawScenario, BaseVertexBeyondIndexTypeRangeMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 70000; // > 0xFFFF
+            std::uint16_t indices[6];
+            for (int i = 0; i < 6; ++i)
+                indices[i] = static_cast<std::uint16_t>(kQuadIndices[i]);
+            BuildScene(kPad, indices, sizeof(indices));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = 6;
+                offsets[i] = reinterpret_cast<const void*>(0);
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLES, counts, GL_UNSIGNED_SHORT, offsets, kColumns, baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElementsBaseVertex(GL_TRIANGLES, counts[i], GL_UNSIGNED_SHORT, offsets[i], baseVertices[i]);
+                }
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "base vertex past the GL_UNSIGNED_SHORT range");
+        }
+
+        // ---- client-memory index arrays ----
+        // No element array buffer, so the indirect tiers have nothing to address and
+        // the compute tier nothing to read; both must decline and hand the batch to
+        // a tier that can replay it.
+
+        TEST_F(MultiDrawScenario, ClientSideIndicesBatchMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 5;
+            BuildScene(kPad, nullptr, 0);
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = 6;
+                offsets[i] = kQuadIndices;
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLES, counts, GL_UNSIGNED_INT, offsets, kColumns, baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElementsBaseVertex(GL_TRIANGLES, counts[i], GL_UNSIGNED_INT, offsets[i], baseVertices[i]);
+                }
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "client-memory index arrays");
+        }
+
+        // ---- primitive restart inside a strip ----
+        // Two things at once: a strip mode, which the flattening tier must decline
+        // because concatenation would weld sub-draws together, and a restart
+        // sentinel, which any tier that rewrites indices must carry across without
+        // adding the base vertex to it.
+
+        TEST_F(MultiDrawScenario, PrimitiveRestartStripBatchMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 5;
+            const std::vector<std::uint32_t> indices = RestartStripIndices<std::uint32_t>(0xFFFFFFFFu);
+            BuildScene(kPad, indices.data(), indices.size() * sizeof(std::uint32_t));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = static_cast<GLsizei>(indices.size());
+                offsets[i] = reinterpret_cast<const void*>(0);
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            glEnable(GL_PRIMITIVE_RESTART_FIXED_INDEX);
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLE_STRIP, counts, GL_UNSIGNED_INT, offsets, kColumns,
+                                              baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElementsBaseVertex(GL_TRIANGLE_STRIP, counts[i], GL_UNSIGNED_INT, offsets[i],
+                                             baseVertices[i]);
+                }
+            });
+            glDisable(GL_PRIMITIVE_RESTART_FIXED_INDEX);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "GL_TRIANGLE_STRIP with primitive restart");
+        }
+
+        // Same, with GL_UNSIGNED_SHORT: the sentinel a rewritten stream has to
+        // recognise is the index TYPE's all-ones value, not the rewritten stream's.
+        TEST_F(MultiDrawScenario, PrimitiveRestartUnsignedShortBatchMatchesUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 5;
+            const std::vector<std::uint16_t> indices = RestartStripIndices<std::uint16_t>(0xFFFFu);
+            BuildScene(kPad, indices.data(), indices.size() * sizeof(std::uint16_t));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                counts[i] = static_cast<GLsizei>(indices.size());
+                offsets[i] = reinterpret_cast<const void*>(0);
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            glEnable(GL_PRIMITIVE_RESTART_FIXED_INDEX);
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLE_STRIP, counts, GL_UNSIGNED_SHORT, offsets, kColumns,
+                                              baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    glDrawElementsBaseVertex(GL_TRIANGLE_STRIP, counts[i], GL_UNSIGNED_SHORT, offsets[i],
+                                             baseVertices[i]);
+                }
+            });
+            glDisable(GL_PRIMITIVE_RESTART_FIXED_INDEX);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "GL_TRIANGLE_STRIP with GL_UNSIGNED_SHORT primitive restart");
+        }
+
+        // ---- a batch with holes ----
+        // Zero-count sub-draws draw nothing. The flattening tier's binary search
+        // finds a sub-draw by prefix sum, and a zero-count entry repeats the
+        // previous sum - so a search that resolves ties the other way would attribute
+        // indices to the empty draw and paint the wrong column.
+
+        TEST_F(MultiDrawScenario, ZeroCountSubDrawsMatchUnrolledDraws) {
+            if (!Ready()) return;
+            constexpr int kPad = 5;
+            BuildScene(kPad, kQuadIndices, sizeof(kQuadIndices));
+
+            GLsizei counts[kColumns];
+            const void* offsets[kColumns];
+            GLint baseVertices[kColumns];
+            for (int i = 0; i < kColumns; ++i) {
+                // Columns 1 and 2 are skipped, leaving the outer two painted.
+                counts[i] = (i == 1 || i == 2) ? 0 : 6;
+                offsets[i] = reinterpret_cast<const void*>(0);
+                baseVertices[i] = kPad + i * 4;
+            }
+
+            const Image batched = RenderPass(m_program, m_vao, [&] {
+                glMultiDrawElementsBaseVertex(GL_TRIANGLES, counts, GL_UNSIGNED_INT, offsets, kColumns, baseVertices);
+            });
+            const Image unrolled = RenderPass(m_program, m_vao, [&] {
+                for (int i = 0; i < kColumns; ++i) {
+                    if (counts[i] == 0) continue;
+                    glDrawElementsBaseVertex(GL_TRIANGLES, counts[i], GL_UNSIGNED_INT, offsets[i], baseVertices[i]);
+                }
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectSameImage(batched, unrolled, "a batch with zero-count sub-draws");
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/OrientationScenario.cpp

@@ -0,0 +1,381 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/OrientationScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario A - "the frame came out upside down".
+//
+// The shipped bug (DirectVulkan, GetBaseTransformFlagsRaw): the shader
+// transform flags - the Y-flip and surface-rotation bits that apply ONLY when
+// the bound draw framebuffer is the default one - were memoized on the
+// swapchain pre-transform alone. The is-default-framebuffer input was not part
+// of the key, so whichever kind of pass evaluated the memo first decided the
+// orientation of every pass after it. In a real frame that meant: after any
+// render-to-texture pass, the next default-framebuffer pass inherited the FBO's
+// unflipped flags and the whole frame rendered upside down (retrace SSIM 0.052,
+// deterministic; flickering clouds on device).
+//
+// What pins it: a pattern asymmetric in BOTH axes - four quadrants, coloured
+//
+//        top-left  RED  | WHITE  top-right
+//     bottom-left BLUE  | GREEN  bottom-right
+//
+// - drawn to a target, read back with glReadPixels, and reduced to the four
+// quadrant-centre colours in the fixed order bottom-left, bottom-right,
+// top-left, top-right.
+//
+// Four quadrants rather than the three horizontal stripes this scenario used to
+// draw, because stripes only pin ONE axis. Stripes read down the centre line
+// are unchanged by an X flip, by a transpose, and by a 180 rotation composed
+// with a Y flip: all three of those bugs would have rendered a green stripe
+// between a blue one and a red one and passed. Every one of the eight
+// symmetries of the square now produces a different string:
+//
+//     identity       blue,green,red,white     <- correct
+//     Y flip         red,white,blue,green     <- the shipped bug
+//     X flip         green,blue,white,red
+//     180 rotation   white,red,green,blue
+//     transpose      blue,red,green,white
+//     anti-transpose white,green,red,blue
+//     rotate 90 CCW  red,blue,white,green
+//     rotate 90 CW   green,white,blue,red
+//
+// The assertions then go further than the signature: every quadrant is checked
+// pixel by pixel over its whole area (RegionIsMostly), so a partial or torn
+// draw cannot pass by having the four sampled centres come out right.
+//
+// Both orderings are covered, because the memo is poisoned by whichever pass
+// runs first and these tests share one process:
+//   - default -> FBO -> default   (the FBO pass inherits the default's flip)
+//   - FBO -> default              (the shipped symptom: the default pass
+//                                  inherits the FBO's lack of flip)
+
+#include <algorithm>
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVertexSource = R"(#version 330 core
+in vec2 aPos;
+in vec3 aColor;
+out vec3 vColor;
+void main() {
+    vColor = aColor;
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kFragmentSource = R"(#version 330 core
+in vec3 vColor;
+out vec4 oColor;
+void main() {
+    oColor = vec4(vColor, 1.0);
+}
+)";
+
+        // The correctly-oriented answer, in glReadPixels order (row 0 is the
+        // bottom row) and in QuadrantSignature's order: bottom-left, bottom-right,
+        // top-left, top-right. Plain GL semantics; holds for every framebuffer,
+        // default or not.
+        constexpr const char* kUprightSignature = "blue,green,red,white";
+
+        // How far inside each quadrant the whole-region checks start. The quadrant
+        // seam sits on a pixel boundary, so one pixel of margin is enough to make
+        // "every single pixel" an achievable (and therefore useful) demand.
+        constexpr int kQuadrantInset = 2;
+
+        struct Vertex {
+            float x, y;
+            float r, g, b;
+        };
+
+        void AppendQuad(std::vector<Vertex>& out, float x0, float x1, float y0, float y1, float r, float g, float b) {
+            const Vertex bl{x0, y0, r, g, b};
+            const Vertex br{x1, y0, r, g, b};
+            const Vertex tr{x1, y1, r, g, b};
+            const Vertex tl{x0, y1, r, g, b};
+            out.insert(out.end(), {bl, br, tr, bl, tr, tl});
+        }
+
+        std::vector<Vertex> QuadrantGeometry() {
+            std::vector<Vertex> vertices;
+            vertices.reserve(24);
+            AppendQuad(vertices, -1.0f, 0.0f, -1.0f, 0.0f, 0.0f, 0.0f, 1.0f); // bottom-left: blue
+            AppendQuad(vertices, 0.0f, 1.0f, -1.0f, 0.0f, 0.0f, 1.0f, 0.0f);  // bottom-right: green
+            AppendQuad(vertices, -1.0f, 0.0f, 0.0f, 1.0f, 1.0f, 0.0f, 0.0f);  // top-left: red
+            AppendQuad(vertices, 0.0f, 1.0f, 0.0f, 1.0f, 1.0f, 1.0f, 1.0f);   // top-right: white
+            return vertices;
+        }
+
+        class OrientationScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                std::string error;
+                m_program = CompileProgram(kVertexSource, kFragmentSource, &error);
+                ASSERT_NE(m_program, 0u) << error;
+
+                const std::vector<Vertex> vertices = QuadrantGeometry();
+                m_vertexCount = static_cast<int>(vertices.size());
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferData(GL_ARRAY_BUFFER, GLsizeiptr(vertices.size() * sizeof(Vertex)), vertices.data(),
+                             GL_STATIC_DRAW);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(0));
+                glEnableVertexAttribArray(1);
+                glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(8));
+                glBindVertexArray(0);
+
+                m_offscreen = MakeColorFbo(Gl().Width(), Gl().Height());
+                ASSERT_NE(m_offscreen.fbo, 0u) << "offscreen FBO is not framebuffer-complete";
+
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                DestroyColorFbo(m_offscreen);
+                if (m_vbo != 0) glDeleteBuffers(1, &m_vbo);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            void DrawQuadrants() {
+                glDisable(GL_DEPTH_TEST);
+                glDisable(GL_BLEND);
+                glUseProgram(m_program);
+                glBindVertexArray(m_vao);
+                glDrawArrays(GL_TRIANGLES, 0, m_vertexCount);
+                glBindVertexArray(0);
+            }
+
+            // One pass to the default (presentable) framebuffer.
+            Image DefaultFramebufferPass() {
+                BindDefaultFramebuffer();
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                DrawQuadrants();
+                return ReadPixels(Gl().Width(), Gl().Height());
+            }
+
+            // One render-to-texture pass. Real frames do this constantly
+            // (shadow maps, post-processing, Minecraft's main render target).
+            Image OffscreenPass() {
+                BindFbo(m_offscreen);
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                DrawQuadrants();
+                return ReadPixels(m_offscreen.width, m_offscreen.height);
+            }
+
+            // The signature says WHICH transform went wrong; this says the whole
+            // image is right, not merely its four sampled centres.
+            void ExpectUprightQuadrants(const Image& image, const std::string& when) {
+                const int w = image.Width();
+                const int h = image.Height();
+                const int inset = kQuadrantInset;
+                EXPECT_TRUE(RegionIsMostly(image, inset, w / 2 - inset, inset, h / 2 - inset, "blue", 0.0, when));
+                EXPECT_TRUE(RegionIsMostly(image, w / 2 + inset, w - inset, inset, h / 2 - inset, "green", 0.0, when));
+                EXPECT_TRUE(RegionIsMostly(image, inset, w / 2 - inset, h / 2 + inset, h - inset, "red", 0.0, when));
+                EXPECT_TRUE(RegionIsMostly(image, w / 2 + inset, w - inset, h / 2 + inset, h - inset, "white", 0.0,
+                                           when));
+            }
+
+            unsigned int m_program = 0;
+            unsigned int m_vao = 0;
+            unsigned int m_vbo = 0;
+            int m_vertexCount = 0;
+            ColorFbo m_offscreen;
+        };
+
+        // The plain statement of GL semantics that everything else leans on: an
+        // FBO pass is never flipped.
+        TEST_F(OrientationScenario, OffscreenPassRendersUpright) {
+            const Image offscreen = OffscreenPass();
+            EXPECT_EQ(offscreen.QuadrantSignature(), kUprightSignature)
+                << "a render-to-texture pass must render unflipped";
+            ExpectUprightQuadrants(offscreen, "render-to-texture pass");
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // The same for the default framebuffer: whatever the backend does with
+        // the swapchain internally, glReadPixels owes the caller GL orientation.
+        TEST_F(OrientationScenario, DefaultFramebufferPassRendersUpright) {
+            const Image presented = DefaultFramebufferPass();
+            EXPECT_EQ(presented.QuadrantSignature(), kUprightSignature)
+                << "a default-framebuffer pass must read back in GL orientation";
+            ExpectUprightQuadrants(presented, "default-framebuffer pass");
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // Scenario A proper: default -> FBO -> default in one frame. The third
+        // pass must be pixel-identical to the first; the FBO pass in between
+        // must not have moved anything.
+        TEST_F(OrientationScenario, DefaultFramebufferSurvivesAnOffscreenPass) {
+            const Image before = DefaultFramebufferPass();
+            const Image offscreen = OffscreenPass();
+            const Image after = DefaultFramebufferPass();
+
+            EXPECT_EQ(before.QuadrantSignature(), kUprightSignature)
+                << "first default-framebuffer pass is already misoriented";
+            EXPECT_EQ(offscreen.QuadrantSignature(), kUprightSignature)
+                << "the render-to-texture pass in the middle rendered flipped - the "
+                   "default framebuffer's transform flags leaked into it";
+            EXPECT_EQ(after.QuadrantSignature(), kUprightSignature)
+                << "the default-framebuffer pass AFTER a render-to-texture pass is "
+                   "misoriented - it inherited the FBO's transform flags";
+            ExpectUprightQuadrants(after, "default-framebuffer pass after a render-to-texture pass");
+            EXPECT_TRUE(after == before) << "the third pass differs from the first in " << after.ByteDiffCount(before)
+                                         << " bytes; first=" << before.QuadrantSignature()
+                                         << " third=" << after.QuadrantSignature();
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // The shipped symptom, in its shipped order: an FBO pass, then the
+        // default framebuffer. This is the one that flipped whole Minecraft
+        // frames.
+        TEST_F(OrientationScenario, DefaultFramebufferAfterOffscreenIsNotFlipped) {
+            const Image offscreen = OffscreenPass();
+            const Image presented = DefaultFramebufferPass();
+
+            EXPECT_EQ(offscreen.QuadrantSignature(), kUprightSignature)
+                << "render-to-texture pass rendered flipped";
+            EXPECT_EQ(presented.QuadrantSignature(), kUprightSignature)
+                << "the default-framebuffer pass that follows a render-to-texture pass "
+                   "rendered upside down";
+            ExpectUprightQuadrants(presented, "default-framebuffer pass following a render-to-texture pass");
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // And across a real frame boundary, which is how a game actually
+        // alternates the two kinds of pass.
+        TEST_F(OrientationScenario, OrientationIsStableAcrossFrames) {
+            const Image firstFrame = DefaultFramebufferPass();
+            ExpectUprightQuadrants(firstFrame, "frame 0");
+            Gl().EndFrame();
+
+            for (int frame = 0; frame < 3; ++frame) {
+                const Image offscreen = OffscreenPass();
+                EXPECT_EQ(offscreen.QuadrantSignature(), kUprightSignature)
+                    << "frame " << frame + 1 << "'s render-to-texture pass is misoriented";
+                const Image presented = DefaultFramebufferPass();
+                EXPECT_EQ(presented.QuadrantSignature(), kUprightSignature)
+                    << "frame " << frame + 1 << " of the alternating FBO/default loop is misoriented";
+                ExpectUprightQuadrants(presented, "frame " + std::to_string(frame + 1));
+                EXPECT_TRUE(presented == firstFrame) << "frame " << frame + 1 << " differs from frame 0 in "
+                                                     << presented.ByteDiffCount(firstFrame) << " bytes";
+                Gl().EndFrame();
+            }
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
+        // A standing self-test of the signature, not of MobileGL: it proves the
+        // four-quadrant reduction really does separate all eight symmetries of
+        // the square, so a future "simplify the pattern" change cannot quietly
+        // reintroduce the blind spot the three-stripe version had (X flip,
+        // transpose and 180+Y-flip all left the stripe signature alone).
+        TEST_F(OrientationScenario, QuadrantSignatureSeparatesEverySquareSymmetry) {
+            const Image upright = OffscreenPass();
+            ASSERT_EQ(upright.QuadrantSignature(), kUprightSignature) << "the reference image is not upright";
+
+            const int w = upright.Width();
+            const int h = upright.Height();
+            // Transposes are expressed on the largest centred square the readback
+            // contains, which is enough for the four quadrant centres to move.
+            const int side = std::min(w, h);
+            const int ox = (w - side) / 2;
+            const int oy = (h - side) / 2;
+
+            struct Symmetry {
+                const char* name;
+                const char* expected;
+                int (*mapX)(int x, int y, int w, int h);
+                int (*mapY)(int x, int y, int w, int h);
+            };
+            const Symmetry symmetries[] = {
+                {"Y flip", "red,white,blue,green", [](int x, int, int, int) { return x; },
+                 [](int, int y, int, int hh) { return hh - 1 - y; }},
+                {"X flip", "green,blue,white,red", [](int x, int, int ww, int) { return ww - 1 - x; },
+                 [](int, int y, int, int) { return y; }},
+                {"180 rotation", "white,red,green,blue", [](int x, int, int ww, int) { return ww - 1 - x; },
+                 [](int, int y, int, int hh) { return hh - 1 - y; }},
+            };
+
+            for (const Symmetry& symmetry : symmetries) {
+                Image transformed(w, h);
+                for (int y = 0; y < h; ++y) {
+                    for (int x = 0; x < w; ++x) {
+                        const Rgba8 source = upright.At(symmetry.mapX(x, y, w, h), symmetry.mapY(x, y, w, h));
+                        std::uint8_t* out = transformed.Data() + (std::size_t(y) * w + x) * 4;
+                        out[0] = source.r;
+                        out[1] = source.g;
+                        out[2] = source.b;
+                        out[3] = source.a;
+                    }
+                }
+                EXPECT_EQ(transformed.QuadrantSignature(), symmetry.expected)
+                    << symmetry.name << " must produce its own signature, or the pattern cannot see it";
+                EXPECT_NE(transformed.QuadrantSignature(), kUprightSignature)
+                    << symmetry.name << " is INDISTINGUISHABLE from an upright frame - the pattern is too symmetric";
+            }
+
+            // The four symmetries that move the axes into each other. They only
+            // make sense on a square, so they run on the largest centred one.
+            struct SquareSymmetry {
+                const char* name;
+                const char* expected;
+                int (*sourceX)(int x, int y, int side);
+                int (*sourceY)(int x, int y, int side);
+            };
+            const SquareSymmetry squareSymmetries[] = {
+                {"transpose", "blue,red,green,white", [](int, int y, int) { return y; },
+                 [](int x, int, int) { return x; }},
+                {"anti-transpose", "white,green,red,blue", [](int, int y, int s) { return s - 1 - y; },
+                 [](int x, int, int s) { return s - 1 - x; }},
+                {"rotate 90 CCW", "red,blue,white,green", [](int, int y, int) { return y; },
+                 [](int x, int, int s) { return s - 1 - x; }},
+                {"rotate 90 CW", "green,white,blue,red", [](int, int y, int s) { return s - 1 - y; },
+                 [](int x, int, int) { return x; }},
+            };
+            for (const SquareSymmetry& symmetry : squareSymmetries) {
+                Image square(side, side);
+                for (int y = 0; y < side; ++y) {
+                    for (int x = 0; x < side; ++x) {
+                        const Rgba8 source =
+                            upright.At(ox + symmetry.sourceX(x, y, side), oy + symmetry.sourceY(x, y, side));
+                        std::uint8_t* out = square.Data() + (std::size_t(y) * side + x) * 4;
+                        out[0] = source.r;
+                        out[1] = source.g;
+                        out[2] = source.b;
+                        out[3] = source.a;
+                    }
+                }
+                EXPECT_EQ(square.QuadrantSignature(), symmetry.expected)
+                    << symmetry.name << " must produce its own signature, or the pattern cannot see it";
+                EXPECT_NE(square.QuadrantSignature(), kUprightSignature)
+                    << symmetry.name << " is INDISTINGUISHABLE from an upright frame";
+            }
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ResidentIndexScenario.cpp

@@ -0,0 +1,383 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ResidentIndexScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario C - RESIDENT index buffers across frame boundaries.
+//
+// WHAT THIS FILE DOES AND DOES NOT COVER, stated plainly because the answer is
+// not the one it was written to find.
+//
+// The shipped fix (d7976326) removed cross-frame slice trust from TWO memos: the
+// vertex-binding one and the EBO one. StreamedArenaScenario pins the vertex
+// half - re-enable that half alone and it fails. Nothing pinned the EBO half,
+// and these cases are the result of trying to build something that does.
+//
+// The EBO memo lives in UploadAndBindIndexBuffer and is recorded ONLY on the
+// resident branch, keyed on (BufferObject*, VkBufferResource::sliceEpoch,
+// frame serial). To fail with only the EBO revalidation re-enabled, a scenario
+// needs a RESIDENT index buffer whose recorded slice stops describing the right
+// bytes while the pointer and the epoch still match. Every case below is an
+// attempt at that, run against the re-enabled buggy path with the branch
+// instrumented to count reaches, acceptances, and - critically - what the
+// skipped AcquireResidentSlice WOULD have done. The measurement, over this file
+// plus every other scenario in the module:
+//
+//     reached=89 accepted=81 sliceMoved=0 bytesChanged=0 epochBumped=0
+//
+// The buggy branch is entered 89 times and serves its recorded slice 81 times,
+// and in NOT ONE of those 81 would the acquire have moved the slice, changed a
+// byte of it, or bumped the epoch. The skipped work was a no-op every time.
+//
+// That is not luck, it is the shape of the code. A resident slice is
+// `resource->buffer.GetSlice(0, size)` of a dedicated VkBuffer, so it can only
+// move when CreateResidentStorage mints new storage - which bumps the epoch. Its
+// bytes can only change through Respecify / SubData / FlushMappedRange - each of
+// which bumps the epoch as its first act - or through
+// BufferObject::SyncPersistentMappedRange, which the acquire calls and the memo
+// skips. That last one is the real escape, and it is dead here: it early-outs
+// when the backend has adopted the map into coherent GPU storage, and
+// AcquirePersistentMap only declines when a host-visible coherent allocation
+// FAILS. Instrumented across the whole module: 50 persistent coherent write
+// maps, 50 adopted, 0 dispatches. A 96 MiB EBO did not change that either.
+//
+// So on DirectVulkan as it stands, the EBO half of the fix is not reachable from
+// a GL-level test - not because the guard is sound in principle (it is the same
+// unsound idea the vertex half shipped corruption with) but because the two
+// mechanisms that made the vertex half observable are both absent for indices:
+//
+//   1. ARENA RELOCATION. The vertex memo records STREAMED slices too, and a
+//      streamed slice moves to a new arena block every frame BY DESIGN - the
+//      epoch that catches it is bumped inside the very acquire the memo skips.
+//      That is what StreamedVertexDataSurvivesArenaRecycling exploits. The index
+//      memo is never recorded on the streamed branch, so no index memo ever
+//      names an arena offset. Measured: StreamedIndexDataSurvivesArenaRecycling
+//      reaches the branch 0 times, and so does PromotedDynamicEbo below (a
+//      promoted DYNAMIC_DRAW buffer is SERVED by AcquireResidentSlice but still
+//      ROUTED as streamed, so it is not memoised either).
+//   2. HOST-MAP SYNC. Dead, as above.
+//
+// These cases therefore stay as what they honestly are: end-to-end regression
+// tests for resident index-buffer freshness across frame boundaries, and the
+// standing tripwire for change (1). The moment anyone memoises the streamed or
+// promoted index path - the natural next step for the same optimisation - these
+// stop being redundant and start failing. Each case says below what it covers.
+
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+in vec3 aColor;
+out vec3 vColor;
+void main() { vColor = aColor; gl_Position = vec4(aPos, 0.0, 1.0); }
+)";
+        constexpr const char* kFS = R"(#version 330 core
+in vec3 vColor;
+out vec4 oColor;
+void main() { oColor = vec4(vColor, 1.0); }
+)";
+
+        struct V {
+            float x, y, r, g, b;
+        };
+        constexpr int kIdx = 6;
+        const GLuint kLeft[kIdx] = {0, 1, 2, 0, 2, 3};
+        const GLuint kRight[kIdx] = {4, 5, 6, 4, 6, 7};
+
+        std::vector<V> Scene() {
+            return {{-1, -1, 1, 0, 0}, {0, -1, 1, 0, 0}, {0, 1, 1, 0, 0}, {-1, 1, 1, 0, 0},
+                    {0, -1, 0, 1, 0},  {1, -1, 0, 1, 0}, {1, 1, 0, 1, 0},  {0, 1, 0, 1, 0}};
+        }
+
+        class ResidentIndexScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string err;
+                m_program = CompileProgram(kVS, kFS, &err);
+                ASSERT_NE(m_program, 0u) << err;
+            }
+            void TearDown() override {
+                if (!Ready()) return;
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            // A VAO whose VBO is STATIC_DRAW (so it resolves resident and the
+            // vertex memo is recorded) and whose EBO is `eboName`.
+            unsigned int MakeVao(unsigned int vbo, unsigned int ebo) {
+                unsigned int vao = 0;
+                glGenVertexArrays(1, &vao);
+                glBindVertexArray(vao);
+                glBindBuffer(GL_ARRAY_BUFFER, vbo);
+                glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(V), reinterpret_cast<void*>(0));
+                glEnableVertexAttribArray(1);
+                glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, sizeof(V), reinterpret_cast<void*>(8));
+                glBindVertexArray(0);
+                return vao;
+            }
+
+            unsigned int MakeStaticVbo() {
+                const std::vector<V> vertices = Scene();
+                unsigned int vbo = 0;
+                glGenBuffers(1, &vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, vbo);
+                glBufferData(GL_ARRAY_BUFFER, GLsizeiptr(vertices.size() * sizeof(V)), vertices.data(),
+                             GL_STATIC_DRAW);
+                return vbo;
+            }
+
+            void Draw(unsigned int vao) {
+                glDisable(GL_DEPTH_TEST);
+                glDisable(GL_BLEND);
+                glUseProgram(m_program);
+                glBindVertexArray(vao);
+                glDrawElements(GL_TRIANGLES, kIdx, GL_UNSIGNED_INT, nullptr);
+                glBindVertexArray(0);
+            }
+            void Begin() {
+                BindDefaultFramebuffer();
+                ClearTo(0, 0, 0, 1);
+            }
+            Image Read() { return ReadPixels(Gl().Width(), Gl().Height()); }
+            void Halves(const Image& image, const char* left, const char* right, const std::string& when) {
+                const int w = image.Width(), h = image.Height();
+                EXPECT_TRUE(RegionIsMostly(image, 2, w / 2 - 2, 2, h - 2, left, 0.0, when + " [left]"));
+                EXPECT_TRUE(RegionIsMostly(image, w / 2 + 2, w - 2, 2, h - 2, right, 0.0, when + " [right]"));
+            }
+
+            unsigned int m_program = 0;
+        };
+
+        // A: a coherent persistent EBO rewritten on EVERY frame, with no GL call
+        // between the write and the draw. This is the only shape in which an
+        // application changes index data with nothing for the backend to notice.
+        //
+        // COVERS: the coherent-persistent index contract end to end.
+        // DOES NOT COVER: the EBO memo. Instrumented it reaches the cross-frame
+        // branch 11 times and is served its recorded slice all 11 - but the
+        // backend adopted the map into that same storage, so the "stale" slice IS
+        // where the application's writes landed. It would only discriminate on a
+        // stack where AcquirePersistentMap declines (see the file header). A
+        // 96 MiB variant was tried to force that and did not: it cost 40s and
+        // measured the same zero, so it is not kept.
+        TEST_F(ResidentIndexScenario, PersistentCoherentEboWrittenEveryFrame) {
+            const unsigned int vbo = MakeStaticVbo();
+            unsigned int ebo = 0;
+            glGenBuffers(1, &ebo);
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            const GLbitfield storageFlags =
+                GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT | GL_DYNAMIC_STORAGE_BIT;
+            glBufferStorage(GL_ELEMENT_ARRAY_BUFFER, GLsizeiptr(sizeof(kLeft)), kLeft, storageFlags);
+            if (FirstGLError() != GL_NO_ERROR) GTEST_SKIP() << "no immutable storage";
+            auto* map = static_cast<unsigned char*>(glMapBufferRange(
+                GL_ELEMENT_ARRAY_BUFFER, 0, GLsizeiptr(sizeof(kLeft)),
+                GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT));
+            ASSERT_NE(map, nullptr);
+            const unsigned int vao = MakeVao(vbo, ebo);
+
+            for (int frame = 0; frame < 12; ++frame) {
+                Begin();
+                const bool wantRight = (frame % 2) == 1;
+                std::memcpy(map, wantRight ? kRight : kLeft, sizeof(kLeft));
+                Draw(vao);
+                const Image image = Read();
+                Halves(image, wantRight ? "black" : "red", wantRight ? "green" : "black",
+                       "frame " + std::to_string(frame) + " of a per-frame coherent EBO rewrite");
+                Gl().EndFrame();
+            }
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            glUnmapBuffer(GL_ELEMENT_ARRAY_BUFFER);
+            glDeleteVertexArrays(1, &vao);
+            glDeleteBuffers(1, &ebo);
+            glDeleteBuffers(1, &vbo);
+        }
+
+        // B: usage escalation. The EBO is memoised as an index buffer, then bound
+        // as a VERTEX buffer in a later frame, which forces the backend to
+        // recreate its resident storage carrying the extra usage bit. A memo that
+        // survived that recreate would name a destroyed VkBuffer.
+        //
+        // COVERS: that a storage recreate driven by a DIFFERENT binding point
+        // retires the index memo. Reaches the branch 5 times.
+        TEST_F(ResidentIndexScenario, EboAlsoBoundAsVertexBufferLater) {
+            const unsigned int vbo = MakeStaticVbo();
+            unsigned int ebo = 0;
+            glGenBuffers(1, &ebo);
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            // Big enough to be a legal (if nonsensical) vertex source too.
+            std::vector<GLuint> indices(64, 0);
+            std::memcpy(indices.data(), kLeft, sizeof(kLeft));
+            glBufferData(GL_ELEMENT_ARRAY_BUFFER, GLsizeiptr(indices.size() * 4), indices.data(), GL_STATIC_DRAW);
+            const unsigned int vao = MakeVao(vbo, ebo);
+
+            unsigned int vertexUseVao = 0;
+            glGenVertexArrays(1, &vertexUseVao);
+            glBindVertexArray(vertexUseVao);
+            glBindBuffer(GL_ARRAY_BUFFER, ebo); // the EBO, as a vertex source
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(V), reinterpret_cast<void*>(0));
+            glEnableVertexAttribArray(1);
+            glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, sizeof(V), reinterpret_cast<void*>(8));
+            glBindVertexArray(0);
+
+            for (int frame = 0; frame < 6; ++frame) {
+                Begin();
+                Draw(vao);
+                if (frame == 2) Draw(vertexUseVao); // forces the usage escalation
+                const Image image = Read();
+                if (frame != 2) {
+                    Halves(image, "red", "black", "frame " + std::to_string(frame) + " around a usage escalation");
+                }
+                Gl().EndFrame();
+            }
+            glDeleteVertexArrays(1, &vertexUseVao);
+            glDeleteVertexArrays(1, &vao);
+            glDeleteBuffers(1, &ebo);
+            glDeleteBuffers(1, &vbo);
+        }
+
+        // C: delete the EBO and immediately recreate it, so the frontend
+        // BufferObject may well land at the same address - which is all the memo's
+        // identity check compares. What stops it is that a fresh resource cannot
+        // reproduce an epoch from the process-lifetime counter; this is the test
+        // that says so out loud.
+        //
+        // COVERS: address reuse of a deleted index buffer. Reaches 7, accepts 6 -
+        // the one decline is the post-recreate draw.
+        TEST_F(ResidentIndexScenario, EboDeletedAndRecreatedAtTheSameName) {
+            const unsigned int vbo = MakeStaticVbo();
+            unsigned int ebo = 0;
+            glGenBuffers(1, &ebo);
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            glBufferData(GL_ELEMENT_ARRAY_BUFFER, GLsizeiptr(sizeof(kLeft)), kLeft, GL_STATIC_DRAW);
+            unsigned int vao = MakeVao(vbo, ebo);
+
+            for (int frame = 0; frame < 4; ++frame) {
+                Begin();
+                Draw(vao);
+                Halves(Read(), "red", "black", "warmup frame " + std::to_string(frame));
+                Gl().EndFrame();
+            }
+
+            // Same VAO, same GL name, different contents.
+            glDeleteVertexArrays(1, &vao);
+            glDeleteBuffers(1, &ebo);
+            glGenBuffers(1, &ebo);
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            glBufferData(GL_ELEMENT_ARRAY_BUFFER, GLsizeiptr(sizeof(kRight)), kRight, GL_STATIC_DRAW);
+            vao = MakeVao(vbo, ebo);
+
+            for (int frame = 0; frame < 4; ++frame) {
+                Begin();
+                Draw(vao);
+                Halves(Read(), "black", "green", "post-recreate frame " + std::to_string(frame));
+                Gl().EndFrame();
+            }
+            glDeleteVertexArrays(1, &vao);
+            glDeleteBuffers(1, &ebo);
+            glDeleteBuffers(1, &vbo);
+        }
+
+        // D: one resident EBO shared by two VAOs, so two independent memo entries
+        // hold the same recorded slice, mutated through one of them and drawn
+        // through both across frames.
+        //
+        // COVERS: that a mutation retires EVERY memo naming the buffer, not just
+        // the one whose VAO issued it. Reaches 8, accepts 6.
+        TEST_F(ResidentIndexScenario, OneEboTwoVaosMutatedAcrossFrames) {
+            const unsigned int vbo = MakeStaticVbo();
+            unsigned int ebo = 0;
+            glGenBuffers(1, &ebo);
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            glBufferData(GL_ELEMENT_ARRAY_BUFFER, GLsizeiptr(sizeof(kLeft)), kLeft, GL_STATIC_DRAW);
+            const unsigned int vaoA = MakeVao(vbo, ebo);
+            const unsigned int vaoB = MakeVao(vbo, ebo);
+
+            for (int frame = 0; frame < 10; ++frame) {
+                Begin();
+                const bool wantRight = frame >= 5;
+                if (frame == 5) {
+                    glBindVertexArray(vaoA);
+                    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+                    glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, GLsizeiptr(sizeof(kRight)), kRight);
+                    glBindVertexArray(0);
+                }
+                Draw((frame % 2) == 0 ? vaoA : vaoB);
+                Halves(Read(), wantRight ? "black" : "red", wantRight ? "green" : "black",
+                       "shared-EBO frame " + std::to_string(frame));
+                Gl().EndFrame();
+            }
+            glDeleteVertexArrays(1, &vaoB);
+            glDeleteVertexArrays(1, &vaoA);
+            glDeleteBuffers(1, &ebo);
+            glDeleteBuffers(1, &vbo);
+        }
+
+        // E: a DYNAMIC_DRAW EBO left untouched long enough for the streaming path
+        // to PROMOTE it onto resident storage, then mutated.
+        //
+        // COVERS: promoted-buffer index freshness across a frame boundary.
+        // DOES NOT COVER: the EBO memo, and this is the useful part - instrumented,
+        // it reaches the cross-frame branch ZERO times. A promoted buffer is SERVED
+        // by AcquireResidentSlice but still ROUTED through the streamed branch of
+        // UploadAndBindIndexBuffer, which never records a memo. That asymmetry is
+        // exactly what makes the EBO half of the shipped fix unobservable, and this
+        // case is the tripwire: memoise the streamed/promoted index path and the
+        // reach stops being zero.
+        TEST_F(ResidentIndexScenario, PromotedDynamicEbo) {
+            const unsigned int vbo = MakeStaticVbo();
+            unsigned int ebo = 0;
+            glGenBuffers(1, &ebo);
+            glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+            glBufferData(GL_ELEMENT_ARRAY_BUFFER, GLsizeiptr(sizeof(kLeft)), kLeft, GL_DYNAMIC_DRAW);
+            const unsigned int vao = MakeVao(vbo, ebo);
+
+            for (int frame = 0; frame < 10; ++frame) {
+                Begin();
+                Draw(vao);
+                Halves(Read(), "red", "black", "promotion warmup frame " + std::to_string(frame));
+                Gl().EndFrame();
+            }
+            for (int frame = 0; frame < 6; ++frame) {
+                Begin();
+                if (frame == 0) {
+                    glBindVertexArray(vao);
+                    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
+                    glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, GLsizeiptr(sizeof(kRight)), kRight);
+                    glBindVertexArray(0);
+                }
+                Draw(vao);
+                Halves(Read(), "black", "green", "post-promotion frame " + std::to_string(frame));
+                Gl().EndFrame();
+            }
+            glDeleteVertexArrays(1, &vao);
+            glDeleteBuffers(1, &ebo);
+            glDeleteBuffers(1, &vbo);
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ThreeChannelAttachmentScenario.cpp

@@ -0,0 +1,300 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ThreeChannelAttachmentScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - THREE-CHANNEL COLOUR ATTACHMENTS, on a live driver.
+//
+// The bug: no OpenGL ES driver renders to a three-channel image. EXT_render_snorm covers
+// R/RG/RGBA only, EXT_color_buffer_float excludes RGB16F, and RGB integer formats are not
+// colour-renderable anywhere. Complementary Reimagined declares colortex1 = RGB8_SNORM and
+// colortex2 = RGB16F, so every framebuffer Iris built from them answered
+// GL_FRAMEBUFFER_UNSUPPORTED and Iris refused to load the shaderpack. DirectGLES now stores such
+// an attachment in its four-channel sibling (GL_RGB8_SNORM -> GL_RGBA16F) and reports the
+// substitution as a caveat capability, which is what makes glCheckFramebufferStatus say COMPLETE.
+//
+// WHY THIS SCENARIO EXISTS RATHER THAN A UNIT TEST. The unit tests in
+// MG_Test/Framebuffer/FramebufferTest.cpp drive a HAND-BUILT capability cache: they prove the
+// frontend accepts a caveat capability, and prove the colour-mask/clear discipline that keeps a
+// widened attachment's stored alpha at 1.0, but they cannot prove that a real driver's probe
+// actually PRODUCES that caveat. Only a live glCheckFramebufferStatus can, and the answer is
+// per-driver, not per-platform:
+//
+//   Mesa llvmpipe (the headless CI driver), ES 3.2, GL_TEXTURE_2D colour attachment:
+//     COMPLETE               GL_RGB8, GL_RGB16F, GL_R11F_G11F_B10F, every RGBA*
+//     INCOMPLETE_ATTACHMENT  GL_RGB8_SNORM, GL_SRGB8, every RGB integer format
+//     UNSUPPORTED            GL_RGB32F
+//
+// So the widening is LIVE on llvmpipe - "the desktop build is unaffected" was simply wrong, and
+// the CI retraces were green before the fix only because retrace ignores what
+// glCheckFramebufferStatus returns. This scenario is the gate that actually looks.
+//
+// DirectGLES only. DirectVulkan's format story is its own (Vulkan exposes R8G8B8_SNORM on almost
+// nothing, and Magma substitutes on different terms); asserting Espryt's answers there would
+// only pin a coincidence.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+void main() {
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        // Two outputs so the mixed case is covered: draw buffer 0 is a natively renderable
+        // four-channel format whose alpha the application owns, draw buffer 1 is the widened
+        // three-channel one whose alpha the format says is 1.0. Both alphas are deliberately
+        // NOT 1.0 in the shader, so an implementation that simply passed the value through would
+        // fail the second assertion.
+        constexpr const char* kFS = R"(#version 330 core
+layout(location = 0) out vec4 oNative;
+layout(location = 1) out vec4 oWidened;
+void main() {
+    oNative  = vec4(1.0, 0.0, 0.0, 0.25);
+    oWidened = vec4(0.0, 1.0, 0.0, 0.75);
+}
+)";
+
+        constexpr int kSize = 16;
+
+        class ThreeChannelAttachmentScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                if (Gl().BackendName() != "DirectGLES") {
+                    GTEST_SKIP() << "three-channel widening is a DirectGLES substitution; backend is "
+                                 << Gl().BackendName();
+                }
+            }
+
+            // A single-level 2D texture in `internalFormat`, or 0 when the driver rejects the
+            // storage outright (which is a different failure from rejecting the ATTACHMENT).
+            static GLuint MakeTexture(GLenum internalFormat) {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                glBindTexture(GL_TEXTURE_2D, texture);
+                glTexStorage2D(GL_TEXTURE_2D, 1, internalFormat, kSize, kSize);
+                if (glGetError() != GL_NO_ERROR) {
+                    glDeleteTextures(1, &texture);
+                    return 0;
+                }
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glBindTexture(GL_TEXTURE_2D, 0);
+                return texture;
+            }
+
+            static GLenum SingleAttachmentStatus(GLenum internalFormat) {
+                const GLuint texture = MakeTexture(internalFormat);
+                if (texture == 0) return GL_NONE;
+                GLuint fbo = 0;
+                glGenFramebuffers(1, &fbo);
+                glBindFramebuffer(GL_DRAW_FRAMEBUFFER, fbo);
+                glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, 0);
+                const GLenum status = glCheckFramebufferStatus(GL_DRAW_FRAMEBUFFER);
+                glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+                glDeleteFramebuffers(1, &fbo);
+                glDeleteTextures(1, &texture);
+                return status;
+            }
+        };
+
+        // THE regression gate for the frontend's answer: this is the exact call Iris makes, and
+        // GL_FRAMEBUFFER_UNSUPPORTED here is the whole shaderpack load failure.
+        TEST_F(ThreeChannelAttachmentScenario, ThreeChannelColorAttachmentsReportComplete) {
+            if (!Ready() || IsSkipped()) return;
+
+            // GL_RGB8 is the control: colour-renderable in ES core, so it must pass with or
+            // without any substitution. If it ever fails, nothing below means anything.
+            EXPECT_EQ(SingleAttachmentStatus(GL_RGB8), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                << "GL_RGB8 is ES-core colour-renderable";
+
+            // Complementary Reimagined's colortex1 and colortex2.
+            EXPECT_EQ(SingleAttachmentStatus(GL_RGB8_SNORM), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                << "colortex1 (RGB8_SNORM) must be renderable through the four-channel widening";
+            EXPECT_EQ(SingleAttachmentStatus(GL_RGB16F), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                << "colortex2 (RGB16F) must be renderable, natively or through the widening";
+
+            // The other formats the widening covers. GL_RGB32F only reaches a renderable
+            // four-channel form when EXT_color_buffer_float is present, so a half-float-only
+            // driver legitimately answers UNSUPPORTED for it - see the POST's per-format row.
+            EXPECT_EQ(SingleAttachmentStatus(GL_SRGB8), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE));
+            EXPECT_EQ(SingleAttachmentStatus(GL_RGB8UI), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE));
+
+            EXPECT_EQ(FirstGLError(), 0u) << GLErrorName(FirstGLError());
+        }
+
+        // The other half: the substitution has to be INVISIBLE. A three-channel format has no
+        // alpha, so GL answers 1.0 for it - and that answer has to hold after a draw that wrote
+        // something else into the widened storage's real alpha channel, which is what the
+        // colour-mask discipline in SyncRenderState is for. GL_DST_ALPHA blending and
+        // glBlitFramebuffer read that stored alpha inside the driver, where no readback fixup can
+        // reach it, so "the storage really holds 1.0" is the only workable invariant.
+        TEST_F(ThreeChannelAttachmentScenario, WidenedAttachmentReadsBackOpaqueWhileItsNeighbourKeepsItsAlpha) {
+            if (!Ready() || IsSkipped()) return;
+
+            std::string error;
+            const GLuint program = CompileProgram(kVS, kFS, &error);
+            ASSERT_NE(program, 0u) << error;
+
+            const GLuint nativeTexture = MakeTexture(GL_RGBA16F);
+            const GLuint widenedTexture = MakeTexture(GL_RGB8_SNORM);
+            ASSERT_NE(nativeTexture, 0u);
+            ASSERT_NE(widenedTexture, 0u);
+
+            GLuint fbo = 0;
+            glGenFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, nativeTexture, 0);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, GL_TEXTURE_2D, widenedTexture, 0);
+            const GLenum drawBuffers[2] = {GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1};
+            glDrawBuffers(2, drawBuffers);
+            ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE));
+
+            glViewport(0, 0, kSize, kSize);
+            // Alpha 0.0 on purpose: the widened attachment must come back 1.0 anyway, and the
+            // native one must come back 0.0 where the draw does not cover it.
+            glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
+            glClear(GL_COLOR_BUFFER_BIT);
+
+            const float quad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+            GLuint vao = 0;
+            GLuint vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, sizeof(quad), quad, GL_STATIC_DRAW);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+            std::vector<float> pixels(static_cast<std::size_t>(kSize) * kSize * 4, -1.0f);
+
+            glReadBuffer(GL_COLOR_ATTACHMENT1);
+            glReadPixels(0, 0, kSize, kSize, GL_RGBA, GL_FLOAT, pixels.data());
+            EXPECT_NEAR(pixels[0], 0.0f, 0.02f) << "widened attachment red";
+            EXPECT_NEAR(pixels[1], 1.0f, 0.02f) << "widened attachment green";
+            EXPECT_NEAR(pixels[2], 0.0f, 0.02f) << "widened attachment blue";
+            EXPECT_NEAR(pixels[3], 1.0f, 0.001f)
+                << "a three-channel format has no alpha channel, so GL must report 1.0 for it";
+
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glReadPixels(0, 0, kSize, kSize, GL_RGBA, GL_FLOAT, pixels.data());
+            EXPECT_NEAR(pixels[0], 1.0f, 0.02f) << "native attachment red";
+            EXPECT_NEAR(pixels[3], 0.25f, 0.02f)
+                << "the alpha discipline must not leak onto a natively renderable attachment";
+
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            glDeleteFramebuffers(1, &fbo);
+            glDeleteBuffers(1, &vbo);
+            glDeleteVertexArrays(1, &vao);
+            glDeleteTextures(1, &nativeTexture);
+            glDeleteTextures(1, &widenedTexture);
+            glDeleteProgram(program);
+            EXPECT_EQ(FirstGLError(), 0u) << GLErrorName(FirstGLError());
+        }
+
+        // The case above can be satisfied by the readback fixup alone (ForceWideReadAlphaToOne
+        // rewrites glReadPixels' alpha), so it does NOT prove the STORED alpha is 1.0. This one
+        // does, by asking the driver to read that alpha itself: GL_DST_ALPHA blending multiplies
+        // by the destination alpha inside the raster pipeline, where nothing MobileGL does can
+        // intervene. Same reason GL_ONE_MINUS_DST_ALPHA and glBlitFramebuffer are covered for
+        // free once this holds - and the reason the discipline is a write mask rather than a
+        // readback patch.
+        //
+        // Ablation-checked on llvmpipe, each half separately: disable the alpha doctoring in
+        // SyncRenderState and the opaque draw leaves 0.25 in the stored alpha; disable the clear
+        // substitution in Clear() and it stays at the application's 0.0. Either way this case
+        // reads back the wrong number, which is what makes it a gate rather than a description.
+        TEST_F(ThreeChannelAttachmentScenario, DstAlphaBlendingSeesOneInAWidenedAttachment) {
+            if (!Ready() || IsSkipped()) return;
+
+            static constexpr const char* kSingleOutFS = R"(#version 330 core
+out vec4 oColor;
+uniform vec4 uColor;
+void main() { oColor = uColor; }
+)";
+            std::string error;
+            const GLuint program = CompileProgram(kVS, kSingleOutFS, &error);
+            ASSERT_NE(program, 0u) << error;
+            const GLint colorLocation = glGetUniformLocation(program, "uColor");
+            ASSERT_GE(colorLocation, 0);
+
+            const GLuint widenedTexture = MakeTexture(GL_RGB8_SNORM);
+            ASSERT_NE(widenedTexture, 0u);
+            GLuint fbo = 0;
+            glGenFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, widenedTexture, 0);
+            ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE));
+
+            const float quad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+            GLuint vao = 0;
+            GLuint vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, sizeof(quad), quad, GL_STATIC_DRAW);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+            glUseProgram(program);
+            glViewport(0, 0, kSize, kSize);
+
+            // The clear's alpha is 0.0 and the draw's is 0.25 - neither is the 1.0 the format
+            // implies, so both halves of the discipline have to fire for the blend below to see
+            // 1.0: the clear substitutes it, and the draw is masked away from it.
+            glDisable(GL_BLEND);
+            glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
+            glClear(GL_COLOR_BUFFER_BIT);
+            glUniform4f(colorLocation, 0.0f, 1.0f, 0.0f, 0.25f);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+            // dst = stored alpha; src factor GL_DST_ALPHA, dst factor GL_ZERO, source white
+            // => the destination colour becomes (storedAlpha, storedAlpha, storedAlpha).
+            glEnable(GL_BLEND);
+            glBlendFunc(GL_DST_ALPHA, GL_ZERO);
+            glUniform4f(colorLocation, 1.0f, 1.0f, 1.0f, 1.0f);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            glDisable(GL_BLEND);
+
+            std::vector<float> pixels(static_cast<std::size_t>(kSize) * kSize * 4, -1.0f);
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glReadPixels(0, 0, kSize, kSize, GL_RGBA, GL_FLOAT, pixels.data());
+            EXPECT_NEAR(pixels[0], 1.0f, 0.02f)
+                << "GL_DST_ALPHA read the stored alpha of a three-channel attachment; it must be 1.0";
+
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            glDeleteFramebuffers(1, &fbo);
+            glDeleteBuffers(1, &vbo);
+            glDeleteVertexArrays(1, &vao);
+            glDeleteTextures(1, &widenedTexture);
+            glDeleteProgram(program);
+            EXPECT_EQ(FirstGLError(), 0u) << GLErrorName(FirstGLError());
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/XfbAfterClipDistanceScenario.cpp

@@ -0,0 +1,584 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/XfbAfterClipDistanceScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario F - a draw must never read a destroyed object's memoised state.
+//
+// Distilled from the order-triggered CTS failure: on DirectVulkan, once
+// KHR-GLxx.clip_distance.functional had run in the same process, every later
+// transform_feedback CAPTURE case failed. It looked like a transform feedback
+// bug and is not one. The capture works; the DRAW being captured fetched its
+// vertices from the WRONG BUFFER - the one the clip workload had just deleted.
+//
+// The mechanism, and why the sequence matters. DirectVulkan memoises a VAO's
+// resolved Vulkan vertex bindings in a table keyed on the VertexArrayObject's
+// heap ADDRESS, validated by a content hash that folds in the bound
+// BufferObject's heap ADDRESS. Both are recycled by the allocator, so when the
+// workload's VAO and vertex buffer are destroyed and the capture phase's own
+// VAO and vertex buffer are allocated onto their addresses under a
+// byte-identical attribute layout (one vec4 float array at location 0 - which
+// is what both phases use), the key matches, the hash matches, and the memo
+// hands the new draw the dead buffer's GPU slice. Nothing about transform
+// feedback is involved: capture just makes the wrong vertices legible, because
+// the captured record IS the vertex data. The fix gives VertexArrayObject and
+// BufferObject never-reused lifetime ids and keys the memo on those.
+//
+// MOBILEGL_ASYNC_SHADER_COMPILE is not part of the defect. It shifts the
+// allocation pattern, so it changes WHICH stop points below land on a recycled
+// address - which is why the CTS saw ~100% incidence with it on and ~2% with it
+// off, and why the sweep case matters more than any single stop point.
+//
+// The shapes are the two CTS cases verbatim in structure:
+//   * the workload is glcClipDistance.cpp FunctionalTest's inner loop (a program
+//     per (redeclaration, clip count), glEnable(GL_CLIP_DISTANCEi), an FBO per
+//     primitive type, a draw and a readback), including its early-return
+//     behaviour: on failure the test returns WITHOUT running its "clip clean"
+//     loop, so GL_CLIP_DISTANCE0..N-1 stay enabled for the rest of the process.
+//     That leftover enable state is NOT the carrier (one of the cases below pins
+//     that); the object churn is.
+//   * the victim is gl3cTransformFeedback3Tests.cpp's skip_components: a
+//     gl_SkipComponents capture layout under GL_RASTERIZER_DISCARD, read back
+//     out of a buffer pre-filled with -1-i so that "captured nothing" is
+//     distinguishable from "captured the wrong thing".
+
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+#ifndef GL_CLIP_DISTANCE0
+#define GL_CLIP_DISTANCE0 0x3000
+#endif
+
+namespace MGITest {
+    namespace {
+
+        GLuint CompileShader(GLenum type, const std::string& source, std::string* log) {
+            const GLuint shader = glCreateShader(type);
+            const char* text = source.c_str();
+            glShaderSource(shader, 1, &text, nullptr);
+            glCompileShader(shader);
+            GLint status = GL_FALSE;
+            glGetShaderiv(shader, GL_COMPILE_STATUS, &status);
+            if (status == GL_FALSE) {
+                GLint length = 0;
+                glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+                std::vector<char> buffer(static_cast<std::size_t>(length) + 1, '\0');
+                glGetShaderInfoLog(shader, length + 1, nullptr, buffer.data());
+                if (log != nullptr) *log = buffer.data();
+                glDeleteShader(shader);
+                return 0;
+            }
+            return shader;
+        }
+
+        // Links a vertex/fragment pair, optionally declaring transform feedback
+        // varyings first (glTransformFeedbackVaryings takes effect at the next link,
+        // exactly as the CTS uses it).
+        GLuint BuildProgram(const std::string& vertexSource, const std::string& fragmentSource,
+                            const std::vector<const char*>& xfbVaryings, GLenum bufferMode, std::string* log) {
+            const GLuint vertexShader = CompileShader(GL_VERTEX_SHADER, vertexSource, log);
+            if (vertexShader == 0) return 0;
+            const GLuint fragmentShader = CompileShader(GL_FRAGMENT_SHADER, fragmentSource, log);
+            if (fragmentShader == 0) {
+                glDeleteShader(vertexShader);
+                return 0;
+            }
+            const GLuint program = glCreateProgram();
+            glAttachShader(program, vertexShader);
+            glAttachShader(program, fragmentShader);
+            if (!xfbVaryings.empty()) {
+                glTransformFeedbackVaryings(program, static_cast<GLsizei>(xfbVaryings.size()), xfbVaryings.data(),
+                                            bufferMode);
+            }
+            glLinkProgram(program);
+            glDeleteShader(vertexShader);
+            glDeleteShader(fragmentShader);
+            GLint status = GL_FALSE;
+            glGetProgramiv(program, GL_LINK_STATUS, &status);
+            if (status == GL_FALSE) {
+                GLint length = 0;
+                glGetProgramiv(program, GL_INFO_LOG_LENGTH, &length);
+                std::vector<char> buffer(static_cast<std::size_t>(length) + 1, '\0');
+                glGetProgramInfoLog(program, length + 1, nullptr, buffer.data());
+                if (log != nullptr) *log = buffer.data();
+                glDeleteProgram(program);
+                return 0;
+            }
+            return program;
+        }
+
+        // ---------------------------------------------------------------- poison
+
+        // glcClipDistance.cpp FunctionalTest::m_vertex_shader_code with the same
+        // three substitutions (redeclaration, clip function, array setter).
+        std::string ClipVertexSource(bool redeclaration, unsigned clipCount, unsigned clipFunction,
+                                     unsigned vertexCount) {
+            const std::string count = std::to_string(clipCount);
+            std::string source = "#version 400 core\n\n";
+            if (redeclaration) {
+                source += "out float gl_ClipDistance[" + count + "];\n";
+            }
+            source += "\n";
+            switch (clipFunction) {
+            case 0:
+                source += "float f(int i)\n{\n    return 0.0;\n}\n";
+                break;
+            case 1:
+                source += "float f(int i)\n{\n    return 0.25 + 0.75 * (float(i) + 1.0) * (float(gl_VertexID) + 1.0)"
+                          " / (float(" + count + ") * float(" + std::to_string(vertexCount) + "));\n}\n";
+                break;
+            default:
+                source += "float f(int i)\n{\n    return - 0.25 - 0.75 * (float(i) + 1.0) * (float(gl_VertexID) + 1.0)"
+                          " / (float(" + count + ") * float(" + std::to_string(vertexCount) + "));\n}\n";
+                break;
+            }
+            source += "\nin vec4 position;\n\nvoid main()\n{\n";
+            if (redeclaration) {
+                // Dynamic array setter.
+                source += "    for(int i = 0; i < " + count + "; i++)\n    {\n"
+                          "        gl_ClipDistance[i] = f(i);\n    }\n";
+            } else {
+                // Static array setter, at the highest index this iteration enables.
+                const std::string index = std::to_string(clipCount - 1);
+                source += "    gl_ClipDistance[" + index + "] = f(" + index + ");\n";
+            }
+            source += "\n    gl_Position  = position;\n}\n";
+            return source;
+        }
+
+        const char* kClipFragmentSource = R"(#version 400 core
+
+out vec4 color;
+
+void main()
+{
+    color = vec4(1.0, 0.0, 0.0, 1.0);
+}
+)";
+
+        // How far into FunctionalTest's loop nest to get before bailing out the way
+        // the CTS does on a failed check: return immediately, skipping the "clip
+        // clean" loop that would have disabled GL_CLIP_DISTANCEi again.
+        struct ClipStopPoint {
+            unsigned primitiveIndex = 0; // 0 = POINTS, 1 = LINES, 2 = TRIANGLES
+            unsigned clipFunction = 0;
+            bool redeclaration = false;
+            unsigned clipCount = 1; // 1..8, the iteration that "fails"
+        };
+
+        // Runs FunctionalTest's loop nest up to and including `stop`, then returns
+        // leaving exactly the state the CTS leaves behind on a failure.
+        void RunClipDistanceWorkload(const ClipStopPoint& stop) {
+            static const GLenum kPrimitiveTypes[] = {GL_POINTS, GL_LINES, GL_TRIANGLES};
+            static const GLsizei kPrimitiveIndices[] = {1, 2, 3};
+            static const float kPositions[3][12] = {
+                {0.0f, 0.0f, 0.0f, 1.0f},
+                {-1.0f, 0.0f, 0.0f, 1.0f, 1.0f, 0.0f, 0.0f, 1.0f},
+                {-1.0f, -1.0f, 0.0f, 1.0f, 1.0f, -1.0f, 0.0f, 1.0f, 0.0f, 1.0f, 0.0f, 1.0f},
+            };
+
+            for (unsigned primitiveIndex = 0; primitiveIndex <= stop.primitiveIndex; ++primitiveIndex) {
+                const GLenum primitiveType = kPrimitiveTypes[primitiveIndex];
+                const GLsizei vertexCount = kPrimitiveIndices[primitiveIndex];
+                const GLsizei framebufferSize = (primitiveType == GL_POINTS) ? 1 : 32;
+
+                GLuint colorBuffer = 0;
+                GLuint framebuffer = 0;
+                glGenRenderbuffers(1, &colorBuffer);
+                glBindRenderbuffer(GL_RENDERBUFFER, colorBuffer);
+                glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, framebufferSize, framebufferSize);
+                glGenFramebuffers(1, &framebuffer);
+                glBindFramebuffer(GL_FRAMEBUFFER, framebuffer);
+                glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, colorBuffer);
+                glViewport(0, 0, framebufferSize, framebufferSize);
+
+                const unsigned lastFunction =
+                    (primitiveIndex == stop.primitiveIndex) ? stop.clipFunction : 2u;
+                for (unsigned clipFunction = 0; clipFunction <= lastFunction; ++clipFunction) {
+                    const bool atStopFunction =
+                        primitiveIndex == stop.primitiveIndex && clipFunction == stop.clipFunction;
+                    for (unsigned redeclaration = 0; redeclaration < 2; ++redeclaration) {
+                        const bool atStopRedeclaration =
+                            atStopFunction && (redeclaration != 0) == stop.redeclaration;
+                        const unsigned lastCount = atStopRedeclaration ? stop.clipCount : 8u;
+                        for (unsigned clipCount = 1; clipCount <= lastCount; ++clipCount) {
+                            std::string log;
+                            const GLuint program =
+                                BuildProgram(ClipVertexSource(redeclaration != 0, clipCount, clipFunction,
+                                                              static_cast<unsigned>(vertexCount)),
+                                             kClipFragmentSource, {}, GL_INTERLEAVED_ATTRIBS, &log);
+                            if (program == 0) continue;
+                            glUseProgram(program);
+
+                            glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
+                            glClear(GL_COLOR_BUFFER_BIT);
+
+                            glEnable(GL_CLIP_DISTANCE0 + clipCount - 1);
+
+                            GLuint vao = 0;
+                            GLuint vbo = 0;
+                            glGenVertexArrays(1, &vao);
+                            glBindVertexArray(vao);
+                            glGenBuffers(1, &vbo);
+                            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+                            glBufferData(GL_ARRAY_BUFFER,
+                                         static_cast<GLsizeiptr>(sizeof(float) * 4 * vertexCount),
+                                         kPositions[primitiveIndex], GL_STATIC_DRAW);
+                            const GLint location = glGetAttribLocation(program, "position");
+                            if (location >= 0) {
+                                glEnableVertexAttribArray(static_cast<GLuint>(location));
+                                glVertexAttribPointer(static_cast<GLuint>(location), 4, GL_FLOAT, GL_FALSE, 0,
+                                                      nullptr);
+                            }
+
+                            glDrawArrays(primitiveType, 0, vertexCount);
+
+                            std::vector<unsigned char> pixels(
+                                static_cast<std::size_t>(framebufferSize) * framebufferSize * 4, 0);
+                            glReadPixels(0, 0, framebufferSize, framebufferSize, GL_RGBA, GL_UNSIGNED_BYTE,
+                                         pixels.data());
+
+                            glBindBuffer(GL_ARRAY_BUFFER, 0);
+                            glBindVertexArray(0);
+                            glUseProgram(0);
+                            // MGL_REPRO_KEEPCLIPOBJ leaks the per-iteration objects so
+                            // no GL name and no heap address can be recycled into the
+                            // capture phase.
+                            // Deleting all three is load-bearing, not tidiness: the defect
+                            // this scenario pins needs the VAO's AND its vertex buffer's heap
+                            // addresses to be freed here so the capture phase's own objects
+                            // can be handed the same ones back.
+                            glDeleteBuffers(1, &vbo);
+                            glDeleteVertexArrays(1, &vao);
+                            glDeleteProgram(program);
+
+                            if (atStopRedeclaration && clipCount == stop.clipCount) {
+                                // The CTS's early return: the "clip clean" loop below
+                                // never runs, so the enables survive.
+                                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                                glDeleteFramebuffers(1, &framebuffer);
+                                glDeleteRenderbuffers(1, &colorBuffer);
+                                return;
+                            }
+                        }
+                        for (unsigned i = 0; i < 8; ++i) {
+                            glDisable(GL_CLIP_DISTANCE0 + i);
+                        }
+                    }
+                }
+
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                glDeleteFramebuffers(1, &framebuffer);
+                glDeleteRenderbuffers(1, &colorBuffer);
+            }
+        }
+
+        // ---------------------------------------------------------------- victim
+
+        // gl3cTransformFeedback3Tests.cpp TransformFeedbackBaseTestCase::m_shader_vert.
+        const char* kXfbVertexSource = R"(#version 400 core
+in vec4 vertex;
+out vec4 value1;
+out vec4 value2;
+out vec4 value3;
+out vec4 value4;
+
+void main (void)
+{
+    vec4 temp = vertex;
+
+    gl_Position = temp;
+
+    value1 = abs(temp) * 1.0;
+    value2 = abs(temp) * 2.0;
+    value3 = abs(temp) * 3.0;
+    value4 = abs(temp) * 4.0;
+}
+)";
+
+        const char* kXfbFragmentSource = R"(#version 400 core
+out vec4 color;
+void main (void)
+{
+    color = vec4(0.0, 0.0, 0.0, 1.0);
+}
+)";
+
+        // The skip_components capture layout, verbatim.
+        std::vector<const char*> SkipComponentsVaryings() {
+            return {"gl_SkipComponents1", "value1",  "gl_SkipComponents2", "gl_SkipComponents1", "value2",
+                    "gl_SkipComponents3", "gl_SkipComponents2", "value3", "gl_SkipComponents4", "value4"};
+        }
+
+        constexpr unsigned kSkipComponentCount = 4 * 4 + (1 + 2 + 3 + 4 + 1 + 2); // 16 values + 13 skipped
+        constexpr unsigned kSkipVertexCount = 6;
+
+        // Runs skip_components and reports what came back. `outCaptured` is the raw
+        // readback so a failure can say whether anything was written at all.
+        void RunSkipComponentsCapture(std::vector<float>& outCaptured, std::string* buildLog) {
+            outCaptured.clear();
+
+            const GLuint program = BuildProgram(kXfbVertexSource, kXfbFragmentSource, SkipComponentsVaryings(),
+                                                GL_INTERLEAVED_ATTRIBS, buildLog);
+            ASSERT_NE(program, 0u) << "skip_components program failed to link: " << (buildLog ? *buildLog : "");
+            glUseProgram(program);
+
+            const std::vector<float> vertices = {
+                -1.0f, -1.0f, -1.0f, 1.0f, 1.0f,  -1.0f, -2.0f, 1.0f, -1.0f, 1.0f, -3.0f, 1.0f,
+                1.0f,  1.0f,  4.0f,  1.0f, -1.0f, 1.0f,  5.0f,  1.0f, 1.0f,  -1.0f, 6.0f, 1.0f,
+            };
+
+            GLuint vao = 0;
+            GLuint vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, static_cast<GLsizeiptr>(sizeof(float) * vertices.size()), vertices.data(),
+                         GL_STATIC_DRAW);
+            const GLint location = glGetAttribLocation(program, "vertex");
+            if (location >= 0) {
+                glEnableVertexAttribArray(static_cast<GLuint>(location));
+                glVertexAttribPointer(static_cast<GLuint>(location), 4, GL_FLOAT, GL_FALSE, 0, nullptr);
+            }
+
+            const unsigned floatCount = kSkipVertexCount * kSkipComponentCount;
+            const GLsizeiptr byteSize = static_cast<GLsizeiptr>(sizeof(float) * floatCount);
+
+            GLuint captureBuffer = 0;
+            glGenBuffers(1, &captureBuffer);
+            glBindBuffer(GL_ARRAY_BUFFER, captureBuffer);
+            glBufferData(GL_ARRAY_BUFFER, byteSize, nullptr, GL_STATIC_READ);
+            glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, captureBuffer);
+            glBindBuffer(GL_ARRAY_BUFFER, 0);
+
+            // The pre-fill that makes "nothing was captured" recognisable.
+            std::vector<float> prefill(floatCount);
+            for (unsigned i = 0; i < floatCount; ++i) {
+                prefill[i] = -1.0f - static_cast<float>(i);
+            }
+            glBindBuffer(GL_ARRAY_BUFFER, captureBuffer);
+            glBufferData(GL_ARRAY_BUFFER, byteSize, prefill.data(), GL_STATIC_DRAW);
+            glBindBuffer(GL_ARRAY_BUFFER, 0);
+
+            glEnable(GL_RASTERIZER_DISCARD);
+            glClearColor(0.1f, 0.0f, 0.5f, 1.0f);
+            glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
+            glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, captureBuffer);
+            glBeginTransformFeedback(GL_TRIANGLES);
+            glDrawArrays(GL_TRIANGLES, 0, static_cast<GLsizei>(kSkipVertexCount));
+            glEndTransformFeedback();
+            glDisable(GL_RASTERIZER_DISCARD);
+
+            outCaptured.resize(floatCount);
+            glBindBufferRange(GL_TRANSFORM_FEEDBACK_BUFFER, 0, captureBuffer, 0, byteSize);
+            const void* mapped = glMapBufferRange(GL_TRANSFORM_FEEDBACK_BUFFER, 0, byteSize, GL_MAP_READ_BIT);
+            if (mapped != nullptr) {
+                std::memcpy(outCaptured.data(), mapped, static_cast<std::size_t>(byteSize));
+                glUnmapBuffer(GL_TRANSFORM_FEEDBACK_BUFFER);
+            }
+
+            glDisableVertexAttribArray(0);
+            glBindBuffer(GL_ARRAY_BUFFER, 0);
+            glDeleteBuffers(1, &vbo);
+            glDeleteBuffers(1, &captureBuffer);
+            glBindVertexArray(0);
+            glDeleteVertexArrays(1, &vao);
+            glUseProgram(0);
+            glDeleteProgram(program);
+            glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, 0);
+        }
+
+        // skip_components' expected buffer: the 13 skipped components keep their
+        // pre-fill, the 16 captured ones carry |vertex| * n.
+        std::vector<float> SkipComponentsExpected() {
+            const std::vector<float> vertices = {
+                -1.0f, -1.0f, -1.0f, 1.0f, 1.0f,  -1.0f, -2.0f, 1.0f, -1.0f, 1.0f, -3.0f, 1.0f,
+                1.0f,  1.0f,  4.0f,  1.0f, -1.0f, 1.0f,  5.0f,  1.0f, 1.0f,  -1.0f, 6.0f, 1.0f,
+            };
+            const unsigned floatCount = kSkipVertexCount * kSkipComponentCount;
+            std::vector<float> expected(floatCount);
+            for (unsigned i = 0; i < floatCount; ++i) {
+                expected[i] = -1.0f - static_cast<float>(i);
+            }
+            // Record layout, in floats:
+            //   [0]      skip1
+            //   [1..4]   value1
+            //   [5..7]   skip2 + skip1
+            //   [8..11]  value2
+            //   [12..16] skip3 + skip2
+            //   [17..20] value3
+            //   [21..24] skip4
+            //   [25..28] value4
+            static const unsigned kValueOffsets[4] = {1, 8, 17, 25};
+            for (unsigned v = 0; v < kSkipVertexCount; ++v) {
+                const unsigned base = v * kSkipComponentCount;
+                for (unsigned value = 0; value < 4; ++value) {
+                    for (unsigned component = 0; component < 4; ++component) {
+                        const float source = vertices[v * 4 + component];
+                        expected[base + kValueOffsets[value] + component] =
+                            std::fabs(source) * static_cast<float>(value + 1);
+                    }
+                }
+            }
+            return expected;
+        }
+
+        // Reports the first mismatch, and whether the readback is byte-for-byte the
+        // pre-fill (i.e. the capture never happened).
+        ::testing::AssertionResult CheckSkipComponents(const std::vector<float>& captured) {
+            const std::vector<float> expected = SkipComponentsExpected();
+            if (captured.size() != expected.size()) {
+                return ::testing::AssertionFailure()
+                       << "readback size " << captured.size() << " != " << expected.size();
+            }
+            bool anyWritten = false;
+            for (std::size_t i = 0; i < captured.size(); ++i) {
+                if (captured[i] != -1.0f - static_cast<float>(i)) {
+                    anyWritten = true;
+                    break;
+                }
+            }
+            for (std::size_t i = 0; i < expected.size(); ++i) {
+                if (std::fabs(captured[i] - expected[i]) > 0.0125f) {
+                    return ::testing::AssertionFailure()
+                           << "capture mismatch at index " << i << ": got " << captured[i] << ", expected "
+                           << expected[i] << (anyWritten ? "" : " (the whole buffer is still the pre-fill: "
+                                                               "NOTHING was captured)");
+                }
+            }
+            return ::testing::AssertionSuccess();
+        }
+
+        // The harness turns "no context came up" into a clean skip, and a skip is
+        // indistinguishable from a pass in a ctest summary. For this scenario that
+        // is a hole rather than a courtesy: the defect it pins is DirectVulkan's
+        // alone, and DirectVulkan now comes up headless on any machine at all - a
+        // surfaceless EGL platform over a software ICD (lavapipe) is enough. So
+        // "DirectVulkan did not initialise" here means the run is MISCONFIGURED,
+        // not that the machine has no GPU, and it must not report green.
+        //
+        // Local on purpose: the harness-wide skip semantics are deliberate
+        // (ScenarioFixture.h states the reasoning), and MOBILEGL_ITEST_REQUIRE_GPU
+        // is the harness-wide lever for the same intent - but that lever also
+        // demands a HARDWARE renderer, which is exactly what a lavapipe-only box
+        // cannot offer. This overrides nothing else: only this scenario, only for
+        // the backend that can regress, and only for the unusable-harness case.
+        class XfbAfterClipDistanceScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                // Ready() is false on the base's skip path AND on its REQUIRE_GPU
+                // failure path; the second one has already failed, so leave it alone
+                // rather than burying its reason under a second message.
+                if (Ready() || HasFatalFailure()) return;
+                if (Gl().BackendName() == "DirectVulkan") {
+                    FAIL() << "DirectVulkan could not be brought up, so the regression this scenario guards - a "
+                              "draw served a destroyed VAO's memoised vertex bindings - was never exercised, and "
+                              "that must be a failure rather than a silent skip. Headless bring-up needs only a "
+                              "Vulkan ICD and a surfaceless EGL platform (a software ICD such as lavapipe "
+                              "qualifies: VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/lvp_icd.x86_64.json with "
+                              "EGL_PLATFORM=surfaceless). Harness reason: "
+                           << Gl().SkipReason();
+                }
+            }
+        };
+
+        // Control: the capture on its own must work.
+        TEST_F(XfbAfterClipDistanceScenario, SkipComponentsCaptureAlone) {
+            if (!Ready()) return;
+            std::vector<float> captured;
+            std::string log;
+            RunSkipComponentsCapture(captured, &log);
+            EXPECT_TRUE(CheckSkipComponents(captured));
+        }
+
+        // Bisection step 1: only the leftover GL_CLIP_DISTANCEi enables.
+        TEST_F(XfbAfterClipDistanceScenario, SkipComponentsCaptureAfterClipDistanceEnables) {
+            if (!Ready()) return;
+            for (unsigned i = 0; i < 8; ++i) {
+                glEnable(GL_CLIP_DISTANCE0 + i);
+            }
+            std::vector<float> captured;
+            std::string log;
+            RunSkipComponentsCapture(captured, &log);
+            for (unsigned i = 0; i < 8; ++i) {
+                glDisable(GL_CLIP_DISTANCE0 + i);
+            }
+            EXPECT_TRUE(CheckSkipComponents(captured));
+        }
+
+        // Bisection step 2: the whole clip_distance.functional workload, stopped
+        // where the CTS stopped in the runs that went on to break the capture.
+        TEST_F(XfbAfterClipDistanceScenario, SkipComponentsCaptureAfterClipDistanceWorkloadLines8) {
+            if (!Ready()) return;
+            RunClipDistanceWorkload({.primitiveIndex = 1, .clipFunction = 0, .redeclaration = false, .clipCount = 8});
+            std::vector<float> captured;
+            std::string log;
+            RunSkipComponentsCapture(captured, &log);
+            for (unsigned i = 0; i < 8; ++i) {
+                glDisable(GL_CLIP_DISTANCE0 + i);
+            }
+            EXPECT_TRUE(CheckSkipComponents(captured));
+        }
+
+        TEST_F(XfbAfterClipDistanceScenario, SkipComponentsCaptureAfterClipDistanceWorkloadPoints1) {
+            if (!Ready()) return;
+            RunClipDistanceWorkload({.primitiveIndex = 0, .clipFunction = 0, .redeclaration = true, .clipCount = 1});
+            std::vector<float> captured;
+            std::string log;
+            RunSkipComponentsCapture(captured, &log);
+            for (unsigned i = 0; i < 8; ++i) {
+                glDisable(GL_CLIP_DISTANCE0 + i);
+            }
+            EXPECT_TRUE(CheckSkipComponents(captured));
+        }
+
+        // A single stop point is not a regression test for this defect: whether the
+        // capture phase's VAO and vertex buffer land on the addresses the workload just
+        // freed is a function of how much the workload allocated, so the two cases above
+        // pin two draws of a lottery. Sweep the grid instead - before the fix, roughly a
+        // third of these stop points came back holding the workload's vertex data.
+        TEST_F(XfbAfterClipDistanceScenario, SkipComponentsCaptureSurvivesEveryClipWorkloadStopPoint) {
+            if (!Ready()) return;
+            for (unsigned primitiveIndex = 0; primitiveIndex < 3; ++primitiveIndex) {
+                for (unsigned redeclaration = 0; redeclaration < 2; ++redeclaration) {
+                    for (const unsigned clipCount : {1u, 4u, 8u}) {
+                        RunClipDistanceWorkload({.primitiveIndex = primitiveIndex,
+                                                 .clipFunction = 0,
+                                                 .redeclaration = redeclaration != 0,
+                                                 .clipCount = clipCount});
+                        std::vector<float> captured;
+                        std::string log;
+                        RunSkipComponentsCapture(captured, &log);
+                        for (unsigned i = 0; i < 8; ++i) {
+                            glDisable(GL_CLIP_DISTANCE0 + i);
+                        }
+                        EXPECT_TRUE(CheckSkipComponents(captured))
+                            << " (stop point: primitive " << primitiveIndex << ", redeclaration " << redeclaration
+                            << ", clip count " << clipCount << ")";
+                    }
+                }
+            }
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/run_integration_test.sh

@@ -0,0 +1,47 @@
+#!/bin/bash
+# Run the headless MobileGL integration scenarios on one backend:
+#   ./run_integration_test.sh espryt [gtest args...]     -> DirectGLES
+#   ./run_integration_test.sh magma  [gtest args...]     -> DirectVulkan
+#
+# The backend is latched at initialization from MOBILEGL_BACKEND_TYPE, so one
+# process is one backend; this script is the dev-box equivalent of the two ctest
+# registrations in CMakeLists.txt.
+#
+# Pin the vendor libraries explicitly, for the same reason
+# MG_Benchmark/Driver/run_driver_bench.sh does: a bare libEGL on a glvnd system
+# resolves to whatever vendor comes first, which is usually Mesa/llvmpipe - a
+# software rasteriser silently replacing the GPU under a GPU test. Override
+# MGL_EGL_VENDOR / MGL_VK_ICD to test another driver.
+#
+# Set MOBILEGL_ITEST_REQUIRE_GPU=1 to turn "the harness is unusable" from a clean
+# skip into a failure. Do that anywhere the machine is supposed to have a GPU: a
+# run that skipped everything and a run that passed everything are otherwise the
+# same green, so without it a broken driver pinning is invisible.
+set -eu
+HERE=$(cd "$(dirname "$0")" && pwd)
+BIN=${MOBILEGL_ITEST_BIN:-$HERE/MobileGLIntegrationTest}
+EGL_VENDOR=${MGL_EGL_VENDOR:-/usr/share/glvnd/egl_vendor.d/10_nvidia.json}
+VK_ICD=${MGL_VK_ICD:-/usr/share/vulkan/icd.d/nvidia_icd.x86_64.json}
+MODE=$1; shift
+
+if [ ! -x "$BIN" ]; then
+  echo "MobileGLIntegrationTest not found at $BIN"
+  echo "configure with -DMOBILEGL_BUILD_INTEGRATION_TEST=ON and set MOBILEGL_ITEST_BIN"
+  exit 1
+fi
+
+[ -r "$EGL_VENDOR" ] && export __EGL_VENDOR_LIBRARY_FILENAMES=$EGL_VENDOR
+export EGL_PLATFORM=${EGL_PLATFORM:-x11}
+
+case "$MODE" in
+  espryt|DirectGLES)
+    export MOBILEGL_BACKEND_TYPE=DirectGLES
+    ;;
+  magma|DirectVulkan)
+    export MOBILEGL_BACKEND_TYPE=DirectVulkan
+    [ -r "$VK_ICD" ] && export VK_ICD_FILENAMES=$VK_ICD
+    ;;
+  *) echo "unknown mode: $MODE (espryt|magma)"; exit 1 ;;
+esac
+export MOBILEGL_ITEST_REQUIRE_GPU=${MOBILEGL_ITEST_REQUIRE_GPU:-}
+exec "$BIN" "$@"

MobileGL/MG_State/GLState/BufferState/BufferObject.cpp

@@ -8,9 +8,17 @@
 
 #include "BufferObject.h"
 
+#include <atomic>
+
 namespace MobileGL::MG_State::GLState {
     namespace {
         const BufferBackendOps* g_bufferBackendOps = nullptr;
+        // Starts at 1 so a zero-initialized cache slot can never carry a live buffer's id.
+        std::atomic<Uint64> g_nextBufferLifetimeId{1};
+    }
+
+    Uint64 BufferObject::AllocateLifetimeId() {
+        return g_nextBufferLifetimeId.fetch_add(1, std::memory_order_relaxed);
     }
 
     void SetBufferBackendOps(const BufferBackendOps* ops) {
@@ -198,8 +206,11 @@ namespace MobileGL::MG_State::GLState {
     }
 
     void BufferObject::UploadSubData(DataPtr data, SizeT atOffset) {
-        MOBILEGL_ASSERT(!m_isMapped || (m_mappingAccess & BufferMappingAccessBit::Persistent),
+        // GL 4.6 core 6.5 forbids only the OVERLAPPING write: a glBufferSubData that stays
-                        "Cannot upload sub data while buffer is non-persistently mapped.");
+        // clear of a non-persistent mapping is legal, and the frontend lets it through.
+        MOBILEGL_ASSERT(!m_isMapped || (m_mappingAccess & BufferMappingAccessBit::Persistent) ||
+                            atOffset >= m_mappedRange.end || atOffset + data.size <= m_mappedRange.start,
+                        "Cannot upload sub data overlapping a non-persistent mapping.");
         MOBILEGL_ASSERT(atOffset + data.size <= m_size,
                         "UploadSubData out of bounds: atOffset (%zu) + data.size (%zu) > m_size (%zu)", atOffset,
                         data.size, m_size);

MobileGL/MG_State/GLState/BufferState/BufferObject.h

@@ -185,6 +185,13 @@ namespace MobileGL {
             Flags<BufferMappingAccessBit> GetMappingAccess() const;
             GLbitfield GetStorageFlags() const;
             Uint GetExternalIndex() const;
+            // Globally-unique, never-reused id for THIS object's lifetime - same contract
+            // and same motivation as ProgramObject::GetLifetimeId() and
+            // VertexArrayObject::GetLifetimeId(). A backend that folds a buffer's IDENTITY
+            // into a cache key must use this, never the GL name (LIFO-recycled by
+            // glGenBuffers) and never the heap address (recycled by the allocator): both
+            // let a deleted-and-recreated buffer answer to a dead one's cache entry.
+            Uint64 GetLifetimeId() const { return m_lifetimeId; }
             // Monotonic counter bumped on every shadow mutation; backends use it to
             // validate cached transient slices.
             Uint64 GetChangeSerial() const;
@@ -207,7 +214,10 @@ namespace MobileGL {
             // SubData transfer to sync the backend's separate GPU copy.
             void NotifyContentWrite(SizeT offset, SizeT size);
 
+            static Uint64 AllocateLifetimeId();
+
             const Uint m_externalIndex = 0;
+            const Uint64 m_lifetimeId = AllocateLifetimeId();
             SizeT m_size = 0;
             BufferUsage m_usage = BufferUsage::StaticDraw;
             // Owns the buffer's bytes (CPU shadow or backend persistent GPU map) and

MobileGL/MG_State/GLState/Core.cpp

@@ -9,6 +9,9 @@
 #include "Core.h"
 #include "MG_State/GLState/RenderbufferState/RenderbufferObject.h"
 #include "MG_State/EGLState/Core.h"
+#include <MG_Backend/BackendObjects.h>
+#include <MG_Util/Async/ShaderCompilePool.h>
+#include <MG_Util/ShaderTranspiler/CompileEnv.h>
 #include <Config.h>
 
 namespace MobileGL::MG_State {
@@ -24,8 +27,27 @@ namespace MobileGL::MG_State {
     }
 
     namespace GLState {
+        const SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv>& GLContext::GetCompileEnv() {
+            const void* backend = static_cast<const void*>(MG_Backend::pActiveBackendObject.get());
+            if (!m_compileEnv || m_compileEnvBackend != backend) {
+                // First use, or the backend was swapped underneath us. Re-capturing rolls the
+                // fingerprint, so every P0b preprocess memo computed against the old backend's
+                // limits becomes structurally unreachable instead of silently reusable.
+                m_compileEnv = MG_Util::ShaderTranspiler::CaptureCompileEnv();
+                m_compileEnvBackend = backend;
+            }
+            return m_compileEnv;
+        }
+
         // Error
         void GLContext::RecordError(ErrorCode code, UniquePtr<ErrorInfo> info) {
+            // Invariant I1, mechanically enforced: the GL error state is GL-thread-owned.
+            // A compile or link body that needs to raise an error must append to its node's
+            // JobDiagnostics and let the join replay it here (see the P1 design section 6);
+            // reaching this from a worker would corrupt the sticky-flag set that
+            // glGetError's ordering depends on.
+            MOBILEGL_ASSERT(!MG_Util::Async::ShaderCompilePool::IsPoolThread(),
+                            "GLContext::RecordError() called from a shader-compile pool thread");
             m_errorState.RecordError(code, Move(info));
         }
 
@@ -335,6 +357,10 @@ namespace MobileGL::MG_State {
             return m_programState.GetShaderObject(index);
         }
 
+        void GLContext::JoinAllPendingShaderWork() {
+            m_programState.JoinAllPendingWork();
+        }
+
         void GLContext::UseProgram(Uint program) {
             return m_programState.UseProgram(program);
         }
@@ -346,11 +372,35 @@ namespace MobileGL::MG_State {
         const SharedPtr<ProgramObject>& GLContext::GetProgramForDraw() {
             static const SharedPtr<ProgramObject> nullProgram = nullptr;
             const auto& currentProgram = m_programState.GetCurrentProgram();
-            if (currentProgram) return currentProgram;
+            if (currentProgram) {
+                // P1 join site J1, plain glUseProgram half. The backends read a program's
+                // lifetimeId / backendStateVersion / UBO content version to decide whether
+                // their per-program caches are still valid, and none of those pass through
+                // ProgramObject's join gate - so a draw could sample a version, join later
+                // inside the same draw when it finally touched an artifact, and cache under a
+                // version the publish had already superseded. Settling here means every
+                // version a backend reads during a draw describes the program it is drawing.
+                // One null check in steady state.
+                currentProgram->JoinLink();
+                return currentProgram;
+            }
             if (m_boundProgramPipeline == 0) return nullProgram;
             const auto& pipeline = GetBoundProgramPipeline();
             if (!pipeline) return nullProgram;
 
+            // P1 join site J1. ComputeDrawProgramSignature() keys the composite cache on each
+            // stage program's lifetimeId and backendStateVersion - NON-artifact fields, so
+            // they do not pass through ProgramObject's join gate and a pending link would
+            // stay pending right through the signature. Since the version is bumped both at
+            // enqueue and at publish, the signature computed inside a pending window is one
+            // that will never be produced again: every draw would miss the cache and rebuild
+            // (and relink) the composite. Join first, so the signature describes settled
+            // programs. In steady state this is a null check per stage.
+            for (SizeT stage = 0; stage < static_cast<SizeT>(ShaderStage::ShaderStageCount); ++stage) {
+                const auto& stageProgram = pipeline->GetStageProgram(static_cast<ShaderStage>(stage));
+                if (stageProgram) stageProgram->JoinLink();
+            }
+
             const auto signature = pipeline->ComputeDrawProgramSignature();
             if (const auto& cached = pipeline->GetCachedDrawProgram(signature)) return cached;
 
@@ -378,6 +428,10 @@ namespace MobileGL::MG_State {
             // A pipeline with no fragment stage still rasterises, so the default fragment
             // shader is wanted here even though the separable stage programs never get one.
             composite->Link(true);
+            // P1 join site J2. The draw that asked for this program is the very next thing to
+            // happen, so enqueueing the composite's link buys nothing and only moves the wait
+            // to whichever backend accessor happens to touch its artifacts first.
+            composite->JoinLink();
             pipeline->SetCachedDrawProgram(signature, Move(composite));
             return pipeline->GetCachedDrawProgram(signature);
         }

MobileGL/MG_State/GLState/Core.h

@@ -21,6 +21,10 @@
 #include "VertexArrayState/VertexArrayState.h"
 #include "RenderbufferState/RenderbufferState.h"
 
+namespace MobileGL::MG_Util::ShaderTranspiler {
+    struct CompileEnv;
+}
+
 namespace MobileGL {
     namespace MG_State {
         void Init();
@@ -120,6 +124,18 @@ namespace MobileGL {
                 // per-draw sampled-texture set.
                 Uint64 GetTextureBindGeneration() const { return m_textureState.GetTextureBindGeneration(); }
                 void BumpTextureBindGeneration() { m_textureState.BumpTextureBindGeneration(); }
+                // Monotonic counter bumped whenever a texture's shape or a sampler object's
+                // parameters change, i.e. whenever a bound texture's mipmap-completeness (and so
+                // whether a backend binds it at all) can have flipped without any bind moving;
+                // see TextureState::GetSamplingResolutionGeneration.
+                Uint64 GetSamplingResolutionGeneration() const {
+                    return m_textureState.GetSamplingResolutionGeneration();
+                }
+                void BumpSamplingResolutionGeneration() { m_textureState.BumpSamplingResolutionGeneration(); }
+                // Never-reused id of this context, for backend memos keyed on the two counters
+                // above: both restart at 0 in a new context, and a recreated context can land on
+                // the old heap address. See TextureState::GetContextId.
+                Uint64 GetTextureContextId() const { return m_textureState.GetContextId(); }
                 Bool ValidateTextureName(Uint index) const;
                 Bool ValidateTextureObject(Uint index) const;
                 Int GetActiveTextureUnit() const;
@@ -137,6 +153,14 @@ namespace MobileGL {
                 Bool ValidateShaderName(Uint index) const;
                 const SharedPtr<ProgramObject>& GetProgramObject(Uint index);
                 const SharedPtr<ShaderObject>& GetShaderObject(Uint index);
+                // Settles every compile and link this context still owns; see
+                // ProgramState::JoinAllPendingWork. Called by glMaxShaderCompilerThreadsKHR(0).
+                void JoinAllPendingShaderWork();
+                // P1 stage 6: the per-context index of adoptable compile nodes, for its
+                // adoption counter. Diagnostics and tests only - no GL entry point reads it.
+                ShaderCompileAdoptionMap& GetShaderCompileAdoptionMap() {
+                    return m_programState.GetShaderCompileAdoptionMap();
+                }
                 void UseProgram(Uint program);
                 const SharedPtr<ProgramObject>& GetCurrentProgram();
                 // What a draw or dispatch actually executes: the program in use, or - when
@@ -368,6 +392,15 @@ namespace MobileGL {
                 Bool ValidateRenderbufferName(Uint index) const;
                 Bool ValidateRenderbufferObject(Uint index) const;
 
+                // P1: the shader compile/link pipeline's snapshot of everything it reads from
+                // outside its own (stage, source) inputs. Captured lazily here because it
+                // cannot be captured in MG_State::Init() - that runs BEFORE MG_Backend::Init(),
+                // so there is no backend to query yet. Re-captured whenever the active backend
+                // object changes, which also rolls the fingerprint and therefore invalidates
+                // every P0b preprocess memo keyed against the old one.
+                // GL thread only.
+                const SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv>& GetCompileEnv();
+
             private:
                 // State Components
                 ErrorState m_errorState;
@@ -425,6 +458,11 @@ namespace MobileGL {
                 FramebufferState m_framebufferState;
                 SamplerState m_samplerState;
                 RenderbufferState m_renderbufferState;
+
+                mutable SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> m_compileEnv;
+                // Identity of the backend object m_compileEnv was captured against; a plain
+                // pointer compare, never dereferenced.
+                const void* m_compileEnvBackend = nullptr;
             };
         } // namespace GLState
 

MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.h

@@ -0,0 +1,111 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+#include <MG_State/GLState/ProgramState/ProgramObject.h>
+#include <MG_State/GLState/ProgramState/ShaderCompileTask.h>
+#include <MG_Util/Async/JobNode.h>
+#include <MG_Util/ShaderTranspiler/CompileEnv.h>
+
+namespace MobileGL::MG_State::GLState {
+    // One attached shader, as the link sees it: never the ShaderObject, always a snapshot.
+    //
+    // The ShaderObject is GL-thread-owned and may be re-sourced, detached or destroyed while
+    // this link is still queued; everything below is either immutable or independently owned,
+    // so none of that can reach the worker.
+    struct LinkShaderInput {
+        ShaderStage stage = ShaderStage::Unknown;
+        // For the compile-error diagnostic and the compute local_size check, both of which
+        // quote the ORIGINAL source rather than the preprocessed one.
+        SharedPtr<const String> source;
+        // The authoritative compiled state. Null, or non-Complete, both read as "this shader
+        // did not compile" - the same verdict ShaderObject's join gate produces.
+        SharedPtr<const ShaderCompileTask> compiled;
+    };
+
+    // The unit of asynchronous linking: one glLinkProgram's worth of pure CPU work - glslang
+    // link + mapIO, SPIR-V generation and optimization, the GL-facing reflection surface, the
+    // global-UBO routing tables, fragment-output validation and transform-feedback
+    // resolution - with every input it needs snapshotted at enqueue.
+    //
+    // Same ownership rule as ShaderCompileTask: the body reads nothing but `in` (all of it
+    // owned or immutable) and writes nothing but `artifacts`. No GL call, no
+    // pActiveBackendObject read, no pGLContext->RecordError(); the device limits arrive
+    // through the CompileEnv snapshot and diagnostics are deferred to the join.
+    //
+    // ONE LINK IS ONE HANDLER. RunBody() runs start to finish inside a single pool handler
+    // and is the only place `artifacts` is written. Do not split it across handlers to
+    // "pipeline" the reflection half: the intermediates that GlslangToSpv and buildReflection
+    // share are mutated in a strict order (see the GenerateSpirv-before-DoReflection comment
+    // in Run()), and a second handler would let a cancel land between them and publish a
+    // program whose SPIR-V and reflection describe different things.
+    class ProgramLinkTask final : public MG_Util::Async::JobNode {
+    public:
+        // ---- inputs, snapshotted on the GL thread in ProgramObject::Link()'s prologue ----
+        struct Inputs {
+            Uint externalIndex = 0; // logs only
+            Vector<LinkShaderInput> shaders; // already stage-sorted
+            SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> env;
+            // The four "takes effect at the next link" request maps. Snapshotted rather than
+            // referenced, which is precisely what makes glBindAttribLocation and friends
+            // legal to call over a pending link without cancelling it: the pending link keeps
+            // linking the inputs it was given.
+            UnorderedMap<String, Uint> explicitAttribLocations; // glBindAttribLocation
+            UnorderedMap<String, Uint> explicitFragDataLocation; // glBindFragDataLocation
+            UnorderedMap<String, Uint> explicitFragDataIndex; // glBindFragDataLocationIndexed
+            Vector<String> requestedXfbVaryings; // glTransformFeedbackVaryings
+            GLenum requestedXfbBufferMode = GL_INTERLEAVED_ATTRIBS;
+            Int maxFragmentOutputColorNumber = 8; // GL_MAX_DRAW_BUFFERS, stamped in by the entry point
+        } in;
+
+        // ---- output: valid iff IsComplete(), immutable afterwards ----
+        // Moved (never copied) into the ProgramObject by EnsureLinkJoined().
+        ProgramObject::LinkArtifacts artifacts;
+
+        // Posts this job once every compile in `deps` is terminal - and not one moment
+        // earlier, so the body never waits on anything (invariant I4: no job body may block
+        // on another job, or the pool could deadlock with all its workers waiting on each
+        // other). `deps` is the subset of the snapshot's compile nodes that were still
+        // in flight; an already-terminal one needs no edge.
+        //
+        // GL thread only, and only after the caller has stored a SharedPtr to this node:
+        // OnDepSettled takes shared_from_this().
+        void SubmitAfter(const Vector<SharedPtr<ShaderCompileTask>>& deps);
+
+    private:
+        void RunBody() override;
+
+        // Runs when one dependency goes terminal - on whichever thread drove it there, which
+        // is a pool worker for a compile that finished on one. Non-throwing by construction;
+        // see the definition.
+        void OnDepSettled();
+
+        // ---- the link body, split exactly as ProgramObject::Link() had it ----
+        // Each returns false to abort the link with `artifacts.infoLog` already set, which is
+        // GL's definition of a failed link: LINK_STATUS false plus a log, never a GL error.
+        Bool ConsumeShaders(Vector<SharedPtr<glslang::TShader>>& outShaders);
+        Bool DoReflection(const MG_Util::ShaderTranspiler::CompileEnv& env);
+        Bool ValidateFragmentOutputLocations();
+        Bool ResolveTransformFeedbackVaryings();
+        void ResolveGsTriangleStripCapture(const glslang::TIntermediate* captureIntermediate);
+        void GenerateSpirv();
+        void BuildGlobalUboRouting();
+
+        // Worker-side MGLOG replacement: appended to diagnostics.logLines and replayed by the
+        // join, on the GL thread, where a serial implementation would have printed it.
+        // Logging straight from a worker interleaves mid-line with the GL thread's output and
+        // lands out of order relative to the glLinkProgram that caused it.
+        void DeferLog(String line);
+
+        // Counts down to zero exactly once. Starts at deps + 1: the extra guard is released
+        // by SubmitAfter itself, so a dependency that settles while the edges are still being
+        // registered cannot post the job from under a half-built dependency list.
+        std::atomic<Int> m_remainingDeps{0};
+    };
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

@@ -14,9 +14,22 @@
 #include <MG_Util/ShaderTranspiler/SpvcSession.h>
 
 namespace MobileGL::MG_State::GLState {
+    // The link job. Only ever held by SharedPtr here, so a forward declaration is enough -
+    // ProgramLinkTask.h includes THIS header (it outputs a LinkArtifacts), so including it
+    // back would be circular. The destructor is therefore out of line.
+    class ProgramLinkTask;
+
     class ProgramObject {
     public:
         ProgramObject(Uint externalIndex) : m_externalIndex(externalIndex), m_lifetimeId(AllocateLifetimeId()) {}
+        // Cancel-not-join, exactly like ~ShaderObject: the link job owns its inputs, so an
+        // in-flight link whose program just went away is safe to abandon where it stands.
+        // Nothing can observe its result any more - this object was the only route to it.
+        // Out of line because ProgramLinkTask is incomplete here.
+        ~ProgramObject();
+        ProgramObject(const ProgramObject&) = delete;
+        ProgramObject& operator=(const ProgramObject&) = delete;
+
         bool ShaderIsAttached(const SharedPtr<ShaderObject>& shader);
         // GL-visible attachment: in the attach list and not pending detach (glDetachShader
         // defers the actual removal to the next link).
@@ -44,31 +57,39 @@ namespace MobileGL::MG_State::GLState {
 
         Vector<SharedPtr<ShaderObject>>& GetAttachedShaders();
         const Vector<SharedPtr<ShaderObject>>& GetAttachedShaders() const;
-        const String& GetInfoLog() const { return m_infoLog; }
+        const String& GetInfoLog() const { return Artifacts().infoLog; }
         // glCreateShaderProgramv folds the shader's compile log into the program's log, which
         // is the only place a caller can read it from once the shader name is gone.
         void AppendInfoLog(const String& text) {
             if (text.empty()) return;
-            if (!m_infoLog.empty() && m_infoLog.back() != '\n') m_infoLog += '\n';
+            if (!Artifacts().infoLog.empty() && Artifacts().infoLog.back() != '\n') Artifacts().infoLog += '\n';
-            m_infoLog += text;
+            Artifacts().infoLog += text;
         }
-        Int GetUniformMaxLength() const { return m_uniformNameMaxLength; }
+        Int GetUniformMaxLength() const { return Artifacts().uniformNameMaxLength; }
-        Uint GetUniformCount() const { return m_activeUniformCount; }
+        Uint GetUniformCount() const { return Artifacts().activeUniformCount; }
-        Uint GetMaxUniformLocation() const { return m_maxUniformLocation; }
+        Uint GetMaxUniformLocation() const { return Artifacts().maxUniformLocation; }
         Int GetUniformLocation(const String& name) const {
-            const auto it = m_uniformLocations.find(name);
+            const auto it = Artifacts().uniformLocations.find(name);
-            if (it != m_uniformLocations.end()) return (Int)it->second;
+            if (it != Artifacts().uniformLocations.end()) return (Int)it->second;
 
             // Reflection stores GL-style names: an array uniform is keyed "arr[0]" (its base
             // location). A bare "arr" query resolves to that entry; an "arr[k]" query resolves
             // to base + k because DoReflection reserves one location per array element.
             if (name.empty()) return -1;
             if (name.back() != ']') {
-                const auto suffixedIt = m_uniformLocations.find(name + "[0]");
+                const auto suffixedIt = Artifacts().uniformLocations.find(name + "[0]");
-                if (suffixedIt != m_uniformLocations.end()) return (Int)suffixedIt->second;
+                if (suffixedIt != Artifacts().uniformLocations.end()) return (Int)suffixedIt->second;
                 return -1;
             }
             if (name.length() < 4) return -1;
+            // An array of arrays is keyed by its full "[0]"-terminated spelling
+            // ("a[2][1][0]"), so a query that already ends in a subscript may still be the
+            // NAME of an array rather than an element of one. Try that first; only then
+            // treat the trailing subscript as an element index.
+            {
+                const auto arrayOfArraysIt = Artifacts().uniformLocations.find(name + "[0]");
+                if (arrayOfArraysIt != Artifacts().uniformLocations.end()) return (Int)arrayOfArraysIt->second;
+            }
             const SizeT bracket = name.rfind('[');
             // Require at least one digit between the brackets.
             if (bracket == String::npos || bracket + 1 >= name.length() - 1) return -1;
@@ -78,21 +99,21 @@ namespace MobileGL::MG_State::GLState {
                 element = element * 10 + static_cast<Uint>(name[i] - '0');
                 if (element > 0x0FFFFFFFu) return -1;
             }
-            auto baseIt = m_uniformLocations.find(name.substr(0, bracket) + "[0]");
+            auto baseIt = Artifacts().uniformLocations.find(name.substr(0, bracket) + "[0]");
-            if (baseIt == m_uniformLocations.end()) {
+            if (baseIt == Artifacts().uniformLocations.end()) {
                 // Legacy key without the "[0]" suffix (defensive; reflection normally
                 // stores the suffixed form for arrays).
-                baseIt = m_uniformLocations.find(name.substr(0, bracket));
+                baseIt = Artifacts().uniformLocations.find(name.substr(0, bracket));
-                if (baseIt == m_uniformLocations.end()) return -1;
+                if (baseIt == Artifacts().uniformLocations.end()) return -1;
             }
             const Int base = (Int)baseIt->second;
             if (!IsValidUniformLocation(base)) return -1;
-            const Int index = m_uniformIndexInTProgram[base];
+            const Int index = Artifacts().uniformIndexInTProgram[base];
             // "[k]" only addresses arrays ("scalar[0]" is not a uniform name), and only
             // in-range elements.
-            const glslang::TType* type = m_program->getUniform(index).getType();
+            const glslang::TType* type = Artifacts().program->getUniform(index).getType();
             if (type == nullptr || !type->isArray()) return -1;
-            if (static_cast<GLint>(element) >= GetActiveUniformArraySize(index)) return -1;
+            if (static_cast<GLint>(element) >= GetUniformArraySizeByTIndex(index)) return -1;
             const Int location = base + (Int)element;
             if (!UniformLocationsAliasSameUniform(base, location)) return -1;
             return location;
@@ -101,14 +122,41 @@ namespace MobileGL::MG_State::GLState {
         // True when both locations are element slots of the same uniform variable.
         Bool UniformLocationsAliasSameUniform(Int a, Int b) const {
             if (!IsValidUniformLocation(a) || !IsValidUniformLocation(b)) return false;
-            return m_uniformIndexInTProgram[a] == m_uniformIndexInTProgram[b];
+            return Artifacts().uniformIndexInTProgram[a] == Artifacts().uniformIndexInTProgram[b];
+        }
+
+        // ---- GL index <-> glslang TProgram index translation ----
+        // The single relaxed parse enumerates artifacts GL must not see: every declared
+        // default-block uniform (even dead ones) as a member of the synthesized
+        // MGL_GLOBAL_UBO, and that block itself. DoReflection builds filtered GL-facing
+        // index spaces; every public "index"-taking getter translates through them, so
+        // GL and backend consumers keep seeing exactly the pre-P0a surface.
+        Int TProgramUniformIndex(Uint glIndex) const {
+            return Artifacts().glUniformIndexToTProgram[glIndex];
+        }
+        Int GlUniformIndexFromTProgram(Int tIndex) const {
+            if (tIndex < 0 || tIndex >= static_cast<Int>(Artifacts().tProgramUniformIndexToGl.size())) return -1;
+            return Artifacts().tProgramUniformIndexToGl[tIndex];
+        }
+        // GL uniform-block index -> glslang TProgram block index (the inverse of
+        // GlBlockIndexFromTProgram). The interface-query layer needs it to reach block
+        // properties glslang exposes but no typed getter here does.
+        Int TProgramBlockIndex(Uint glBlockIndex) const {
+            return glBlockIndex < Artifacts().glBlockIndexToTProgram.size()
+                ? Artifacts().glBlockIndexToTProgram[glBlockIndex]
+                : -1;
+        }
+        Int GlBlockIndexFromTProgram(Int tBlockIndex) const {
+            if (tBlockIndex < 0 || tBlockIndex >= static_cast<Int>(Artifacts().tProgramBlockIndexToGl.size())) return -1;
+            return Artifacts().tProgramBlockIndexToGl[tBlockIndex];
         }
 
         Int GetActiveUniformIndex(const String& name) const {
-            const Int uniformIndex = m_program->getUniformIndex(name.c_str());
+            const Int tProgramCount = static_cast<Int>(Artifacts().tProgramUniformIndexToGl.size());
-            if (uniformIndex >= 0 && uniformIndex < m_activeUniformCount &&
+            const Int uniformIndex = Artifacts().program->getUniformIndex(name.c_str());
-                m_program->getUniform(uniformIndex).name == name) {
+            if (uniformIndex >= 0 && uniformIndex < tProgramCount &&
-                return uniformIndex;
+                Artifacts().program->getUniform(uniformIndex).name == name) {
+                return GlUniformIndexFromTProgram(uniformIndex);
             }
 
             // Reflection stores an array uniform under "arr[0]"; accept the bare "arr"
@@ -116,61 +164,60 @@ namespace MobileGL::MG_State::GLState {
             // robustness against non-suffixed reflection entries.
             if (!name.empty() && name.back() != ']') {
                 const String suffixedName = name + "[0]";
-                const Int suffixedIndex = m_program->getUniformIndex(suffixedName.c_str());
+                const Int suffixedIndex = Artifacts().program->getUniformIndex(suffixedName.c_str());
-                if (suffixedIndex >= 0 && suffixedIndex < m_activeUniformCount &&
+                if (suffixedIndex >= 0 && suffixedIndex < tProgramCount &&
-                    m_program->getUniform(suffixedIndex).name == suffixedName) {
+                    Artifacts().program->getUniform(suffixedIndex).name == suffixedName) {
-                    return suffixedIndex;
+                    return GlUniformIndexFromTProgram(suffixedIndex);
                 }
                 return -1;
             }
 
             if (name.length() <= 3 || name.compare(name.length() - 3, 3, "[0]") != 0) return -1;
             const String baseName = name.substr(0, name.length() - 3);
-            const Int baseIndex = m_program->getUniformIndex(baseName.c_str());
+            const Int baseIndex = Artifacts().program->getUniformIndex(baseName.c_str());
-            if (baseIndex < 0 || baseIndex >= m_activeUniformCount) return -1;
+            if (baseIndex < 0 || baseIndex >= tProgramCount) return -1;
-            return m_program->getUniform(baseIndex).name == baseName ? baseIndex : -1;
+            return Artifacts().program->getUniform(baseIndex).name == baseName ? GlUniformIndexFromTProgram(baseIndex)
+                                                                     : -1;
         }
 
-        Bool IsValidUniformLocation(Int location) const {
+        Bool IsValidUniformLocation(Int location) const { return IsValidUniformLocation(Artifacts(), location); }
-            if (location < 0 || location > static_cast<Int>(m_maxUniformLocation)) return false;
-            if (static_cast<SizeT>(location) >= m_uniformIndexInTProgram.size()) return false;
-            const Int uniformIndexInProgram = m_uniformIndexInTProgram[location];
-            return uniformIndexInProgram != glslang::TQualifier::layoutLocationEnd &&
-                   uniformIndexInProgram >= 0 && uniformIndexInProgram < m_activeUniformCount;
-        }
 
         GLenum GetUniformType(Uint location) const {
-            auto& uniform = m_program->getUniform(m_uniformIndexInTProgram[location]);
+            auto& uniform = Artifacts().program->getUniform(Artifacts().uniformIndexInTProgram[location]);
             return uniform.glDefineType;
         }
 
         GLenum GetActiveUniformType(Uint index) const {
-            auto& uniform = m_program->getUniform(static_cast<Int>(index));
+            auto& uniform = Artifacts().program->getUniform(TProgramUniformIndex(index));
             return uniform.glDefineType;
         }
 
         // Number of active array elements (GL_UNIFORM_SIZE / GL_ARRAY_SIZE); 1 for a non-array.
         // glslang's TObjectReflection.size only carries the element count for a NON-block array; for
         // a block array member it reports 1, so take the count from the TType, which is authoritative
-        // for both. GL 3.3 core uniforms are always sized.
+        // for both. GL 3.3 core uniforms are always sized. Takes a TProgram uniform index (the space
+        // the artifacts' uniformIndexInTProgram stores).
+        GLint GetUniformArraySizeByTIndex(Int tIndex) const {
+            return GetUniformArraySizeByTIndex(Artifacts(), tIndex);
+        }
+
         GLint GetActiveUniformArraySize(Uint index) const {
-            const auto& uniform = m_program->getUniform(static_cast<Int>(index));
+            return GetUniformArraySizeByTIndex(TProgramUniformIndex(index));
-            const glslang::TType* type = uniform.getType();
-            if (type != nullptr && type->isSizedArray()) {
-                return type->getOuterArraySize();
-            }
-            return uniform.size < 1 ? 1 : uniform.size;
         }
 
         Int GetActiveUniformBlockIndex(Uint index) const {
-            auto& uniform = m_program->getUniform(static_cast<Int>(index));
+            auto& uniform = Artifacts().program->getUniform(TProgramUniformIndex(index));
-            return uniform.index;
+            // Members of the synthesized global UBO are default-block uniforms to GL: -1.
+            return GlBlockIndexFromTProgram(uniform.index);
         }
 
-        // GL_UNIFORM_OFFSET: byte offset within the owning named block. glslang already reports -1
+        // GL_UNIFORM_OFFSET: byte offset within the owning named block; -1 for a default-block
-        // for a default-block uniform, which is exactly the spec value there.
+        // uniform. The relaxed parse gives global-UBO members real byte offsets, but GL must keep
+        // seeing them as default-block uniforms, so gate on the GL-visible block index.
         GLint GetActiveUniformOffset(Uint index) const {
-            return m_program->getUniform(static_cast<Int>(index)).offset;
+            const auto& uniform = Artifacts().program->getUniform(TProgramUniformIndex(index));
+            if (GlBlockIndexFromTProgram(uniform.index) < 0) return -1;
+            return uniform.offset;
         }
 
         // GL_UNIFORM_ARRAY_STRIDE: byte stride of an array member in a named block; 0 for a non-array
@@ -182,8 +229,8 @@ namespace MobileGL::MG_State::GLState {
         // generated SPIR-V lay the array out with std140 16-byte-rounded strides. MobileGL's UBO
         // layout is always std140, where every array element stride rounds up to a vec4.
         GLint GetActiveUniformArrayStride(Uint index) const {
-            const auto& uniform = m_program->getUniform(static_cast<Int>(index));
+            const auto& uniform = Artifacts().program->getUniform(TProgramUniformIndex(index));
-            if (uniform.index < 0) return -1;
+            if (GlBlockIndexFromTProgram(uniform.index) < 0) return -1;
             const glslang::TType* type = uniform.getType();
             if (type == nullptr || !type->isArray()) return 0;
             if (type->isMatrix()) {
@@ -202,13 +249,13 @@ namespace MobileGL::MG_State::GLState {
         // check suffices; the getUniformBlock() fallback is defensive for a config that instead leaves
         // an inheriting member's layoutMatrix == ElmNone.
         GLint GetActiveUniformIsRowMajor(Uint index) const {
-            const auto& uniform = m_program->getUniform(static_cast<Int>(index));
+            const auto& uniform = Artifacts().program->getUniform(TProgramUniformIndex(index));
-            if (uniform.index < 0) return 0;
+            if (GlBlockIndexFromTProgram(uniform.index) < 0) return 0;
             const glslang::TType* type = uniform.getType();
             if (type == nullptr || !type->isMatrix()) return 0;
             glslang::TLayoutMatrix layoutMatrix = type->getQualifier().layoutMatrix;
             if (layoutMatrix == glslang::ElmNone) {
-                layoutMatrix = m_program->getUniformBlock(uniform.index).getType()->getQualifier().layoutMatrix;
+                layoutMatrix = Artifacts().program->getUniformBlock(uniform.index).getType()->getQualifier().layoutMatrix;
             }
             return (layoutMatrix == glslang::ElmRowMajor) ? 1 : 0;
         }
@@ -220,13 +267,13 @@ namespace MobileGL::MG_State::GLState {
         // out as std140 (packed/shared are coerced), so this matches the offsets glslang reports. For
         // every GL 3.3 float matrix this evaluates to 16, independent of majorness.
         GLint GetActiveUniformMatrixStride(Uint index) const {
-            const auto& uniform = m_program->getUniform(static_cast<Int>(index));
+            const auto& uniform = Artifacts().program->getUniform(TProgramUniformIndex(index));
-            if (uniform.index < 0) return -1;
+            if (GlBlockIndexFromTProgram(uniform.index) < 0) return -1;
             const glslang::TType* type = uniform.getType();
             if (type == nullptr || !type->isMatrix()) return 0;
             glslang::TLayoutMatrix layoutMatrix = type->getQualifier().layoutMatrix;
             if (layoutMatrix == glslang::ElmNone) {
-                layoutMatrix = m_program->getUniformBlock(uniform.index).getType()->getQualifier().layoutMatrix;
+                layoutMatrix = Artifacts().program->getUniformBlock(uniform.index).getType()->getQualifier().layoutMatrix;
             }
             const bool rowMajor = (layoutMatrix == glslang::ElmRowMajor);
             const int strideVectorComponents = rowMajor ? type->getMatrixCols() : type->getMatrixRows();
@@ -238,50 +285,50 @@ namespace MobileGL::MG_State::GLState {
         }
 
         const glslang::TType* GetUniformTType(Uint location) const {
-            auto& uniform = m_program->getUniform(m_uniformIndexInTProgram[location]);
+            auto& uniform = Artifacts().program->getUniform(Artifacts().uniformIndexInTProgram[location]);
             return uniform.getType();
         }
 
         Bool IsUniformOpaqueAtLocation(Uint location) const { return GetUniformTType(location)->isOpaque(); }
 
         const String& GetUniformName(Uint location) const {
-            auto& uniform = m_program->getUniform(m_uniformIndexInTProgram[location]);
+            auto& uniform = Artifacts().program->getUniform(Artifacts().uniformIndexInTProgram[location]);
             return uniform.name;
         }
 
         const String& GetActiveUniformName(Uint index) const {
-            auto& uniform = m_program->getUniform(static_cast<Int>(index));
+            auto& uniform = Artifacts().program->getUniform(TProgramUniformIndex(index));
             return uniform.name;
         }
         // Sentinel for a uniform location without global-UBO backing storage (should not
         // survive linking: GenerateBinary falls back to tail-allocated scratch storage).
         static constexpr Uint kInvalidUniformOffset = ~0u;
-        Uint GetUniformOffset(Uint location) const { return m_uniformOffsets[location]; }
+        Uint GetUniformOffset(Uint location) const { return Artifacts().uniformOffsets[location]; }
         Uint GetUniformSizesInBytes(Uint location) const { return MG_Util::GetGLTypeSize(GetUniformType(location)); }
 
         Int GetAttributeLocation(const String& name) {
-            const auto it = std::find(m_attribs.begin(), m_attribs.end(), name);
+            const auto it = std::find(Artifacts().attribs.begin(), Artifacts().attribs.end(), name);
-            return (it == m_attribs.end()) ? -1 : (Int)std::distance(m_attribs.begin(), it);
+            return (it == Artifacts().attribs.end()) ? -1 : (Int)std::distance(Artifacts().attribs.begin(), it);
         }
         Uint32 GetActiveAttributeLocationMask() const {
             Uint32 mask = 0;
-            const SizeT count = std::min<SizeT>(m_attribs.size(), 32);
+            const SizeT count = std::min<SizeT>(Artifacts().attribs.size(), 32);
             for (SizeT index = 0; index < count; ++index) {
-                if (!m_attribs[index].empty()) {
+                if (!Artifacts().attribs[index].empty()) {
                     mask |= (1u << index);
                 }
             }
             return mask;
         }
         Uint32 GetActiveFragmentOutputLocationMask() const {
-            if (!m_program) {
+            if (!Artifacts().program) {
                 return 0;
             }
 
             Uint32 mask = 0;
-            const Int outputCount = m_program->getNumPipeOutputs();
+            const Int outputCount = Artifacts().program->getNumPipeOutputs();
             for (Int index = 0; index < outputCount; ++index) {
-                const Int location = static_cast<Int>(m_program->getPipeOutput(index).layoutLocation());
+                const Int location = static_cast<Int>(Artifacts().program->getPipeOutput(index).layoutLocation());
                 if (location >= 0 && location < 32) {
                     mask |= (1u << location);
                 }
@@ -289,47 +336,59 @@ namespace MobileGL::MG_State::GLState {
             return mask;
         }
         Int GetActiveFragmentOutputCount() const {
-            return m_program ? m_program->getNumPipeOutputs() : 0;
+            return Artifacts().program ? Artifacts().program->getNumPipeOutputs() : 0;
         }
         const String& GetActiveFragmentOutputName(Uint index) const {
-            MOBILEGL_ASSERT(m_program != nullptr, "ProgramObject::GetActiveFragmentOutputName: program is null");
+            MOBILEGL_ASSERT(Artifacts().program != nullptr, "ProgramObject::GetActiveFragmentOutputName: program is null");
-            MOBILEGL_ASSERT(index < static_cast<Uint>(m_program->getNumPipeOutputs()),
+            MOBILEGL_ASSERT(index < static_cast<Uint>(Artifacts().program->getNumPipeOutputs()),
                             "ProgramObject::GetActiveFragmentOutputName: index=%u out of range", index);
-            return m_program->getPipeOutput(static_cast<Int>(index)).name;
+            return Artifacts().program->getPipeOutput(static_cast<Int>(index)).name;
         }
         Int GetFragmentOutputLocation(Uint index) const {
-            MOBILEGL_ASSERT(m_program != nullptr, "ProgramObject::GetFragmentOutputLocation: program is null");
+            MOBILEGL_ASSERT(Artifacts().program != nullptr, "ProgramObject::GetFragmentOutputLocation: program is null");
-            MOBILEGL_ASSERT(index < static_cast<Uint>(m_program->getNumPipeOutputs()),
+            MOBILEGL_ASSERT(index < static_cast<Uint>(Artifacts().program->getNumPipeOutputs()),
                             "ProgramObject::GetFragmentOutputLocation: index=%u out of range",
                             index);
-            return static_cast<Int>(m_program->getPipeOutput(static_cast<Int>(index)).layoutLocation());
+            return static_cast<Int>(Artifacts().program->getPipeOutput(static_cast<Int>(index)).layoutLocation());
         }
         GLint GetActiveFragmentOutputArraySize(Uint index) const {
-            MOBILEGL_ASSERT(m_program != nullptr, "ProgramObject::GetActiveFragmentOutputArraySize: program is null");
+            MOBILEGL_ASSERT(Artifacts().program != nullptr, "ProgramObject::GetActiveFragmentOutputArraySize: program is null");
-            MOBILEGL_ASSERT(index < static_cast<Uint>(m_program->getNumPipeOutputs()),
+            MOBILEGL_ASSERT(index < static_cast<Uint>(Artifacts().program->getNumPipeOutputs()),
                             "ProgramObject::GetActiveFragmentOutputArraySize: index=%u out of range", index);
-            return m_program->getPipeOutput(static_cast<Int>(index)).size;
+            return Artifacts().program->getPipeOutput(static_cast<Int>(index)).size;
         }
         GLenum GetFragmentOutputType(Uint index) const {
-            MOBILEGL_ASSERT(m_program != nullptr, "ProgramObject::GetFragmentOutputType: program is null");
+            MOBILEGL_ASSERT(Artifacts().program != nullptr, "ProgramObject::GetFragmentOutputType: program is null");
-            MOBILEGL_ASSERT(index < static_cast<Uint>(m_program->getNumPipeOutputs()),
+            MOBILEGL_ASSERT(index < static_cast<Uint>(Artifacts().program->getNumPipeOutputs()),
                             "ProgramObject::GetFragmentOutputType: index=%u out of range",
                             index);
-            return m_program->getPipeOutput(static_cast<Int>(index)).glDefineType;
+            return Artifacts().program->getPipeOutput(static_cast<Int>(index)).glDefineType;
         }
-        GLenum GetAttribType(Uint index) const { return m_attribTypes[index]; }
+        GLenum GetAttribType(Uint index) const { return Artifacts().attribTypes[index]; }
-        const String& GetAttribName(Uint index) const { return m_attribs[index]; }
+        const String& GetAttribName(Uint index) const { return Artifacts().attribs[index]; }
-        GLenum GetActiveAttribType(Uint index) const { return m_program->getPipeInput(static_cast<Int>(index)).glDefineType; }
+        GLenum GetActiveAttribType(Uint index) const { return Artifacts().program->getPipeInput(static_cast<Int>(index)).glDefineType; }
-        GLint GetActiveAttribArraySize(Uint index) const { return m_program->getPipeInput(static_cast<Int>(index)).size; }
+        GLint GetActiveAttribArraySize(Uint index) const { return Artifacts().program->getPipeInput(static_cast<Int>(index)).size; }
-        const String& GetActiveAttribName(Uint index) const { return m_program->getPipeInput(static_cast<Int>(index)).name; }
+        // The Vulkan-semantics parse reflects the vertex builtins under their SPIR-V names;
-        void* MapUBO() { return m_globalUboScratch.data(); }
+        // GL must keep reporting the GL spellings (glGetActiveAttrib and the program-input
-        const void* GetUBOData() const { return m_globalUboScratch.data(); }
+        // resource queries enumerate builtins).
-        Uint GetUBOSize() const { return static_cast<Uint>(m_globalUboScratch.size()); }
+        static const String& NormalizeBuiltinPipeInputName(const String& name) {
+            static const String kGlVertexId = "gl_VertexID";
+            static const String kGlInstanceId = "gl_InstanceID";
+            if (name == "gl_VertexIndex") return kGlVertexId;
+            if (name == "gl_InstanceIndex") return kGlInstanceId;
+            return name;
+        }
+        const String& GetActiveAttribName(Uint index) const {
+            return NormalizeBuiltinPipeInputName(Artifacts().program->getPipeInput(static_cast<Int>(index)).name);
+        }
+        void* MapUBO() { return Artifacts().globalUboScratch.data(); }
+        const void* GetUBOData() const { return Artifacts().globalUboScratch.data(); }
+        Uint GetUBOSize() const { return static_cast<Uint>(Artifacts().globalUboScratch.size()); }
         // Content version of the CPU-side global-UBO shadow: writers bump it so backends
         // can skip re-uploading an unchanged UBO on every draw. ~0u is reserved as the
         // backends' "never uploaded" sentinel, so skip over it on wrap.
         Uint32 GetUBOContentVersion() const { return m_uboContentVersion; }
-        void MarkUBOContentDirty() {
+        void MarkUBOContentDirty() const {
             if (++m_uboContentVersion == ~0u) m_uboContentVersion = 0;
         }
         Uint32 GetBackendStateVersion() const { return m_backendStateVersion; }
@@ -370,20 +429,20 @@ namespace MobileGL::MG_State::GLState {
         }
 
         void SetUniformSamplerOrImageUnitIndex(Uint location, Int unit) {
-            if (location >= m_uniformSamplerOrImageUnitIndex.size() ||
+            if (location >= Artifacts().uniformSamplerOrImageUnitIndex.size() ||
-                m_uniformSamplerOrImageUnitIndex[location] == unit) {
+                Artifacts().uniformSamplerOrImageUnitIndex[location] == unit) {
                 return;
             }
-            m_uniformSamplerOrImageUnitIndex[location] = unit;
+            Artifacts().uniformSamplerOrImageUnitIndex[location] = unit;
             ++m_backendStateVersion;
         }
 
         Int GetUniformSamplerOrImageUnitIndex(Uint location) const {
-            return m_uniformSamplerOrImageUnitIndex[location];
+            return Artifacts().uniformSamplerOrImageUnitIndex[location];
         }
 
         Bool GetDeleteStatus() const { return m_deleteStatus; }
-        Bool GetLinkStatus() const { return m_linkStatus; }
+        Bool GetLinkStatus() const { return Artifacts().linkStatus; }
         // GL_PROGRAM_BINARY_RETRIEVABLE_HINT. MobileGL exposes no program binary format
         // (GL_NUM_PROGRAM_BINARY_FORMATS is 0), so the hint is pure state - which is all
         // ARB_get_program_binary requires of it.
@@ -397,24 +456,32 @@ namespace MobileGL::MG_State::GLState {
         // glProgramBinary always fails here (there is no format it could accept) and the
         // spec then requires the program's LINK_STATUS to read FALSE.
         void MarkLinkFailedByProgramBinary() {
-            ResetLinkArtifacts();
+            // Before anything reads m_artifacts: a pending link would otherwise publish its
-            m_infoLog = "No program binary format is supported.";
+            // (possibly successful) result over the failure this call is required to install
+            // - and Artifacts() below would be the thing that let it. Cancel-not-join: GL
+            // gives glProgramBinary no reason to wait for a link it is about to invalidate.
+            CancelLink();
+            BumpLinkObservableVersions();
+            ResetLinkArtifacts(Artifacts());
+            Artifacts().infoLog = "No program binary format is supported.";
         }
         Bool GetValidateStatus() const { return m_validateStatus; }
-        Int GetActiveAtomicCounterCount() const { return m_program->getNumAtomicCounters(); }
+        Int GetActiveAtomicCounterCount() const { return Artifacts().program->getNumAtomicCounters(); }
-        Int GetActiveAttributesCount() const { return m_program->getNumPipeInputs(); }
+        Int GetActiveAttributesCount() const { return Artifacts().program->getNumPipeInputs(); }
-        Int GetActiveUniformBlocksCount() const { return m_program->getNumUniformBlocks(); }
+        // GL-visible uniform blocks only: the synthesized MGL_GLOBAL_UBO the relaxed parse
-        GLuint GetComputeLocalSize(Uint dim) const { return m_program->getLocalSize(static_cast<Int>(dim)); }
+        // materializes for default-block uniforms is filtered out by DoReflection.
-        Int GetActiveAttributesMaxLength() const { return m_attribInNameMaxLength; }
+        Int GetActiveUniformBlocksCount() const { return static_cast<Int>(Artifacts().glBlockIndexToTProgram.size()); }
-        Int GetActiveUniformBlocksMaxNameLength() const { return m_uniformBlockNameMaxLength; }
+        GLuint GetComputeLocalSize(Uint dim) const { return Artifacts().program->getLocalSize(static_cast<Int>(dim)); }
+        Int GetActiveAttributesMaxLength() const { return Artifacts().attribInNameMaxLength; }
+        Int GetActiveUniformBlocksMaxNameLength() const { return Artifacts().uniformBlockNameMaxLength; }
         Uint GetUniformBlockIndex(const char* name) const {
-            auto it = m_uniformBlockIndexByName.find(name);
+            auto it = Artifacts().uniformBlockIndexByName.find(name);
-            if (it != m_uniformBlockIndexByName.end()) return it->second;
+            if (it != Artifacts().uniformBlockIndexByName.end()) return it->second;
             // Instances of an arrayed block are reflected as "Block[0]".."Block[N-1]";
             // a bare "Block" query resolves to the first instance per GL semantics.
             const String suffixedName = String(name) + "[0]";
-            it = m_uniformBlockIndexByName.find(suffixedName);
+            it = Artifacts().uniformBlockIndexByName.find(suffixedName);
-            if (it != m_uniformBlockIndexByName.end()) return it->second;
+            if (it != Artifacts().uniformBlockIndexByName.end()) return it->second;
             return 0xFFFFFFFFu; // GL_INVALID_INDEX
         }
         Bool IsActiveUniformBlock(Uint index) const {
@@ -427,11 +494,11 @@ namespace MobileGL::MG_State::GLState {
             // (like a std140 struct) occupies a vec4-rounded size, and that is what the
             // backend compiles: ES drivers reject draws whose bound UBO range is smaller
             // than the block (a block ending in ivec3 reported 12 while the driver needs 16).
-            return (m_program->getUniformBlock((Int)index).size + 15u) & ~15u;
+            return (Artifacts().program->getUniformBlock(Artifacts().glBlockIndexToTProgram[index]).size + 15u) & ~15u;
         }
 
         const String& GetUniformBlockName(Uint index) const {
-            auto& ubo = m_program->getUniformBlock((Int)index);
+            auto& ubo = Artifacts().program->getUniformBlock(Artifacts().glBlockIndexToTProgram[index]);
             return ubo.name;
         }
 
@@ -443,8 +510,8 @@ namespace MobileGL::MG_State::GLState {
             if (name.empty() || name.back() != ']') return index;
             const SizeT bracket = name.rfind('[');
             if (bracket == String::npos) return index;
-            const auto it = m_uniformBlockIndexByName.find(name.substr(0, bracket) + "[0]");
+            const auto it = Artifacts().uniformBlockIndexByName.find(name.substr(0, bracket) + "[0]");
-            if (it != m_uniformBlockIndexByName.end()) return it->second;
+            if (it != Artifacts().uniformBlockIndexByName.end()) return it->second;
             return index;
         }
 
@@ -455,31 +522,64 @@ namespace MobileGL::MG_State::GLState {
         Int GetUniformBlockActiveUniformCount(Uint index) const {
             const Int ownerIndex = static_cast<Int>(GetUniformBlockMemberOwnerIndex(index));
             Int count = 0;
-            for (Uint uniformIndex = 0; uniformIndex < m_activeUniformCount; ++uniformIndex) {
+            for (Uint uniformIndex = 0; uniformIndex < Artifacts().activeUniformCount; ++uniformIndex) {
                 if (GetActiveUniformBlockIndex(uniformIndex) == ownerIndex) ++count;
             }
             return count;
         }
 
         Bool IsUniformBlockReferencedByStage(Uint index, EShLanguage stage) const {
-            const auto& ubo = m_program->getUniformBlock((Int)index);
+            const auto& ubo = Artifacts().program->getUniformBlock(Artifacts().glBlockIndexToTProgram[index]);
             const auto stageMask = static_cast<EShLanguageMask>(1 << stage);
             return (ubo.stages & stageMask) != 0;
         }
 
         // Set by glUniformBlockBinding
         void SetUniformBlockBinding(Uint index, Uint binding) {
-            if (index >= m_uniformBlockBinding.size() || m_uniformBlockBinding[index] == static_cast<Int>(binding)) {
+            if (index >= Artifacts().uniformBlockBinding.size() || Artifacts().uniformBlockBinding[index] == static_cast<Int>(binding)) {
                 return;
             }
-            m_uniformBlockBinding[index] = static_cast<Int>(binding);
+            Artifacts().uniformBlockBinding[index] = static_cast<Int>(binding);
             ++m_backendStateVersion;
         }
 
-        Uint GetUniformBlockBinding(Uint index) const { return m_uniformBlockBinding[index]; }
+        Uint GetUniformBlockBinding(Uint index) const { return Artifacts().uniformBlockBinding[index]; }
 
-        Vector<Vector<unsigned>>& GetGeneratedSpirv() { return m_generatedSpirv; }
+        // Set by glShaderStorageBlockBinding, keyed by the block's GL name rather than by any
-        const Vector<Vector<unsigned>>& GetGeneratedSpirv() const { return m_generatedSpirv; }
+        // index. A shader storage block has THREE index spaces - the frontend interface-query
+        // enumeration, DirectVulkan's SPIR-V descriptor order and DirectGLES's real-driver
+        // order - and the name is the only coordinate all three agree on. Absent from the map
+        // means "never rebound", and the shader's declared binding still stands.
+        void SetShaderStorageBlockBinding(const String& blockName, Uint binding) {
+            Artifacts().shaderStorageBlockBinding[blockName] = static_cast<Int>(binding);
+        }
+        // -1 when the block has never been rebound. `blockName` is the interface-query
+        // spelling; an arrayed block's elements ("B[0]", "B[1]") are separate GL resources
+        // with separate bindings, so they are separate keys.
+        Int GetShaderStorageBlockBindingOverride(const String& blockName) const {
+            const auto it = Artifacts().shaderStorageBlockBinding.find(blockName);
+            if (it != Artifacts().shaderStorageBlockBinding.end()) return it->second;
+            // A backend that collapses an arrayed block down to one resource knows it only by
+            // the bare block name; answer that with element zero's binding.
+            const auto zeroth = Artifacts().shaderStorageBlockBinding.find(blockName + "[0]");
+            return zeroth != Artifacts().shaderStorageBlockBinding.end() ? zeroth->second : -1;
+        }
+        // Every rebinding recorded so far, for a backend that has to REPLAY them onto a
+        // driver program it just (re)built. Empty for the overwhelming majority of programs -
+        // check .empty() before doing any per-block work.
+        const UnorderedMap<String, Int>& GetShaderStorageBlockBindingOverrides() const {
+            return Artifacts().shaderStorageBlockBinding;
+        }
+
+        Vector<Vector<unsigned>>& GetGeneratedSpirv() { return Artifacts().generatedSpirv; }
+        const Vector<Vector<unsigned>>& GetGeneratedSpirv() const { return Artifacts().generatedSpirv; }
+
+        // The linked glslang reflection itself, for the ONE consumer that needs resource
+        // lists no typed getter above exposes: the GL program-interface query layer
+        // (MG_Impl/GLImpl/Program/ProgramInterface.cpp), which has to enumerate buffer
+        // blocks, buffer variables, atomic counters and per-stage reference masks. Null
+        // until a link has succeeded. Read through the join gate like everything else.
+        const glslang::TProgram* GetReflection() const { return Artifacts().program.get(); }
 
         Int GetShaderIndexByStage(ShaderStage stage) const {
             auto it = std::find_if(m_shaders.begin(), m_shaders.end(), [stage](const SharedPtr<ShaderObject>& shader) {
@@ -501,41 +601,202 @@ namespace MobileGL::MG_State::GLState {
             // layout captures into; see NeedsScatteredTransformFeedbackCapture.
             Uint32 packedOffsetBytes = 0;
         };
+
+        // ---- P1: everything a link PRODUCES, in one movable block ----
+        //
+        // The membership rule is mechanical, not editorial: this is exactly the field list
+        // ResetLinkArtifacts() clears (plus the four it forgot to - infoLog,
+        // linkedFragDataLocation/Index and the geometry strip-capture pair - which are just
+        // as much link output). Nothing else belongs here.
+        //
+        // Why a struct: once glLinkProgram runs on a worker (P1 stage 4) the worker writes
+        // its OWN LinkArtifacts and the GL thread publishes it with a single move, instead
+        // of thirty cross-thread field assignments. Until then this is a pure refactor.
+        //
+        // Access rule (invariant I5): the member below is private and reachable ONLY
+        // through ProgramObject::Artifacts(), which calls EnsureLinkJoined() first. That is
+        // what makes "every read of link output joins the pending link" a property the
+        // compiler checks rather than a review item - a new reader cannot spell the field
+        // without going through the gate.
+        struct LinkArtifacts {
+            SharedPtr<glslang::TProgram> program;
+            Vector<Vector<unsigned>> generatedSpirv;
+
+            // Attributes (Vertex in)
+            Vector<String> attribs;
+            Vector<GLenum> attribTypes;
+
+            // FragData (Frag out): the per-link snapshot of the explicit request maps.
+            UnorderedMap<String, Uint> linkedFragDataLocation;
+            UnorderedMap<String, Uint> linkedFragDataIndex;
+
+            // GL-facing index spaces (see the translation helpers above): GL active-uniform
+            // index <-> glslang TProgram uniform index, GL uniform-block index <-> TProgram
+            // block index. -1 marks a TProgram entry GL does not expose (dead default-block
+            // uniforms swept into MGL_GLOBAL_UBO by the relaxed parse, and that block itself).
+            Vector<Int> glUniformIndexToTProgram;
+            Vector<Int> tProgramUniformIndexToGl;
+            Vector<Int> glBlockIndexToTProgram;
+            Vector<Int> tProgramBlockIndexToGl;
+            // Per-link merged snapshot of the attached shaders' lexically extracted
+            // layout(location = N) default-block uniform qualifiers (the relaxed parse drops
+            // them from reflection; the DoReflection assigner restores them from here).
+            UnorderedMap<String, Int> linkedExplicitUniformLocations;
+            UnorderedMap<String, Uint> uniformLocations;
+            // Ordered by location,
+            // aka. uniformIndexInTProgram[loc] == "uniform index of TProgram at location `loc`"
+            Vector<Int> uniformIndexInTProgram;
+            // ditto. Will be set at glUniform1i
+            Vector<Int> uniformSamplerOrImageUnitIndex;
+            UnorderedMap<String, Uint> explicitOpaqueUniformBindings;
+
+            // Ordered by uniform block index
+            // index is DIFFERENT from binding!!!
+            //
+            // Let's define UniformBlockIndex == the order at glslang getUniformBlock()
+            // aka `i = glGetUniformBlockIndex(prog, "BlockName")` implies:
+            // `prog->getUniformBlock(i) == "BlockName"`
+            // These stuff are present for GL semantics, not for backend inspection
+            // These may change after-link (because GL spec decided to have `glUniformBlockBinding`)
+            UnorderedMap<String, Uint> uniformBlockIndexByName;
+            Vector<Int> uniformBlockBinding;
+            // glShaderStorageBlockBinding overrides, keyed by GL block name. See
+            // SetShaderStorageBlockBinding for why this one is by name and not by index.
+            UnorderedMap<String, Int> shaderStorageBlockBinding;
+
+            // Need to be reflected after linking of SPIR-V binary
+            Vector<Uint> uniformOffsets;
+            Vector<Uint> uniformSizesInBytes;
+            Vector<Uint8> globalUboScratch;
+
+            Uint activeUniformCount = 0;
+            Uint maxUniformLocation = 0;
+            Int uniformNameMaxLength = 0;
+            Int attribInNameMaxLength = 0;
+            Int uniformBlockNameMaxLength = 0;
+
+            String infoLog;
+            Bool linkStatus = false;
+
+            // Transform feedback: the linked snapshot (the request lives outside, on the
+            // GL-thread-owned side).
+            Vector<XfbVarying> xfbVaryings;
+            // The glTransformFeedbackVaryings request list exactly as this link consumed it,
+            // INCLUDING the gl_NextBuffer / gl_SkipComponentsN pseudo-varyings that
+            // xfbVaryings deliberately drops (they steer the capture layout and must never
+            // reach a backend's varying list). GL_TRANSFORM_FEEDBACK_VARYING enumerates the
+            // full request, pseudo-varyings and all, so the interface query needs its own copy.
+            Vector<String> xfbInterfaceNames;
+            Vector<Uint32> xfbStrides;
+            Vector<Uint32> gsStripTriangles;
+            Bool gsStripCaptureFixup = false;
+            GLenum gsInputPrimitive = GL_NONE;
+            GLenum xfbBufferMode = GL_INTERLEAVED_ATTRIBS;
+            Int xfbVaryingNameMaxLength = 0;
+            Bool xfbNeedsScatteredCapture = false;
+            Uint32 xfbPackedStride = 0;
+        };
+
+        // ---- artifacts-only helpers, shared with ProgramLinkTask ----
+        // Static and taking the block explicitly, because from stage 4 the link BODY needs
+        // them while its artifacts still live on the job node, not on any ProgramObject. The
+        // member overloads above are the same functions read through the join gate.
+
+        // Clears every field one link produces, EXCEPT infoLog, linkedFragDataLocation/Index
+        // and the geometry strip-capture pair. That exception is load-bearing: the callers
+        // that survive (glProgramBinary's mandated failure, and the link body's own mid-link
+        // aborts) write infoLog immediately AFTER calling here. Link()'s prologue does not
+        // use this at all - it assigns a whole default-constructed LinkArtifacts, where the
+        // ordering is explicit and nothing is exempt.
+        static void ResetLinkArtifacts(LinkArtifacts& artifacts);
+
+        static Bool IsValidUniformLocation(const LinkArtifacts& artifacts, Int location) {
+            if (location < 0 || location > static_cast<Int>(artifacts.maxUniformLocation)) return false;
+            if (static_cast<SizeT>(location) >= artifacts.uniformIndexInTProgram.size()) return false;
+            const Int uniformIndexInProgram = artifacts.uniformIndexInTProgram[location];
+            return uniformIndexInProgram != glslang::TQualifier::layoutLocationEnd &&
+                   uniformIndexInProgram >= 0 &&
+                   uniformIndexInProgram < static_cast<Int>(artifacts.tProgramUniformIndexToGl.size());
+        }
+
+        // Number of active array elements (GL_UNIFORM_SIZE / GL_ARRAY_SIZE); 1 for a non-array.
+        // glslang's TObjectReflection.size only carries the element count for a NON-block array; for
+        // a block array member it reports 1, so take the count from the TType, which is authoritative
+        // for both. GL 3.3 core uniforms are always sized. Takes a TProgram uniform index (the space
+        // the artifacts' uniformIndexInTProgram stores).
+        static GLint GetUniformArraySizeByTIndex(const LinkArtifacts& artifacts, Int tIndex) {
+            const auto& uniform = artifacts.program->getUniform(tIndex);
+            const glslang::TType* type = uniform.getType();
+            if (type != nullptr && type->isSizedArray()) {
+                return type->getOuterArraySize();
+            }
+            return uniform.size < 1 ? 1 : uniform.size;
+        }
+
+        // Blocks until a pending link has published its artifacts. Public because a few call
+        // sites have to join without reading anything - see the explicit-join list (J1-J8) in
+        // the P1 design. GL thread only.
+        void JoinLink() const { EnsureLinkJoined(); }
+
+        // Drops a link that is still in flight, without waiting for it. Called at the points
+        // where the pending link's result stops being the answer to "what did this program
+        // link to": a re-link supersedes it, glProgramBinary must force LINK_STATUS false,
+        // and a destroyed program has no observers left.
+        //
+        // Deliberately NOT called by the "takes effect at the next link" setters
+        // (glBindAttribLocation, glBindFragDataLocation(Indexed), glTransformFeedbackVaryings,
+        // glProgramParameteri) NOR by glAttachShader/glDetachShader. Every one of those is
+        // defined by GL to leave the CURRENT link result alone, and the pending link already
+        // snapshotted its own inputs at enqueue, so it is computing exactly the answer GL
+        // requires. Cancelling on any of them would make
+        //   glLinkProgram(p); <setter>; glGetProgramiv(p, GL_LINK_STATUS)
+        // report FALSE for a link that succeeded - and for the attach/detach pair it would
+        // additionally break glCreateShaderProgramv, which detaches immediately after linking.
+        void CancelLink();
+
+        // MUST NOT JOIN - this is what GL_COMPLETION_STATUS_KHR reads when the extension
+        // surface lands. "No job at all" counts as complete: there is nothing outstanding to
+        // wait for.
+        Bool IsLinkComplete() const { return m_pendingLink == nullptr || IsPendingLinkTerminal(); }
+
         void SetTransformFeedbackVaryings(Vector<String>&& names, GLenum bufferMode) {
             m_requestedXfbVaryings = Move(names);
             m_requestedXfbBufferMode = bufferMode;
         }
-        GLenum GetTransformFeedbackBufferMode() const { return m_xfbBufferMode; }
+        GLenum GetTransformFeedbackBufferMode() const { return Artifacts().xfbBufferMode; }
-        SizeT GetTransformFeedbackVaryingCount() const { return m_xfbVaryings.size(); }
+        SizeT GetTransformFeedbackVaryingCount() const { return Artifacts().xfbVaryings.size(); }
         const XfbVarying* GetTransformFeedbackVarying(SizeT index) const {
-            return index < m_xfbVaryings.size() ? &m_xfbVaryings[index] : nullptr;
+            return index < Artifacts().xfbVaryings.size() ? &Artifacts().xfbVaryings[index] : nullptr;
         }
-        const Vector<XfbVarying>& GetTransformFeedbackVaryings() const { return m_xfbVaryings; }
+        const Vector<XfbVarying>& GetTransformFeedbackVaryings() const { return Artifacts().xfbVaryings; }
+        // The GL_TRANSFORM_FEEDBACK_VARYING resource list: every name the last successful
+        // link was asked to capture, in request order, pseudo-varyings included.
+        const Vector<String>& GetTransformFeedbackInterfaceNames() const { return Artifacts().xfbInterfaceNames; }
         // Stride of one captured vertex in the given capture buffer slot.
         Uint32 GetTransformFeedbackStride(Uint32 bufferIndex) const {
-            return bufferIndex < m_xfbStrides.size() ? m_xfbStrides[bufferIndex] : 0;
+            return bufferIndex < Artifacts().xfbStrides.size() ? Artifacts().xfbStrides[bufferIndex] : 0;
         }
-        SizeT GetTransformFeedbackBufferCount() const { return m_xfbStrides.size(); }
+        SizeT GetTransformFeedbackBufferCount() const { return Artifacts().xfbStrides.size(); }
-        Int GetTransformFeedbackVaryingMaxLength() const { return m_xfbVaryingNameMaxLength; }
+        Int GetTransformFeedbackVaryingMaxLength() const { return Artifacts().xfbVaryingNameMaxLength; }
         // True when the capture layout uses gl_SkipComponents / gl_NextBuffer
         // (ARB_transform_feedback3), which no ES driver can express: it can only pack every
         // captured varying into one record with no gaps. A backend that captures through
         // such a driver has to capture into scratch storage and scatter the records into the
         // application's buffers itself, using packedOffsetBytes as the source offset and
         // (bufferIndex, offsetBytes, stride) as the destination.
-        Bool NeedsScatteredTransformFeedbackCapture() const { return m_xfbNeedsScatteredCapture; }
+        Bool NeedsScatteredTransformFeedbackCapture() const { return Artifacts().xfbNeedsScatteredCapture; }
         // Bytes one gap-free captured record occupies.
-        Uint32 GetTransformFeedbackPackedStride() const { return m_xfbPackedStride; }
+        Uint32 GetTransformFeedbackPackedStride() const { return Artifacts().xfbPackedStride; }
         // True when the capture stage is a triangle-strip geometry shader with a
         // statically-known emit sequence: the Vulkan capture order then needs the GL
         // odd-triangle vertex swap after EndTransformFeedback.
-        Bool HasGsTriangleStripCaptureFixup() const { return m_gsStripCaptureFixup; }
+        Bool HasGsTriangleStripCaptureFixup() const { return Artifacts().gsStripCaptureFixup; }
         // Triangles per strip, in emission order, for ONE geometry invocation.
-        const Vector<Uint32>& GetGsStripTriangles() const { return m_gsStripTriangles; }
+        const Vector<Uint32>& GetGsStripTriangles() const { return Artifacts().gsStripTriangles; }
         // GL_GEOMETRY_INPUT_TYPE of the linked geometry stage (GL_POINTS, GL_LINES,
         // GL_LINES_ADJACENCY, GL_TRIANGLES or GL_TRIANGLES_ADJACENCY), or GL_NONE when the
         // program has no geometry stage. Draws must present a compatible primitive type.
-        GLenum GetGeometryInputType() const { return m_gsInputPrimitive; }
+        GLenum GetGeometryInputType() const { return Artifacts().gsInputPrimitive; }
 
         Uint GetExternalIndex() const { return m_externalIndex; }
         // Globally-unique, never-reused id for this program object's lifetime. Unlike the GL
@@ -545,81 +806,74 @@ namespace MobileGL::MG_State::GLState {
         Uint64 GetLifetimeId() const { return m_lifetimeId; }
 
     private:
-        void ResetLinkArtifacts();
+        // ---- The one and only join gate for link output (P1 invariant I5) ----
-        void DoReflection();
+        // Blocks until a pending link has finished and its LinkArtifacts have been
-        // Resolves the requested transform feedback varyings against the linked
+        // published into m_artifacts. It exists so that the ~120 readers of link output are
-        // vertex stage; fails the link (GL semantics) on unknown or duplicate
+        // routed through it by the compiler rather than by review: m_artifacts is private
-        // names or exceeded capture limits.
+        // and Artifacts() is the only spelling that reaches it.
-        Bool ResolveTransformFeedbackVaryings();
+        //
-        void ResolveGsTriangleStripCapture(const glslang::TIntermediate* captureIntermediate);
+        // The fast path - no pending link - is one predictable branch and stays inline: it
-        void GenerateBinary();
+        // runs on every Artifacts() read (~1200 call sites project-wide) and the project
-        void WaitUntilGenerationCompleted() const;
+        // never builds with LTO (MOBILEGL_ENABLE_LTO=OFF), so an out-of-line body would be a
+        // real cross-TU call at every one of them. The blocking half is out of line.
+        void EnsureLinkJoined() const {
+            if (m_pendingLink) JoinPendingLink();
+        }
+        void JoinPendingLink() const;
+        // ProgramLinkTask is incomplete here, so IsLinkComplete()'s non-joining peek at the
+        // node's state goes through this out-of-line helper.
+        Bool IsPendingLinkTerminal() const;
+
+        LinkArtifacts& Artifacts() {
+            EnsureLinkJoined();
+            return m_artifacts;
+        }
+        const LinkArtifacts& Artifacts() const {
+            EnsureLinkJoined();
+            return m_artifacts;
+        }
+
+        // GL-thread-only companion to ResetLinkArtifacts (see its definition). Const because
+        // the publish half of the join calls it; see the mutable counters below.
+        void BumpLinkObservableVersions() const;
         void AddDefaultFragmentShaderIfMissing();
-        Bool ValidateFragmentOutputLocations();
 
         static Uint64 AllocateLifetimeId();
 
+        // ---- GL-thread-owned state: never joins ----
+        // Most of this is never produced by a link at all. The three version counters
+        // (m_backendStateVersion / m_uboContentVersion / m_linkVersion) ARE
+        // link-observable, but they are bumped exclusively on the GL thread
+        // (BumpLinkObservableVersions in Link()'s prologue and glProgramBinary's
+        // failure path) - the link BODY, which stage 4 moves to a worker, never
+        // writes them.
         const Uint m_externalIndex = 0;
         const Uint64 m_lifetimeId = 0;
+        // The attach lists are mutated only in Link()'s GL-thread prologue, which is why
+        // glGetAttachedShaders / GL_ATTACHED_SHADERS / the orphan-shader sweep need no join.
         Vector<SharedPtr<ShaderObject>> m_shaders;
         Vector<SharedPtr<ShaderObject>> m_detachedShaders; // Store detached shaders and remove on next link
 
-        SharedPtr<glslang::TProgram> m_program;
+        // Link INPUTS (all "take effect at the next link" per GL): glBindAttribLocation,
-
+        // glBindFragDataLocation(Indexed), glTransformFeedbackVaryings, and the draw-buffer
-        Vector<Vector<unsigned>> m_generatedSpirv;
+        // count stamped in by the entry point. A pending link snapshots these at enqueue.
-
-        // Attributes (Vertex in)
         UnorderedMap<String, Uint> m_explicitAttribLocations;
-        Vector<String> m_attribs;
-        Vector<GLenum> m_attribTypes;
-
-        // FragData (Frag out)
         UnorderedMap<String, Uint> m_explicitFragDataLocation;
-        UnorderedMap<String, Uint> m_linkedFragDataLocation;
         // Dual-source blend color index per output name (glBindFragDataLocationIndexed); snapshotted
         // into the linked map at link time, like the location maps above.
         UnorderedMap<String, Uint> m_explicitFragDataIndex;
-        UnorderedMap<String, Uint> m_linkedFragDataIndex;
         Int m_maxFragmentOutputColorNumber = 8;
+        Vector<String> m_requestedXfbVaryings;
+        GLenum m_requestedXfbBufferMode = GL_INTERLEAVED_ATTRIBS;
 
-        // Uniforms
-        UnorderedMap<String, Uint> m_uniformLocations;
-        // Ordered by location,
-        // aka. m_uniformIndexInTProgram[loc] == "uniform index of TProgram at location `loc`"
-        Vector<Int> m_uniformIndexInTProgram;
-        // ditto. Will be set at glUniform1i
-        Vector<Int> m_uniformSamplerOrImageUnitIndex;
-        UnorderedMap<String, Uint> m_explicitOpaqueUniformBindings;
-
-        // Ordered by uniform block index
-        // index is DIFFERENT from binding!!!
-        //
-        // Let's define UniformBlockIndex == the order at glslang getUniformBlock()
-        // aka `i = glGetUniformBlockIndex(prog, "BlockName")` implies:
-        // `prog->getUniformBlock(i) == "BlockName"`
-        // These stuff are present for GL semantics, not for backend inspection
-        // These may change after-link (because GL spec decided to have `glUniformBlockBinding`)
-        UnorderedMap<String, Uint> m_uniformBlockIndexByName;
-        Vector<Int> m_uniformBlockBinding;
-
-        // Need to be reflected after linking of SPIR-V binary
-        Vector<Uint> m_uniformOffsets;
-        Vector<Uint> m_uniformSizesInBytes;
-        Vector<Uint8> m_globalUboScratch;
-
-        Uint m_activeUniformCount = 0;
-        Uint m_maxUniformLocation = 0;
-        Int m_uniformNameMaxLength = 0;
-        Int m_attribInNameMaxLength = 0;
-        Int m_uniformBlockNameMaxLength = 0;
-
-        String m_infoLog;
         Bool m_deleteStatus = false;
-        Bool m_linkStatus = false;
         Bool m_binaryRetrievableHint = false;
         Bool m_separable = false;
         Bool m_validateStatus = true;
-        Uint32 m_backendStateVersion = 0;
+        // Mutable, like m_artifacts and for the same reason: publishing a pending link is a
+        // READ-side operation (the first gated getter is what pulls the result in), and the
+        // publish has to bump these. Still GL-thread-only - a worker never touches them.
+        mutable Uint32 m_backendStateVersion = 0;
 
         // Backend-owned content-hash memo (see GetBackendHashMemo): valid only while
         // m_backendStateVersion matches. Several slots, not one: a backend may resolve the same
@@ -635,20 +889,20 @@ namespace MobileGL::MG_State::GLState {
         mutable Array<BackendHashMemoSlot, kBackendHashMemoSlotCount> m_backendHashMemoSlots{};
         mutable SizeT m_backendHashMemoNextSlot = 0;
         mutable Uint32 m_backendHashMemoVersion = ~0u;
-        Uint32 m_uboContentVersion = 0;
+        mutable Uint32 m_uboContentVersion = 0;
-        Uint32 m_linkVersion = 0;
+        mutable Uint32 m_linkVersion = 0;
 
-        // Transform feedback: request (applies at next link) and linked snapshot.
+        // ---- Link OUTPUT ----
-        Vector<String> m_requestedXfbVaryings;
+        // Written by the link and by the post-link setters GL allows (glUniform1i's sampler
-        GLenum m_requestedXfbBufferMode = GL_INTERLEAVED_ATTRIBS;
+        // unit, glUniformBlockBinding). Reachable only through Artifacts(); see LinkArtifacts.
-        Vector<XfbVarying> m_xfbVaryings;
+        //
-        Vector<Uint32> m_xfbStrides;
+        // Mutable because publishing is a READ-side operation: a const getter has to be able
-        Vector<Uint32> m_gsStripTriangles;
+        // to settle an outstanding link before answering it.
-        Bool m_gsStripCaptureFixup = false;
+        mutable LinkArtifacts m_artifacts;
-        GLenum m_gsInputPrimitive = GL_NONE;
+
-        GLenum m_xfbBufferMode = GL_INTERLEAVED_ATTRIBS;
+        // The link job, from enqueue until the first observable read pulls its result. Null
-        Int m_xfbVaryingNameMaxLength = 0;
+        // means m_artifacts is already the answer - which is the state every reader outside
-        Bool m_xfbNeedsScatteredCapture = false;
+        // the pending window sees, and the whole reason the gate above is one branch.
-        Uint32 m_xfbPackedStride = 0;
+        mutable SharedPtr<ProgramLinkTask> m_pendingLink;
     };
 } // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ProgramState.cpp

@@ -11,7 +11,7 @@
 namespace MobileGL::MG_State::GLState {
     Uint ProgramState::CreateProgram() {
         Uint programId = 0;
-        m_programIndexGenerator.Generate(1, &programId);
+        m_programShaderNameGenerator.Generate(1, &programId);
         EnsureIndexAvail(programId, m_programObjects);
         auto programObject = MakeShared<ProgramObject>(programId);
         if (programObject == nullptr) return 0;
@@ -39,11 +39,19 @@ namespace MobileGL::MG_State::GLState {
 
     void ProgramState::DestroyProgramSlot(const Uint program) {
         auto& programObject = m_programObjects[program];
+        // P1 join site J4/J5 (glDeleteProgram, and the deferred destroy UseProgram performs
+        // when a deletion-flagged program stops being current). The program's name is about
+        // to go, so nothing can observe its link any more: cancel-not-join, so a delete never
+        // blocks the GL thread on a worker. Explicit rather than left to ~ProgramObject,
+        // because the reset below only destroys the object if this table held the last
+        // reference - a program still bound as current, or still referenced by a pipeline,
+        // outlives it, and its link should stop the moment the name does.
+        programObject->CancelLink();
         // Snapshot the attachments: deleting the program is a detach point for shaders
         // that were flagged with glDeleteShader while still attached.
         const Vector<SharedPtr<ShaderObject>> attachedShaders = programObject->GetAttachedShaders();
         programObject.reset();
-        m_programIndexGenerator.Delete(program);
+        m_programShaderNameGenerator.Delete(program);
         for (const auto& shader : attachedShaders) {
             const Uint shaderName = shader->GetExternalIndex();
             if (CheckIndexAvail(shaderName, m_shaderObjects) && m_shaderObjects[shaderName] == shader) {
@@ -77,9 +85,10 @@ namespace MobileGL::MG_State::GLState {
 
     Uint ProgramState::CreateShader(ShaderStage stage) {
         Uint shaderId = 0;
-        m_shaderIndexGenerator.Generate(1, &shaderId);
+        m_programShaderNameGenerator.Generate(1, &shaderId);
         EnsureIndexAvail(shaderId, m_shaderObjects);
-        auto shaderObject = MakeShared<ShaderObject>(stage, shaderId);
+        auto shaderObject =
+            MakeShared<ShaderObject>(stage, shaderId, m_shaderPreprocessCache, m_shaderCompileAdoptionMap);
         if (shaderObject == nullptr) return 0;
         m_shaderObjects[shaderId] = shaderObject;
         return shaderId;
@@ -91,6 +100,31 @@ namespace MobileGL::MG_State::GLState {
         return m_shaderObjects[shader];
     }
 
+    void ProgramState::JoinAllPendingWork() {
+        // Programs first: a link joins the compiles it depends on, so the shader pass that
+        // follows finds most of them already settled. The reverse order would be correct but
+        // would wait on each compile twice - once here, once inside the link's own prologue.
+        //
+        // A copy of each slot rather than a reference into the vector, and an index rather
+        // than an iterator: publishing a link replays deferred diagnostics, which reach
+        // pGLContext->RecordError. That does not touch these tables today, but it is a sink
+        // that can grow, and a reallocation underneath this loop would be a use-after-free
+        // that only shows up on the one GL call that walks the whole table. The copy costs a
+        // refcount bump on a path a mode switch takes at most once.
+        for (SizeT i = 0; i < m_programObjects.size(); ++i) {
+            const SharedPtr<ProgramObject> program = m_programObjects[i];
+            if (program) program->JoinLink();
+        }
+        for (SizeT i = 0; i < m_shaderObjects.size(); ++i) {
+            const SharedPtr<ShaderObject> shader = m_shaderObjects[i];
+            if (shader) shader->JoinCompile();
+        }
+        // The currently-used program is reachable through m_programObjects unless
+        // glDeleteProgram already freed its slot while it stayed current. Nothing else holds
+        // a GL-visible name for it, but a draw would still join it, so settle it here too.
+        if (m_currentProgram) m_currentProgram->JoinLink();
+    }
+
     void ProgramState::MarkShaderObjectForDeletion(Uint shader) {
         if (!CheckIndexAvail(shader, m_shaderObjects)) return;
         auto& shaderObject = m_shaderObjects[shader];
@@ -120,8 +154,15 @@ namespace MobileGL::MG_State::GLState {
         auto& shaderObject = m_shaderObjects[shader];
         if (shaderObject == nullptr || !shaderObject->GetDeleteStatus()) return;
         if (ShaderHasGLVisibleAttachment(shaderObject)) return;
+        // The name is about to go, so nothing can observe this shader's compile through THIS
+        // object any more and a job still in flight for it is pure waste - unless another
+        // shader object adopted the same node (stage 6) or a pending link pinned it, which is
+        // exactly what ReleaseCompileNode weighs before it cancels anything. Cancel-not-join
+        // either way: the job owns its inputs, so dropping the object out from under it is
+        // safe and the GL thread never blocks on a delete.
+        shaderObject->ReleaseCompileNode();
         shaderObject.reset();
-        m_shaderIndexGenerator.Delete(shader);
+        m_programShaderNameGenerator.Delete(shader);
     }
 
     Bool ProgramState::ValidateShaderObject(Uint shader) const {

MobileGL/MG_State/GLState/ProgramState/ProgramState.h

@@ -10,6 +10,8 @@
 #include <Includes.h>
 #include <MG_Util/Miscellany/IndexGenerator.h>
 #include "ProgramObject.h"
+#include "ShaderCompileAdoptionMap.h"
+#include "ShaderPreprocessCache.h"
 
 namespace MobileGL::MG_State::GLState {
     class ProgramState {
@@ -33,6 +35,29 @@ namespace MobileGL::MG_State::GLState {
 
         const SharedPtr<ProgramObject>& GetCurrentProgram() const { return m_currentProgram; }
 
+        // Joins every outstanding compile and link this context still owns, publishing each
+        // one's artifacts through the ordinary gates. The single caller is
+        // glMaxShaderCompilerThreadsKHR(0): GL_KHR_parallel_shader_compile requires a zero
+        // count to leave nothing in flight, so that every subsequent
+        // GL_COMPLETION_STATUS_KHR reads GL_TRUE.
+        //
+        // NOT a teardown path and NOT ShaderCompilePool::StopAndDrain(): the pool keeps its
+        // threads and stays usable, because a later nonzero count has to bring asynchronous
+        // compilation straight back. Nodes belonging to objects this context has already
+        // dropped are not joined - nothing can observe them, and waiting on them would make
+        // a GL call's cost depend on garbage.
+        void JoinAllPendingWork();
+
+        // P0b layer 2. Exposed for tests and diagnostics; the GL frontend never touches it
+        // directly - shader objects reach it through the pointer they are handed at
+        // CreateShader().
+        ShaderPreprocessCache& GetShaderPreprocessCache() { return *m_shaderPreprocessCache; }
+
+        // P1 stage 6, same deal: exposed for tests and diagnostics only. Its adoption counter
+        // is the one number that says how many glCompileShader calls this context turned into
+        // no work at all; nothing in the GL frontend branches on it.
+        ShaderCompileAdoptionMap& GetShaderCompileAdoptionMap() { return *m_shaderCompileAdoptionMap; }
+
     private:
         Bool ShaderHasGLVisibleAttachment(const SharedPtr<ShaderObject>& shaderObject) const;
         // Frees the name slot and releases orphaned attached shaders; the immediate half
@@ -52,10 +77,24 @@ namespace MobileGL::MG_State::GLState {
             vec.resize(idx + 1);
         }
 
-        IndexGenerator<Uint> m_programIndexGenerator;
+        // Programs and shaders share one GL name space (GL 3.3 core 2.11: a shader
-        Vector<SharedPtr<ProgramObject>> m_programObjects;
+        // name passed where a program is expected must be recognized as a shader and
+        // rejected with INVALID_OPERATION, and vice versa). One generator for both
+        // object kinds keeps the names disjoint; the object tables stay separate.
+        IndexGenerator<Uint> m_programShaderNameGenerator;
 
-        IndexGenerator<Uint> m_shaderIndexGenerator;
+        // P0b layer 2: every shader object created here is handed shared ownership of this
+        // cache, so its lifetime no longer depends on member destruction order (P1: an
+        // in-flight compile job may outlive the context). The FIRST-member declaration is
+        // kept anyway - it costs nothing and documents the intent.
+        SharedPtr<ShaderPreprocessCache> m_shaderPreprocessCache = MakeShared<ShaderPreprocessCache>();
+        // P1 stage 6: the GL-thread-only index of adoptable compile nodes. Shared ownership
+        // for the same reason as the cache above - a ShaderObject held by a ProgramObject can
+        // outlive these tables, and its destructor releases a node - though unlike the cache
+        // no worker ever sees this one, which is why it carries no lock.
+        SharedPtr<ShaderCompileAdoptionMap> m_shaderCompileAdoptionMap = MakeShared<ShaderCompileAdoptionMap>();
+
+        Vector<SharedPtr<ProgramObject>> m_programObjects;
         Vector<SharedPtr<ShaderObject>> m_shaderObjects;
 
         SharedPtr<ProgramObject> m_currentProgram;

MobileGL/MG_State/GLState/ProgramState/ShaderCompileAdoptionMap.cpp

@@ -0,0 +1,90 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderCompileAdoptionMap.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "ShaderCompileAdoptionMap.h"
+
+#include "ShaderCompileTask.h"
+
+namespace MobileGL::MG_State::GLState {
+    SharedPtr<ShaderCompileTask> ShaderCompileAdoptionMap::FindAdoptable(const ShaderStage stage,
+                                                                        const Uint64 sourceHash, const String& source,
+                                                                        const Uint64 envFingerprint) {
+        const ShaderSourceKey key{.stage = stage,
+                                  .sourceHash = sourceHash,
+                                  .sourceLength = source.length(),
+                                  .envFingerprint = envFingerprint};
+
+        const auto it = m_entries.find(key);
+        if (it == m_entries.end()) return nullptr;
+
+        SharedPtr<ShaderCompileTask> node = it->second.lock();
+        // Expired (every shader object that held it has released it), settled as Cancelled
+        // (the enqueue lost a race with teardown, or the body threw), or CANCELLATION
+        // REQUESTED but not yet settled (a releaser fired Cancel() while a worker was still
+        // inside RunBody(), so the node is stuck at Running until the body returns - see
+        // JobNode::Run: once m_cancelled is set, the node is DOOMED to end up Cancelled no
+        // matter how the body finishes, it just has not gotten there yet). All three can
+        // never publish artifacts a caller may rely on, so all three are misses. Only the
+        // first two are dead weight worth pruning from the index here - a cancellation-
+        // requested-but-still-running node is still reachable from its own (about to
+        // release) ShaderObject and will get pruned once it actually settles, so leave the
+        // entry alone and just refuse to hand this node out.
+        if (!node || node->IsCancelled()) {
+            m_entries.erase(it);
+            return nullptr;
+        }
+        if (node->IsCancellationRequested()) return nullptr;
+
+        // Never let correctness ride on a 64-bit hash. Lengths already matched (they are part
+        // of the key), so this is a plain memcmp - and it is the ONLY thing that authorizes
+        // two GL shader names to share one compile.
+        if (*node->source != source) return nullptr;
+
+        ++m_adoptionCount;
+        return node;
+    }
+
+    void ShaderCompileAdoptionMap::Register(const SharedPtr<ShaderCompileTask>& node) {
+        if (!node) return;
+
+        SweepIfCrowded();
+        // operator[] rather than a find/insert pair: an existing entry for this key is either
+        // a re-registration of the same source (the previous node expired or was cancelled)
+        // or an astronomically rare hash collision. The newcomer wins in both cases.
+        m_entries[ShaderSourceKey{.stage = node->stage,
+                                  .sourceHash = node->sourceHash,
+                                  .sourceLength = node->source->length(),
+                                  .envFingerprint = node->env->fingerprint}] = node;
+    }
+
+    void ShaderCompileAdoptionMap::Clear() {
+        m_entries.clear();
+        m_sweepThreshold = kMinSweepThreshold;
+    }
+
+    void ShaderCompileAdoptionMap::SweepIfCrowded() {
+        if (m_entries.size() < m_sweepThreshold) return;
+
+        // Collect first, erase after: FastSTL::unordered_map is open-addressed, so erasing
+        // through an iterator that the same loop is still advancing is not worth reasoning
+        // about on a path this cold.
+        Vector<ShaderSourceKey> dead;
+        for (const auto& entry : m_entries) {
+            const SharedPtr<ShaderCompileTask> node = entry.second.lock();
+            if (!node || node->IsCancelled()) dead.push_back(entry.first);
+        }
+        for (const ShaderSourceKey& key : dead) {
+            m_entries.erase(key);
+        }
+
+        // Amortization: after a sweep the map holds exactly the nodes still reachable from
+        // some shader object, so letting it double before the next sweep makes the whole
+        // scheme O(1) per Register() while keeping the map O(live nodes).
+        m_sweepThreshold = std::max(kMinSweepThreshold, m_entries.size() * 2);
+    }
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ShaderCompileAdoptionMap.h

@@ -0,0 +1,97 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderCompileAdoptionMap.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+#include <MG_State/GLState/ProgramState/ShaderSourceKey.h>
+
+namespace MobileGL::MG_State::GLState {
+    class ShaderCompileTask;
+
+    // P1 stage 6: the per-context index of compile job nodes that a NEW shader object may
+    // adopt instead of enqueueing a duplicate of.
+    //
+    // Why it is not the P0b preprocess cache. That cache only helps once a compile has
+    // FINISHED - it memoizes the source-only half of the pipeline, and a worker consults it
+    // from inside the job body. Under asynchronous compilation the dominant shape is
+    // different: a shaderpack load hands N different shader objects byte-identical source
+    // within the same GL-thread burst (measured across bsl/complementary/bliss, ~21% of all
+    // Compile() calls are such cross-object duplicates), and all N are enqueued before any of
+    // them completes. Every one of those workers then misses the cache, runs the whole
+    // pipeline, and races the others to insert the same entry. This map closes that window on
+    // the GL thread, at enqueue: the second object through takes the FIRST object's node.
+    //
+    // What "adopt" means: the two shader objects end up holding the same SharedPtr in their
+    // m_compiled. They are two distinct GL names with two distinct info-log/COMPILE_STATUS
+    // queries, but both queries read one set of artifacts - which is exactly right, because
+    // the pipeline is a pure function of the key below and the full source text. Nothing is
+    // copied and no worker ever waits (P1 invariant I4 is untouched: this only ever REMOVES
+    // work from the pool). The single consume-once resource, the glslang parse, is already
+    // guarded for sharing by ShaderCompileTask::ClaimParsedShader's CAS, which stage 4 built
+    // for exactly this shape - one node, several links.
+    //
+    // ---- Threading: GL thread only, and therefore lock-free ----
+    // Every entry point below is reached from glCompileShader (ShaderObject::Compile) and
+    // from nowhere else. That is one GL entry point on the application's context thread, so
+    // the map needs no mutex, unlike the preprocess cache which several workers hit at once.
+    // The weak pointers are the ONLY thing this class stores, precisely so it can never keep
+    // a node - or the artifacts a node owns - alive past its last real holder.
+    //
+    // ---- Lifetime and pruning ----
+    // WeakPtr, never SharedPtr: the map is an index, not an owner. An entry whose node has
+    // been released by every shader object simply expires, and a node that was CANCELLED
+    // carries no result at all, so both are treated as misses and pruned where they are
+    // found. Pruning is otherwise amortized: Register() sweeps the whole map whenever it has
+    // grown past twice its size at the last sweep, which bounds the map at O(live nodes)
+    // without a per-call cost.
+    class ShaderCompileAdoptionMap {
+    public:
+        // Never sweep below this: a shaderpack burst is a few hundred distinct sources, and
+        // an entry is a key plus a weak pointer.
+        static constexpr SizeT kMinSweepThreshold = 256;
+
+        // The adoptable node for this exact source under this exact environment, or null.
+        //
+        // A hit is honored only after the FULL source text has been compared byte for byte
+        // against the candidate node's own snapshot: the hash in the key is a lookup
+        // accelerator, never the answer (ShaderSourceKey). A node that has settled as
+        // Cancelled is never handed out - it published nothing, so adopting it would give the
+        // new object a compile that can never report anything but GL_FALSE. Nor is a node
+        // whose cancellation has merely been REQUESTED but not yet settled (still Running,
+        // with IsCancellationRequested() true): JobNode::Run forces such a node to Cancelled
+        // the moment its body returns regardless of how the body finished, so it is already
+        // doomed and handing it out would just move the same GL_FALSE-with-no-log outcome to
+        // a second, unrelated shader object.
+        //
+        // A COMPLETED node is adoptable, and deliberately so: the new object gets the right
+        // answer for zero work, which is the same deal the P0b cache offers one layer down.
+        SharedPtr<ShaderCompileTask> FindAdoptable(ShaderStage stage, Uint64 sourceHash, const String& source,
+                                                   Uint64 envFingerprint);
+
+        // Indexes `node` as the adoptable one for its key. A key already present is
+        // overwritten: the newcomer is at least as fresh as whatever was there, and one entry
+        // per key keeps this a plain map.
+        void Register(const SharedPtr<ShaderCompileTask>& node);
+
+        void Clear();
+
+        // ---- diagnostics only; nothing in the GL frontend branches on these ----
+        // Monotonic count of nodes handed out by FindAdoptable, i.e. of glCompileShader calls
+        // that did NOT enqueue a job because an equivalent one already existed. Tests read it
+        // as a delta across a burst.
+        Uint64 GetAdoptionCount() const { return m_adoptionCount; }
+        SizeT GetEntryCount() const { return m_entries.size(); }
+
+    private:
+        void SweepIfCrowded();
+
+        UnorderedMap<ShaderSourceKey, WeakPtr<ShaderCompileTask>, ShaderSourceKeyHasher> m_entries;
+        SizeT m_sweepThreshold = kMinSweepThreshold;
+        Uint64 m_adoptionCount = 0;
+    };
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ShaderCompileTask.cpp

@@ -0,0 +1,338 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderCompileTask.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "ShaderCompileTask.h"
+
+#include <MG_Util/Converters/MGToGL/ProgramEnumConverter.h>
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+#include <MG_Util/ShaderTranspiler/ShaderSourceProcessor.h>
+#include <MG_Util/ShaderTranspiler/Types.h>
+
+#include <glslang/Include/PoolAlloc.h>
+
+#include <charconv>
+
+namespace {
+    struct ComputeLocalSize {
+        MobileGL::Uint x = 1;
+        MobileGL::Uint y = 1;
+        MobileGL::Uint z = 1;
+        bool declared = false;
+    };
+
+    static MobileGL::String StripGlslComments(const MobileGL::String& source) {
+        MobileGL::String result;
+        result.reserve(source.length());
+
+        bool inLineComment = false;
+        bool inBlockComment = false;
+        for (MobileGL::SizeT i = 0; i < source.length(); ++i) {
+            if (inLineComment) {
+                if (source[i] == '\n') {
+                    inLineComment = false;
+                    result.push_back(source[i]);
+                } else {
+                    result.push_back(' ');
+                }
+                continue;
+            }
+
+            if (inBlockComment) {
+                if (source[i] == '*' && i + 1 < source.length() && source[i + 1] == '/') {
+                    inBlockComment = false;
+                    result.append("  ");
+                    ++i;
+                } else {
+                    result.push_back(source[i] == '\n' ? '\n' : ' ');
+                }
+                continue;
+            }
+
+            if (source[i] == '/' && i + 1 < source.length()) {
+                if (source[i + 1] == '/') {
+                    inLineComment = true;
+                    result.append("  ");
+                    ++i;
+                    continue;
+                }
+                if (source[i + 1] == '*') {
+                    inBlockComment = true;
+                    result.append("  ");
+                    ++i;
+                    continue;
+                }
+            }
+
+            result.push_back(source[i]);
+        }
+
+        return result;
+    }
+
+    // Hoisted out of ParseComputeLocalSize: constructing a std::regex costs far more than
+    // running it over a small source, and it was being rebuilt on every compute compile. A
+    // const regex carries no mutable state, so sharing one instance across workers is safe.
+    static const std::regex kComputeLocalSizePattern(R"(local_size_([xyz])\s*=\s*([0-9]+))");
+
+    static ComputeLocalSize ParseComputeLocalSize(const MobileGL::String& source) {
+        ComputeLocalSize localSize;
+        const MobileGL::String uncommentedSource = StripGlslComments(source);
+
+        for (std::sregex_iterator it(uncommentedSource.begin(), uncommentedSource.end(), kComputeLocalSizePattern),
+             end;
+             it != end; ++it) {
+            const char axis = (*it)[1].str()[0];
+            // The [0-9]+ capture is unbounded, so `local_size_x = 99999999999999999999999`
+            // is a legal match. std::stoull would throw std::out_of_range on it and let the
+            // exception escape glCompileShader; std::from_chars reports the overflow instead.
+            // An overflowing literal saturates to UINT_MAX, which the device-limit check
+            // below rejects anyway - the same verdict a non-overflowing huge value gets.
+            const MobileGL::String digits = (*it)[2].str();
+            unsigned long long value = 0;
+            const std::from_chars_result parsed =
+                std::from_chars(digits.data(), digits.data() + digits.size(), value);
+            const MobileGL::Uint clampedValue = (parsed.ec != std::errc() || value > UINT_MAX)
+                                                    ? UINT_MAX
+                                                    : static_cast<MobileGL::Uint>(value);
+
+            // TODO: Replace this literal layout scanner with parser/AST-backed validation so expressions and
+            // specialization-id layouts are handled consistently with glslang.
+            localSize.declared = true;
+            if (axis == 'x') {
+                localSize.x = clampedValue;
+            } else if (axis == 'y') {
+                localSize.y = clampedValue;
+            } else {
+                localSize.z = clampedValue;
+            }
+        }
+
+        return localSize;
+    }
+
+    // The device limits come from the CompileEnv snapshot, never from a live driver query.
+    // GL_MAX_COMPUTE_WORK_GROUP_SIZE is a real GLES call on the DirectGLES backend: issued
+    // off the context thread it would silently no-op and turn a legal local_size_z into
+    // COMPILE_STATUS=FALSE. CaptureCompileEnv() issues it once, on the GL thread.
+    static std::optional<MobileGL::String> ValidateComputeLocalSizeLimits(
+        const MobileGL::String& source, const MobileGL::MG_Util::ShaderTranspiler::CompileEnv& env) {
+        const ComputeLocalSize localSize = ParseComputeLocalSize(source);
+        if (!localSize.declared) return std::nullopt;
+
+        if (localSize.x > env.maxComputeWorkGroupSize[0] || localSize.y > env.maxComputeWorkGroupSize[1] ||
+            localSize.z > env.maxComputeWorkGroupSize[2]) {
+            return "Compute shader local_size exceeds GL_MAX_COMPUTE_WORK_GROUP_SIZE.";
+        }
+
+        const unsigned long long invocations = static_cast<unsigned long long>(localSize.x) * localSize.y * localSize.z;
+        if (invocations > env.maxComputeWorkGroupInvocations) {
+            return "Compute shader local_size product exceeds GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS.";
+        }
+
+        return std::nullopt;
+    }
+
+    // The half of a compile that depends on nothing but the source text, the stage and the
+    // environment snapshot: preprocessing, the two lexical rejections, and the two lexical
+    // side-channel extractions. Split out so P0b layer 2 can memoize exactly this and
+    // nothing else - the glslang parse stays per-object because its TShader is consume-once.
+    // Deliberately free of any per-object state so the memo is sound.
+    //
+    // The compute local-size verdict reads `env` rather than the live backend, and
+    // env.fingerprint is part of the P0b cache key, so a memo can never be returned against
+    // limits other than the ones it was computed against.
+    static MobileGL::MG_State::GLState::ShaderPreprocessResult RunSourceOnlyPipeline(
+        const MobileGL::ShaderStage stage, const MobileGL::String& source,
+        const MobileGL::MG_Util::ShaderTranspiler::CompileEnv& env) {
+        using namespace MobileGL;
+        using namespace MobileGL::MG_Util::ShaderTranspiler;
+        using MobileGL::MG_State::GLState::ShaderPreprocessOutcome;
+
+        MobileGL::MG_State::GLState::ShaderPreprocessResult result;
+        result.preprocessedSource = source;
+        PreprocessShaderSource(stage, result.preprocessedSource, env);
+
+        if (stage == ShaderStage::Compute) {
+            if (const std::optional<String> localSizeError =
+                    ValidateComputeLocalSizeLimits(result.preprocessedSource, env)) {
+                result.outcome = ShaderPreprocessOutcome::ComputeLocalSizeRejected;
+                result.infoLog = *localSizeError;
+                return result;
+            }
+        }
+
+        if (const std::optional<String> reservedError = FindReservedIdentifierViolation(result.preprocessedSource)) {
+            result.outcome = ShaderPreprocessOutcome::ReservedIdentifierRejected;
+            result.infoLog = *reservedError;
+            return result;
+        }
+
+        // The parse this feeds runs in the link-compatible configuration (Vulkan-client
+        // env with relaxed rules): the TShader it produces is what glLinkProgram links and
+        // what the backends' SPIR-V is generated from - there is no second, GL-client
+        // parse. The GL frontend semantics the relaxed parse cannot provide are restored
+        // on top: explicit default-block uniform locations through the lexical
+        // side-channels below, dead-uniform/global-UBO filtering in
+        // ProgramObject::DoReflection.
+        result.explicitUniformLocations = ExtractExplicitUniformLocations(result.preprocessedSource);
+        result.explicitOpaqueBindings = ExtractExplicitOpaqueBindings(result.preprocessedSource);
+        result.outcome = ShaderPreprocessOutcome::Preprocessed;
+        return result;
+    }
+
+} // namespace
+
+namespace MobileGL::MG_State::GLState {
+    // glslang has no "detach this thread" API in the vendored revision (there is no
+    // InitThread/DetachThread pair any more; thread attachment is implicit through
+    // thread_local state, and glslang::InitializeProcess() is process-wide, refcounted and
+    // mutex-guarded, so it needs no per-worker counterpart). The pool allocator is the part
+    // that needs undoing; see the declaration in ShaderCompileTask.h.
+    GlslangThreadAllocatorGuard::~GlslangThreadAllocatorGuard() { glslang::SetThreadPoolAllocator(nullptr); }
+
+    // Pure CPU work only. Everything this reads is either an input the node owns or a
+    // process-wide constant; everything it writes is `artifacts`. Do not add a GL/EGL call,
+    // a pActiveBackendObject read, or a pGLContext->RecordError() here - the first two are
+    // what CompileEnv exists to replace, and the third is why the design's section 6
+    // deferral mechanism (and JobNode's debug assert on it) exists.
+    void ShaderCompileTask::RunBody() {
+        // Own the failure rather than letting JobNode's backstop settle the node as
+        // Cancelled: an abandoned node publishes nothing, so the shader would report
+        // COMPILE_STATUS false with an EMPTY info log. GL models a failed compile as
+        // status + log, so turn a throw into exactly that - a completed job whose result
+        // is "this shader did not compile", with a log the application can read.
+        // (JobNode still catches: it is the last resort for anything below.)
+        try {
+            RunCompilePipeline();
+        } catch (const std::exception& e) {
+            artifacts = {};
+            artifacts.env = env;
+            artifacts.compileStatus = false;
+            artifacts.infoLog = std::format("Error: shader compilation failed: {}", e.what());
+        } catch (...) {
+            artifacts = {};
+            artifacts.env = env;
+            artifacts.compileStatus = false;
+            artifacts.infoLog = "Error: shader compilation failed: unknown exception";
+        }
+    }
+
+    void ShaderCompileTask::RunCompilePipeline() {
+        using namespace MG_Util::ShaderTranspiler;
+        const GlslangThreadAllocatorGuard glslangGuard;
+
+        const CompileEnv& compileEnv = *env;
+        artifacts.env = env;
+
+        // P0b layer 2: another shader object in this context may already have run the
+        // source-only half over byte-identical text under the same environment.
+        ShaderPreprocessResultPtr cached =
+            cache ? cache->Find(stage, sourceHash, *source, compileEnv.fingerprint) : nullptr;
+        SharedPtr<ShaderPreprocessResult> fresh;
+        if (!cached) fresh = MakeShared<ShaderPreprocessResult>(RunSourceOnlyPipeline(stage, *source, compileEnv));
+        const ShaderPreprocessResult& shared = cached ? *cached : *fresh;
+        const Bool shouldPopulateCache = !cached && cache != nullptr;
+
+        if (!shared.Preprocessed()) {
+            // Rejected lexically, or a glslang failure this context has already seen for
+            // this exact source (ParseFailed) - either way the parse can be skipped.
+            artifacts.infoLog = shared.infoLog;
+            if (shouldPopulateCache) {
+                cache->Insert(stage, sourceHash, *source, compileEnv.fingerprint, Move(fresh));
+            }
+            return;
+        }
+
+        ShaderAttrib attrib{.shaderType = MG_Util::ConvertShaderStageToGLEnum(stage),
+                            .sourceStr = shared.preprocessedSource,
+                            .flags = 0,
+                            .env = &compileEnv};
+
+        auto result = ShaderCompiler::CompileShader(attrib);
+        if (result) {
+            artifacts.compileStatus = true;
+            artifacts.shader = result.value();
+            // Copy, not move: `shared` may alias a cache entry that has to outlive us, and
+            // `fresh` is about to be handed to the cache.
+            artifacts.preprocessedSource = shared.preprocessedSource;
+            artifacts.explicitUniformLocations = shared.explicitUniformLocations;
+            artifacts.explicitOpaqueBindings = shared.explicitOpaqueBindings;
+            artifacts.infoLog.clear();
+            if (shouldPopulateCache) {
+                cache->Insert(stage, sourceHash, *source, compileEnv.fingerprint, Move(fresh));
+            }
+        } else {
+            artifacts.infoLog = result.error().log;
+            // Deferred, not logged here, for two reasons. MGLOG from a pool thread interleaves
+            // mid-line with the GL thread's own output and lands out of order relative to the
+            // glCompileShader that caused it; diagnostics.logLines is replayed by the join, on
+            // the GL thread, exactly where a serial implementation would have printed it.
+            // And a one-line summary rather than the old full source dump: a shaderpack stage
+            // is ~100KB, so the dump was the single largest thing this driver ever wrote to
+            // the log, for every failing shader. The info log is what names the offending
+            // line; the source is recoverable from the application.
+            const SizeT firstLineEnd = artifacts.infoLog.find('\n');
+            diagnostics.logLines.push_back(std::format(
+                "ShaderCompileTask: shader {} (stage {}) failed to compile; compileStatus = false. "
+                "Preprocessed source: {} bytes. First log line: {}",
+                externalIndex, static_cast<Int>(stage), shared.preprocessedSource.length(),
+                artifacts.infoLog.substr(0, firstLineEnd == String::npos ? artifacts.infoLog.length()
+                                                                        : firstLineEnd)));
+            if (shouldPopulateCache) {
+                fresh->outcome = ShaderPreprocessOutcome::ParseFailed;
+                fresh->infoLog = artifacts.infoLog;
+                fresh->explicitUniformLocations.clear();
+                fresh->explicitOpaqueBindings.clear();
+                cache->Insert(stage, sourceHash, *source, compileEnv.fingerprint, Move(fresh));
+            }
+        }
+    }
+
+    SharedPtr<glslang::TShader> ShaderCompileTask::ClaimParsedShader(String& outReparseLog) const {
+        MOBILEGL_ASSERT(IsComplete(),
+                        "ShaderCompileTask::ClaimParsedShader() on a job that has not completed; its artifacts "
+                        "are still being written");
+
+        if (artifacts.shader) {
+            // The whole race, in one instruction. Acquire-release because the winner is about
+            // to hand the TShader to glslang's linker on a possibly different thread from the
+            // one that parsed it - the node's terminal transition already published the
+            // parse, and this orders the two claimants against each other.
+            Bool expected = false;
+            if (m_parseClaimed.compare_exchange_strong(expected, true, std::memory_order_acq_rel,
+                                                       std::memory_order_acquire)) {
+                return artifacts.shader;
+            }
+        }
+
+        // Either another link already consumed the stored parse (and mapIO mutated its
+        // intermediate), or there never was one. Re-parse the preprocessed source through the
+        // identical configuration; that costs one glslang parse, which is what GenerateBinary
+        // used to spend here on EVERY link rather than only on reuse.
+        //
+        // The guard is not optional on this path: from stage 4 this runs on a pool worker,
+        // and TShader::parse would leave that worker's TLS allocator pointing at a pool the
+        // GL thread is about to free. (ProgramLinkTask::RunBody holds one too; they nest
+        // harmlessly - both just reset the thread to its own default.)
+        const GlslangThreadAllocatorGuard glslangGuard;
+        using namespace MG_Util::ShaderTranspiler;
+        ShaderAttrib attrib{.shaderType = MG_Util::ConvertShaderStageToGLEnum(stage),
+                            .sourceStr = artifacts.preprocessedSource,
+                            .flags = 0,
+                            // Re-parse against the SAME environment the original parse used,
+                            // not against whatever the backend reports now.
+                            .env = artifacts.env.get()};
+        auto result = ShaderCompiler::CompileShader(attrib);
+        if (!result) {
+            // Should be unreachable: the same source parsed successfully at Compile().
+            outReparseLog = result.error().log;
+            return nullptr;
+        }
+        return result.value();
+    }
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ShaderCompileTask.h

@@ -0,0 +1,186 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderCompileTask.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+#include <MG_Util/Async/JobNode.h>
+#include <MG_Util/ShaderTranspiler/CompileEnv.h>
+#include <MG_State/GLState/ProgramState/ShaderPreprocessCache.h>
+
+namespace MobileGL::MG_State::GLState {
+    // glslang has no "detach this thread" API in the vendored revision, but TShader::parse
+    // leaves the calling thread's TLS pool allocator pointing at the shader's own pool and
+    // never restores it. Left there, the next allocation this thread makes - in an unrelated
+    // job, or in glslang code reached from a different object - would come out of a pool the
+    // GL thread may already have deleted with the TShader. SetThreadPoolAllocator(nullptr)
+    // reverts the thread to its own thread_local default and is the documented idiom.
+    //
+    // A scope guard, so it also runs when a body throws. Declared here rather than kept
+    // file-local because stage 4 gave it a second user: ProgramLinkTask's body parses (the
+    // claim-CAS loser's re-parse), links and emits SPIR-V, all on a pool thread.
+    struct GlslangThreadAllocatorGuard {
+        GlslangThreadAllocatorGuard() = default;
+        ~GlslangThreadAllocatorGuard();
+        GlslangThreadAllocatorGuard(const GlslangThreadAllocatorGuard&) = delete;
+        GlslangThreadAllocatorGuard& operator=(const GlslangThreadAllocatorGuard&) = delete;
+    };
+
+    // Everything one glCompileShader PRODUCES, in one block.
+    //
+    // This is exactly the set a single run of the compile pipeline writes, which is what
+    // makes "discard the artifacts" a complete invalidation and "move the artifacts" a
+    // complete publish. It lives on the job node rather than on ShaderObject: a worker fills
+    // it in, and the GL thread reads it through ShaderObject's join gate.
+    struct ShaderCompileArtifacts {
+        // The CompileEnv snapshot this compile ran against. Held so the consume-once
+        // re-parse in ClaimParsedShader() reproduces the original parse exactly, instead of
+        // re-reading whatever the backend says now.
+        SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> env;
+        SharedPtr<glslang::TShader> shader;
+        // The source the parse actually consumed (after PreprocessShaderSource), kept for
+        // ClaimParsedShader's re-parse so a later link never depends on the preprocessor
+        // being deterministic across backend-state changes.
+        String preprocessedSource;
+        UnorderedMap<String, Int> explicitUniformLocations;
+        UnorderedMap<String, Uint> explicitOpaqueBindings;
+        String infoLog;
+        Bool compileStatus = false;
+    };
+
+    // The unit of asynchronous shader compilation: one glCompileShader's worth of pure CPU
+    // work - preprocess, the two lexical rejections, the two lexical extractions, and the
+    // glslang parse - with every input it needs owned by the node itself.
+    //
+    // That ownership is the whole point. The node reads no GL-thread state (the source is a
+    // SharedPtr<const String> snapshot, the device limits come from the CompileEnv snapshot,
+    // the P0b cross-object memo is shared-owned and internally locked) and writes nothing
+    // but its own `artifacts`. So a node whose ShaderObject was re-sourced, deleted, or
+    // destroyed while it was still running is safe to simply abandon - no wait, no
+    // synchronization with the GL thread beyond the node's own terminal state.
+    class ShaderCompileTask final : public MG_Util::Async::JobNode {
+    public:
+        ShaderCompileTask(const ShaderStage stage, SharedPtr<const String> source, const Uint64 sourceHash,
+                          SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> env,
+                          SharedPtr<ShaderPreprocessCache> cache, const Uint externalIndex)
+            : stage(stage), source(Move(source)), sourceHash(sourceHash), env(Move(env)), cache(Move(cache)),
+              externalIndex(externalIndex) {}
+
+        // ---- inputs: immutable after construction, all owned by the node ----
+        const ShaderStage stage;
+        // The exact text at enqueue. ShaderObject compares this pointer against its own
+        // m_source to decide whether its layer-1 memo is armed, which is why glShaderSource
+        // only ever swaps the pointer when the text genuinely differs.
+        const SharedPtr<const String> source;
+        const Uint64 sourceHash;
+        const SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> env;
+        // P0b layer 2, or null. Null is the "no context" case (the default fragment shader,
+        // the backends' internal blit/mipmap shaders) and doubles as the marker for
+        // "compile inline regardless of the async flag" - see ShaderObject::Compile().
+        const SharedPtr<ShaderPreprocessCache> cache;
+        const Uint externalIndex; // logs only
+
+        // ---- output: valid iff IsComplete(), immutable afterwards ----
+        ShaderCompileArtifacts artifacts;
+
+        // Hands out a link-consumable TShader, exactly once for the stored parse.
+        //
+        // glslang's mapIO mutates the TShader's aliased intermediate, so the parse this node
+        // produced may feed exactly ONE link; every later link (a relink, or the same shader
+        // attached to a second program) needs a fresh parse. The claim is a CAS on this
+        // shared node rather than a flag on the ShaderObject because from stage 4 the two
+        // callers can be two ProgramLinkTasks running on two workers: two programs sharing
+        // one shader, linked back to back. Copying the parse out and tracking consumed-ness
+        // per program would let both of them decide they were the first, run mapIO over the
+        // same intermediate twice, and ship silently corrupt SPIR-V.
+        //
+        // The CAS loser re-parses artifacts.preprocessedSource against THIS node's own
+        // CompileEnv (not against whatever the backend reports now), through the identical
+        // CompileShader path - so winner and loser produce byte-identical SPIR-V. Callable
+        // only once IsComplete() and compileStatus are true. Returns null only if that
+        // re-parse fails, and outReparseLog then carries its diagnostics.
+        //
+        // Const because the claim is the node's own synchronization, not a mutation of its
+        // published artifacts: a claim that is taken and then abandoned (its link was
+        // cancelled) costs one extra re-parse later and nothing else.
+        SharedPtr<glslang::TShader> ClaimParsedShader(String& outReparseLog) const;
+
+        // Sticky marker for "a ProgramLinkTask has this node in its input snapshot".
+        //
+        // It exists to keep a cancel from eating a result someone still needs. A pending link
+        // holds its dependencies by SharedPtr, so the NODE always outlives the ShaderObject -
+        // but Cancel() is not about lifetime, it discards the result. The reachable sequence
+        // is the ordinary one: compile, attach, glLinkProgram (enqueued), glDetachShader,
+        // glDeleteShader. The detach makes the shader GL-invisible, so the delete frees its
+        // name, and ReleaseShaderNameIfOrphaned would cancel a compile the enqueued link is
+        // waiting on - turning a link that must report GL_TRUE into GL_FALSE. Set on the GL
+        // thread in Link()'s prologue, read on the GL thread by
+        // ShaderObject::ReleaseCompileNode - which from stage 6 weighs it together with the
+        // adopter count below, because a node can now have both kinds of observer at once.
+        //
+        // Never cleared: the worst case is one stale node compiling to completion for nobody,
+        // which is exactly what the pre-stage-3 implementation always did.
+        void MarkLinkReferenced() { m_linkReferenced.store(true, std::memory_order_release); }
+        Bool IsLinkReferenced() const { return m_linkReferenced.load(std::memory_order_acquire); }
+
+        // ---- P1 stage 6: the adopter count ----
+        // How many live ShaderObjects currently hold this node in their m_compiled.
+        //
+        // It exists because stage 6 lets a node be SHARED: before it, a node had exactly one
+        // shader object, so "this object stopped caring" and "nothing can observe this
+        // result" were the same statement and ShaderObject::CancelCompile could cancel
+        // unconditionally. Once two GL shader names hold one node, that cancel would kill the
+        // other one's pending compile - a compile that must still report GL_TRUE. So a cancel
+        // is now authorized by TWO conditions, both checked by the releaser:
+        //   * this release brings the count to zero (no shader object is left), AND
+        //   * IsLinkReferenced() is false (no enqueued link took the node into its snapshot).
+        // The second is the stage-4 pin, unchanged; the first is what stage 6 adds.
+        //
+        // ---- Why a plain Int and not an atomic ----
+        // Every mutation is made from ShaderObject, and every ShaderObject mutation site is a
+        // GL entry point on the application's context thread: glCompileShader (adopt/create),
+        // glShaderSource with different text, glDeleteShader's orphan sweep, and
+        // ~ShaderObject. All of them are the SAME thread, so the count is never concurrently
+        // mutated and an atomic would only buy an unneeded lock prefix on the hottest compile
+        // path. Workers cannot touch it by construction: a job body's entire contract (see
+        // this class's header comment) is that it reads only the node's inputs and writes only
+        // `artifacts`, and a plain Int here makes that contract grep-checkable in a way an
+        // atomic would quietly hide.
+        //
+        // The CANCEL that the count authorizes still races the worker, and deliberately so -
+        // that is the settled cancel-not-join semantics from stage 3: JobNode::Cancel is
+        // cooperative and non-blocking, a node already running settles as Cancelled when its
+        // body returns, and a node that has already gone terminal ignores the request.
+        // Nothing about that changes here.
+        //
+        // Exactness under that race: ShaderObject::ReleaseCompileNode returns EARLY, without
+        // decrementing and without dropping its reference, when the node is already terminal
+        // (there is nothing left to stop). Terminality is sticky, so if a releaser observes a
+        // node as NON-terminal then no holder has ever taken that early return on it, and the
+        // count it reads is exactly the number of holders. If the worker finishes in the
+        // window between that observation and the Cancel(), the Cancel is a no-op on a
+        // terminal node - and the count was zero, so there was no other holder to harm.
+        void AddAdopter() { ++m_adopters; }
+        void ReleaseAdopter() {
+            MOBILEGL_ASSERT(m_adopters > 0,
+                            "ShaderCompileTask adopter count underflow; a ShaderObject released a node it did not "
+                            "hold (every release must pair with exactly one AddAdopter)");
+            --m_adopters;
+        }
+        Int AdopterCount() const { return m_adopters; }
+
+    private:
+        void RunBody() override;
+        // The real body; RunBody wraps it so a throw becomes a GL-visible compile failure.
+        void RunCompilePipeline();
+
+        mutable std::atomic<Bool> m_parseClaimed{false};
+        std::atomic<Bool> m_linkReferenced{false};
+        // GL-thread-owned; see AddAdopter above for why this is not an atomic.
+        Int m_adopters = 0;
+    };
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ShaderObject.cpp

@@ -7,197 +7,208 @@
 // End of Source File Header
 
 #include "ShaderObject.h"
-#include <MG_Util/ShaderTranspiler/Types.h>
+#include "ShaderPreprocessCache.h"
-#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+#include <MG_Util/Async/ShaderCompilePool.h>
 #include <MG_Util/Converters/MGToGL/ProgramEnumConverter.h>
-#include <MG_Util/ShaderTranspiler/ShaderSourceProcessor.h>
+#include <MG_Util/ShaderTranspiler/CompileEnv.h>
-#include <MG_Util/ShaderTranspiler/glslang/UniformTraverser.h>
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
-#include <MG_Backend/BackendObjects.h>
+#include <MG_Util/ShaderTranspiler/Types.h>
-
-namespace {
-    struct ComputeLocalSize {
-        MobileGL::Uint x = 1;
-        MobileGL::Uint y = 1;
-        MobileGL::Uint z = 1;
-        bool declared = false;
-    };
-
-    static MobileGL::String StripGlslComments(const MobileGL::String& source) {
-        MobileGL::String result;
-        result.reserve(source.length());
-
-        bool inLineComment = false;
-        bool inBlockComment = false;
-        for (MobileGL::SizeT i = 0; i < source.length(); ++i) {
-            if (inLineComment) {
-                if (source[i] == '\n') {
-                    inLineComment = false;
-                    result.push_back(source[i]);
-                } else {
-                    result.push_back(' ');
-                }
-                continue;
-            }
-
-            if (inBlockComment) {
-                if (source[i] == '*' && i + 1 < source.length() && source[i + 1] == '/') {
-                    inBlockComment = false;
-                    result.append("  ");
-                    ++i;
-                } else {
-                    result.push_back(source[i] == '\n' ? '\n' : ' ');
-                }
-                continue;
-            }
-
-            if (source[i] == '/' && i + 1 < source.length()) {
-                if (source[i + 1] == '/') {
-                    inLineComment = true;
-                    result.append("  ");
-                    ++i;
-                    continue;
-                }
-                if (source[i + 1] == '*') {
-                    inBlockComment = true;
-                    result.append("  ");
-                    ++i;
-                    continue;
-                }
-            }
-
-            result.push_back(source[i]);
-        }
-
-        return result;
-    }
-
-    static ComputeLocalSize ParseComputeLocalSize(const MobileGL::String& source) {
-        ComputeLocalSize localSize;
-        const MobileGL::String uncommentedSource = StripGlslComments(source);
-        const std::regex localSizePattern(R"(local_size_([xyz])\s*=\s*([0-9]+))");
-
-        for (std::sregex_iterator it(uncommentedSource.begin(), uncommentedSource.end(), localSizePattern), end;
-             it != end; ++it) {
-            const char axis = (*it)[1].str()[0];
-            const auto value = static_cast<unsigned long long>(std::stoull((*it)[2].str()));
-            const MobileGL::Uint clampedValue = value > UINT_MAX ? UINT_MAX : static_cast<MobileGL::Uint>(value);
-
-            // TODO: Replace this literal layout scanner with parser/AST-backed validation so expressions and
-            // specialization-id layouts are handled consistently with glslang.
-            localSize.declared = true;
-            if (axis == 'x') {
-                localSize.x = clampedValue;
-            } else if (axis == 'y') {
-                localSize.y = clampedValue;
-            } else {
-                localSize.z = clampedValue;
-            }
-        }
-
-        return localSize;
-    }
-
-    static MobileGL::Uint GetComputeWorkGroupSizeLimit(MobileGL::Uint index) {
-        constexpr MobileGL::Uint kFrontendMinComputeWorkGroupSizes[] = {1024, 1024, 64};
-        MobileGL::Int backendValue = 0;
-        if (MobileGL::MG_Backend::gBackendFunctionsTable.GL.GetIntegeri_v) {
-            MobileGL::MG_Backend::gBackendFunctionsTable.GL.GetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_SIZE, index,
-                                                                          &backendValue);
-        }
-
-        // TODO: Share these exposed compute limit helpers with GL_Getter.cpp instead of duplicating the frontend minima.
-        return std::max(static_cast<MobileGL::Uint>(std::max(backendValue, 0)),
-                        kFrontendMinComputeWorkGroupSizes[index]);
-    }
-
-    static unsigned long long GetComputeWorkGroupInvocationLimit() {
-        constexpr unsigned long long kFrontendMaxComputeWorkGroupInvocations = 1024;
-        if (!MobileGL::MG_Backend::pActiveBackendObject) return kFrontendMaxComputeWorkGroupInvocations;
-
-        return std::max(static_cast<unsigned long long>(std::max(
-                            MobileGL::MG_Backend::pActiveBackendObject->GetDynamicParameters()
-                                .MaxComputeWorkGroupInvocations,
-                            0)),
-                        kFrontendMaxComputeWorkGroupInvocations);
-    }
-
-    static std::optional<MobileGL::String> ValidateComputeLocalSizeLimits(const MobileGL::String& source) {
-        const ComputeLocalSize localSize = ParseComputeLocalSize(source);
-        if (!localSize.declared) return std::nullopt;
-
-        if (localSize.x > GetComputeWorkGroupSizeLimit(0) || localSize.y > GetComputeWorkGroupSizeLimit(1) ||
-            localSize.z > GetComputeWorkGroupSizeLimit(2)) {
-            return "Compute shader local_size exceeds GL_MAX_COMPUTE_WORK_GROUP_SIZE.";
-        }
-
-        const unsigned long long invocations = static_cast<unsigned long long>(localSize.x) * localSize.y * localSize.z;
-        if (invocations > GetComputeWorkGroupInvocationLimit()) {
-            return "Compute shader local_size product exceeds GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS.";
-        }
-
-        return std::nullopt;
-    }
-}
 
 namespace MobileGL::MG_State::GLState {
     void ShaderObject::SetShaderSource(const String& source) {
-        m_source = source;
+        // P0b layer 1. glShaderSource always REPLACES the source, but replacing it with a
-        m_shader.reset();
+        // byte-identical one cannot change what a compile would produce: the whole
-        m_compileStatus = false;
+        // pipeline (preprocess -> lexical checks -> glslang parse) is a pure function of
-        m_infoLog.clear();
+        // (stage, source, CompileEnv). So keeping the compiled state is not an optimization
+        // that changes observable behaviour - the COMPILE_STATUS, the info log and the
+        // reflection a caller can query are exactly what a real recompile would have
+        // rebuilt, byte for byte. A compile still IN FLIGHT is left running for the same
+        // reason: it is computing the right answer for text this object still holds.
+        if (SourceMatchesCompiledState(source)) return;
+        // The text genuinely changed, so whatever a running job is computing is now about
+        // an old source. Give up our claim on it - it owns its own copy of that old string,
+        // so swapping the pointer below cannot race its storage. Note "our claim", not "the
+        // job": another shader object may have adopted the same node and still be waiting for
+        // exactly this answer, which is what ReleaseCompileNode's count discipline protects.
+        ReleaseCompileNode();
+        m_source = MakeShared<const String>(source);
+        InvalidateCompiledState();
     }
 
     void ShaderObject::SetShaderSource(String&& source) {
-        m_source = Move(source);
+        if (SourceMatchesCompiledState(source)) return;
-        m_shader.reset();
+        ReleaseCompileNode();
-        m_compileStatus = false;
+        m_source = MakeShared<const String>(Move(source));
-        m_infoLog.clear();
+        InvalidateCompiledState();
+    }
+
+    Bool ShaderObject::SourceMatchesCompiledState(const String& candidate) const {
+        // The memo is armed exactly while a job exists that was built from the string this
+        // object still points at - pending or finished, success or failure.
+        if (!HasMemoizedCompile()) return false;
+        if (candidate.length() != m_source->length()) return false;
+        // Never let correctness ride on a hash: the answer is the full text comparison.
+        // (The stored hash on the node is a cache-lookup accelerator, not a substitute.)
+        return candidate == *m_source;
+    }
+
+    void ShaderObject::JoinPendingCompile() const {
+        MOBILEGL_ASSERT(!MG_Util::Async::ShaderCompilePool::IsPoolThread(),
+                        "ShaderObject::EnsureCompileJoined() reached from a pool thread; a job body must never read "
+                        "GL-thread-owned objects");
+        m_compiled->Wait();
+        m_compileJoined = true;
+        // Errors and worker-side log lines are raised HERE, on the GL thread, at the first
+        // join of the job that produced them - which for a single shader is trivially the
+        // order a serial implementation would have produced them in.
+        //
+        // ApplyDeferredDiagnostics DRAINS, so a node shared by several shader objects
+        // (stage 6) replays its worker-side log line exactly once, at whichever object joins
+        // first. That is the honest report - one compile ran - and it is log text only: the
+        // GL-observable half of a failure, COMPILE_STATUS and the info log, lives in
+        // `artifacts` and every sharer reads the identical copy of it.
+        MG_Util::Async::ApplyDeferredDiagnostics(*m_compiled);
+        // A node that settled as Cancelled published nothing. Dropping it here is what keeps
+        // the object's state machine to two reachable cases - "no job" and "a job that
+        // completed" - so every reader below can treat a live node as authoritative.
+        //
+        // Through DropCompileNode, not a bare reset: this object is letting the node go, so
+        // its adopter slot has to go with it. A node shared with another object stays alive
+        // and gets dropped once more when that object joins - once per holder, never twice
+        // for the same one, because DropCompileNode is null-guarded.
+        if (!m_compiled->IsComplete()) DropCompileNode();
+    }
+
+    void ShaderObject::AdoptCompileNode(SharedPtr<ShaderCompileTask> node) const {
+        // Never overwrite a hold without giving its slot back first.
+        DropCompileNode();
+        m_compiled = Move(node);
+        m_compiled->AddAdopter();
+        // Re-arm the join gate: whether this node was just created or just adopted from
+        // another object, THIS object has not pulled its result yet. (An adopted node may
+        // already be terminal - the join then only replays what is left of its diagnostics.)
+        m_compileJoined = false;
+    }
+
+    void ShaderObject::DropCompileNode() const {
+        if (!m_compiled) return;
+        m_compiled->ReleaseAdopter();
+        m_compiled.reset();
+    }
+
+    void ShaderObject::InvalidateCompiledState() {
+        // The job node holds exactly what one Compile() produces, so discarding it IS the
+        // invalidation - and it re-arms nothing, so the next Compile() genuinely recompiles.
+        DropCompileNode();
+    }
+
+    void ShaderObject::ReleaseCompileNode() {
+        if (!m_compiled) return;
+        // Already terminal: there is nothing left to stop, so this is not a release at all -
+        // the node and this object's claim on it both stay. That early return is older than
+        // stage 6 and it is load-bearing: ProgramState::ReleaseShaderNameIfOrphaned calls
+        // this on a shader whose name is going away but whose object a ProgramObject may
+        // still hold, and dropping a COMPLETED compile there would turn that program's link
+        // into GL_FALSE.
+        if (m_compiled->IsTerminal()) return;
+        // Two independent claimants have to be checked before a cancel, and this object is
+        // authorized to cancel only if BOTH say the result has become unobservable.
+        //
+        // 1. Other shader objects. From stage 6 a node can be SHARED by several GL shader
+        //    names that were handed byte-identical source; cancelling here would turn a
+        //    compile they must still see as GL_TRUE into GL_FALSE. Only the releaser that
+        //    takes the count to zero - i.e. the last holder - may cancel. See
+        //    ShaderCompileTask::AddAdopter for why a plain Int is sound here and for the
+        //    exactness argument under the worker race.
+        // 2. A pending LINK. An enqueued ProgramLinkTask holds the node in its input snapshot
+        //    and a cancel would turn its link into GL_FALSE; reached by the ordinary
+        //    link-then-detach-then-delete shader teardown. See MarkLinkReferenced. Never
+        //    cleared, so this is a one-way pin.
+        //
+        // The cancel itself is cooperative and non-blocking, exactly as before: a node no
+        // worker has picked up settles immediately, a running one is flagged and settles when
+        // its body returns, writing only into itself the whole time.
+        if (m_compiled->AdopterCount() == 1 && !m_compiled->IsLinkReferenced()) m_compiled->Cancel();
+        DropCompileNode();
     }
 
     void ShaderObject::Compile() {
-        using namespace MG_Util::ShaderTranspiler;
+        // P0b layer 1, as a tri-state: the memo is "the node in m_compiled was built from
-        String compileSource = m_source;
+        // the string m_source still points at". SetShaderSource only swaps that pointer when
-        MG_Util::ShaderTranspiler::PreprocessShaderSource(m_stage, compileSource);
+        // the text actually differs, so this is a pointer compare, and it covers Pending as
+        // well as Complete - a second glCompileShader on an in-flight object is a no-op, not
+        // a duplicate job racing to write the same fields.
+        //
+        // The failure case is covered too: the info log stays queryable because nothing is
+        // cleared. And if the stored TShader already fed a link, the no-op leaves
+        // preprocessedSource and both side-channel maps intact, which is precisely what
+        // ClaimParsedShader's on-demand re-parse needs - a real recompile would have handed
+        // the next link a fresh parse, the no-op hands it a fresh re-parse of the identical
+        // source instead. Same result, one parse either way.
+        if (HasMemoizedCompile()) return;
 
-        if (m_stage == ShaderStage::Compute) {
+        // Two reasons to stay on this thread, one rule. Without the async flag the whole
-            const std::optional<String> localSizeError = ValidateComputeLocalSizeLimits(compileSource);
+        // path must be byte-identical to the synchronous implementation, and a cache-less
-            if (localSizeError) {
+        // object is an internal shader that compiles and reads its status in the same
-                m_compileStatus = false;
+        // breath (see the constructor comment) - a job would only add a round trip.
-                m_shader.reset();
+        // AsyncShaderCompileActive(), not ...Enabled(): a glMaxShaderCompilerThreadsKHR(0)
-                m_infoLog = *localSizeError;
+        // has to put compilation back on this thread even though the extension is still
+        // advertised, and that is exactly what makes the GL_COMPLETION_STATUS_KHR the
+        // extension mandates after a zero count (immediately GL_TRUE) fall out for free.
+        //
+        // Hoisted above the node construction because stage 6 keys off it too: this same
+        // answer decides whether the adoption map is consulted at all, so a
+        // glMaxShaderCompilerThreadsKHR(0) and a flag-off build both bypass sharing exactly
+        // as they bypass the pool, and their behaviour stays byte-identical to pre-stage-6.
+        const Bool runOnPool = m_preprocessCache && MG_Util::Async::AsyncShaderCompileActive();
+
+        // The compile-environment snapshot is taken HERE, on the GL thread, and handed to
+        // the job. Everything the pipeline needs to know about the device comes through it,
+        // never through pActiveBackendObject - that is what makes the body movable.
+        const SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> env =
+            MG_Util::ShaderTranspiler::GetCurrentCompileEnv();
+        const Uint64 sourceHash = ShaderPreprocessCache::HashSource(*m_source);
+
+        // ---- P1 stage 6: adopt an equivalent compile instead of enqueueing a duplicate ----
+        // ~21% of all glCompileShader calls in the shaderpack corpus are a DIFFERENT shader
+        // object handed byte-identical source. P0b's memo only pays off once one of them has
+        // finished; under async they are all enqueued in the same burst, so without this each
+        // one runs the whole pipeline on its own worker. The map hands back the node the
+        // first of them created - in flight or already complete - and this object simply
+        // holds it too.
+        if (runOnPool && m_adoptionMap) {
+            if (SharedPtr<ShaderCompileTask> shared =
+                    m_adoptionMap->FindAdoptable(m_stage, sourceHash, *m_source, env->fingerprint)) {
+                // Take the node's own source snapshot as ours. FindAdoptable just compared
+                // the two strings in full, so this changes nothing observable - but it is not
+                // optional: the layer-1 memo (HasMemoizedCompile) is a POINTER comparison
+                // against the node's snapshot, so leaving our own equal-but-distinct copy in
+                // place would make the very next glCompileShader on this object decide it had
+                // no memo and enqueue the duplicate this whole stage exists to avoid - and
+                // would make an identical glShaderSource re-source cancel a shared compile.
+                // It also collapses N copies of a ~100 KB shaderpack stage into one.
+                m_source = shared->source;
+                AdoptCompileNode(Move(shared));
                 return;
             }
         }
 
-        const std::optional<String> reservedError =
+        AdoptCompileNode(MakeShared<ShaderCompileTask>(m_stage, m_source, sourceHash, env, m_preprocessCache,
-            MG_Util::ShaderTranspiler::FindReservedIdentifierViolation(compileSource);
+                                                       m_externalIndex));
-        if (reservedError) {
+
-            m_compileStatus = false;
+        if (!runOnPool) {
-            m_shader.reset();
+            m_compiled->RunInline();
-            m_infoLog = *reservedError;
+            // Inline means the node is already terminal, so this join only replays
+            // diagnostics; it is here so the synchronous and asynchronous paths publish
+            // through the identical code.
+            EnsureCompileJoined();
             return;
         }
-
+        // Registered BEFORE the post, so the very next glCompileShader in this burst can
-        // Compile for OpenGL here, so that we can do validation and link
+        // adopt it however fast a worker picks it up. Registration is an index entry only -
-        // like a real OpenGL driver at linking stage
+        // the map holds a WeakPtr and never keeps a node alive.
-        // Will compile for other backends later.
+        if (m_adoptionMap) m_adoptionMap->Register(m_compiled);
-        ShaderAttrib attrib{.shaderType = MG_Util::ConvertShaderStageToGLEnum(m_stage),
+        MG_Util::Async::ShaderCompilePool::Get().Post(m_compiled);
-                            .sourceStr = compileSource,
-                            .flags = ShaderCompileBits::CompileForOpenGL};
-
-        auto result = ShaderCompiler::CompileShader(attrib);
-        if (result) {
-            m_compileStatus = true;
-            m_shader = result.value();
-            m_infoLog.clear();
-        } else {
-            m_compileStatus = false;
-            m_shader.reset();
-            m_infoLog = result.error().log;
-            MGLOG_D("ShaderObject::Compile: Shader %d compilation failed.\nSource:\n%s\nInfoLog:\n%s\nSetting "
-                    "m_compileStatus = false as a result.",
-                    m_externalIndex, compileSource.c_str(), m_infoLog.c_str());
-        }
     }
 
     void ShaderObject::MarkAsDeleted() {

MobileGL/MG_State/GLState/ProgramState/ShaderObject.h

@@ -8,48 +8,229 @@
 
 #pragma once
 #include <Includes.h>
+#include <MG_State/GLState/ProgramState/ShaderStage.h>
+#include <MG_State/GLState/ProgramState/ShaderCompileTask.h>
+#include <MG_State/GLState/ProgramState/ShaderCompileAdoptionMap.h>
 
 namespace MobileGL {
-    enum class ShaderStage {
-        Vertex,
-        TessControl,
-        TessEval,
-        Geometry,
-        Fragment,
-        Compute,
-        ShaderStageCount,
-        Unknown = -1
-    };
-
     namespace MG_State::GLState {
+        // The GL-visible shader name. It owns the source text and one compile job node; the
+        // job node owns everything a compile produces.
+        //
+        // Every member below is GL-thread-owned, and every read of worker-produced state
+        // goes through Compiled(), which joins first. That is invariant I5 of the P1 design:
+        // because Compiled() is the SOLE accessor of the node's artifacts, the compiler
+        // enumerates every reader for us and none can be forgotten.
         class ShaderObject {
         public:
-            ShaderObject(const ShaderStage stage, Uint externalIndex)
+            // `preprocessCache` is the owning context's cross-object memo (P0b layer 2).
-                : m_stage(stage), m_externalIndex(externalIndex) {}
+            // Null is fully supported and means two things at once: "no sharing", and
+            // "compile inline, never on a worker". Those coincide exactly - the only
+            // cache-less shader objects are the internal ones (ProgramObject's default
+            // fragment shader, the DirectVulkan blit and depth-mipmap shaders) and every one
+            // of them compiles and reads its status in the same breath, so a job would only
+            // add a round trip. Shared ownership rather than a raw pointer: a compile job
+            // outlives neither the object nor the context deterministically, and the cache
+            // has to stay alive for whoever is still reading it.
+            //
+            // `adoptionMap` is the same context's stage-6 index of adoptable compile nodes.
+            // It is non-null exactly when `preprocessCache` is (ProgramState hands both out
+            // together, and nobody else hands out either), which is what makes "no cache"
+            // keep meaning "compile inline, share nothing": an internal shader object has
+            // neither, so it neither adopts nor registers and its path is byte-identical to
+            // the pre-stage-6 one. GL-thread-only, so unlike the cache it carries no lock -
+            // shared ownership only because a ShaderObject may outlive the context's tables.
+            ShaderObject(const ShaderStage stage, Uint externalIndex,
+                         SharedPtr<ShaderPreprocessCache> preprocessCache = nullptr,
+                         SharedPtr<ShaderCompileAdoptionMap> adoptionMap = nullptr)
+                : m_stage(stage), m_externalIndex(externalIndex), m_preprocessCache(Move(preprocessCache)),
+                  m_adoptionMap(Move(adoptionMap)) {}
+            // Cancel-not-join: the node owns its inputs, so an in-flight compile whose
+            // object just went away is safe to abandon where it stands. Nothing can observe
+            // its result any more - unless another shader object adopted the same node, or a
+            // link pinned it, which is precisely what ReleaseCompileNode() checks.
+            ~ShaderObject() {
+                ReleaseCompileNode();
+                // ReleaseCompileNode KEEPS a node that has already gone terminal - there is
+                // nothing left to stop, so it is not a release at all. This object is going
+                // away regardless, so hand the adopter slot back here. That is what keeps
+                // ShaderCompileTask::AdopterCount() exactly "how many live ShaderObjects hold
+                // this node" instead of merely an upper bound.
+                DropCompileNode();
+            }
+
+            ShaderObject(const ShaderObject&) = delete;
+            ShaderObject& operator=(const ShaderObject&) = delete;
+
             void SetShaderSource(const String& source);
             void SetShaderSource(String&& source);
             void Compile();
+            // Gives up this object's claim on its compile node, cancelling the node only if
+            // this object was its LAST claimant. Called at the points where the object's
+            // compiled state stops being observable through THIS name: a real source change,
+            // and the release of an orphaned shader name.
+            //
+            // Named for what it does rather than for what it used to do: before stage 6 a
+            // node had exactly one shader object, so giving up the claim and cancelling the
+            // compile were the same act and this was CancelCompile(). They are not the same
+            // act any more - see ShaderCompileTask::AddAdopter for the count discipline and
+            // its single-threadedness argument. Never waits, in either case.
+            void ReleaseCompileNode();
             void MarkAsDeleted();
 
+            // The compile job node itself, for ProgramObject::Link()'s input snapshot.
+            // DELIBERATELY DOES NOT JOIN, and that is the entire point of stage 4: the link
+            // takes the node as a dependency and is posted only once the node is terminal,
+            // so glLinkProgram never blocks on glCompileShader. Null means this object has
+            // never been compiled (or its last compile was abandoned), which the link reads
+            // as COMPILE_STATUS false - the same verdict the joining path produces.
+            //
+            // The caller must MarkLinkReferenced() whatever it keeps: from here on the node's
+            // result has an observer this object knows nothing about (see the marker's
+            // comment in ShaderCompileTask.h).
+            const SharedPtr<ShaderCompileTask>& CompiledNodeForLink() const { return m_compiled; }
+
             Uint GetExternalIndex() const { return m_externalIndex; }
             ShaderStage GetShaderStage() const { return m_stage; }
-            const String& GetShaderSource() const { return m_source; }
+            // No join: the source is GL-thread-owned, and a worker only ever reads the
-            const SharedPtr<glslang::TShader>& GetCompiledShader() const { return m_shader; }
+            // immutable snapshot it was handed at enqueue.
-            const String& GetInfoLog() const { return m_infoLog; }
+            const String& GetShaderSource() const { return *m_source; }
-            const UnorderedMap<String, Uint>& GetUniformLocations() const { return m_uniforms; }
+            // The snapshot itself, for whoever needs to hand it to a job.
-            Bool GetCompileStatus() const { return m_compileStatus; }
+            const SharedPtr<const String>& GetShaderSourcePtr() const { return m_source; }
+
+            const SharedPtr<glslang::TShader>& GetCompiledShader() const { return Compiled().shader; }
+            const String& GetInfoLog() const { return Compiled().infoLog; }
+            // Explicit layout(location = N) qualifiers on this shader's default-block
+            // uniforms, captured lexically at Compile() because the relaxed parse drops
+            // them from reflection (see ExtractExplicitUniformLocations).
+            const UnorderedMap<String, Int>& GetExplicitUniformLocations() const {
+                return Compiled().explicitUniformLocations;
+            }
+            // Explicit layout(binding = N) on sampler/image uniforms - their initial
+            // texture/image units - captured lexically for the same reason (see
+            // ExtractExplicitOpaqueBindings).
+            const UnorderedMap<String, Uint>& GetExplicitOpaqueBindings() const {
+                return Compiled().explicitOpaqueBindings;
+            }
+            Bool GetCompileStatus() const { return Compiled().compileStatus; }
             Bool GetDeleteStatus() const { return m_deleteStatus; }
 
+            // Blocks until a pending compile has published its artifacts. Public for the
+            // sites that must join without reading anything - ProgramObject::Link's
+            // prologue, which needs every attached shader settled before it runs.
+            void JoinCompile() const { EnsureCompileJoined(); }
+
+            // True while this object holds the outcome (success OR failure) of a Compile()
+            // of exactly the source it currently holds - i.e. while the P0b layer-1 memo is
+            // armed and a glCompileShader would be a no-op. Diagnostics and tests only;
+            // nothing in the GL frontend branches on it.
+            //
+            // Tri-state, and deliberately NOT joining: an in-flight compile of the current
+            // source counts as memoized (a second glCompileShader must not enqueue a
+            // duplicate job), but asking that question must never block.
+            // A node that settled as Cancelled (the job body threw, or the enqueue failed)
+            // carries no result, so it must NOT satisfy the memo: otherwise a second
+            // glCompileShader on the same source enqueues nothing and the eventual join
+            // reports GL_FALSE forever. The synchronous path retries in exactly this case.
+            Bool HasMemoizedCompile() const {
+                return m_compiled != nullptr && m_compiled->source == m_source && !m_compiled->IsCancelled();
+            }
+
+            // MUST NOT JOIN - this is what GL_COMPLETION_STATUS_KHR will read when the
+            // extension surface lands. "No job at all" counts as complete: there is nothing
+            // outstanding to wait for.
+            Bool IsCompileComplete() const { return m_compiled == nullptr || m_compiled->IsTerminal(); }
+
         private:
+            // ---- The one and only join gate for compile output (P1 invariant I5) ----
+            // The fast path - no job, or a job whose result this object has already pulled -
+            // is two predictable branches and stays inline: it runs on every Compiled() read
+            // and the project never builds with LTO, so an out-of-line body would be a real
+            // cross-TU call at each of those sites. The blocking half is out of line.
+            //
+            // The gate keys on "has this object pulled the job's result yet", NOT on "is the
+            // job terminal". Those differ in the case that matters: a worker can finish a
+            // compile before the GL thread ever looks at it, and the pull is where deferred
+            // diagnostics get replayed and an abandoned node gets dropped. Keying on
+            // terminality would silently skip both.
+            void EnsureCompileJoined() const {
+                if (m_compiled && !m_compileJoined) JoinPendingCompile();
+            }
+            void JoinPendingCompile() const;
+
+            // The artifacts of a compile that ran to completion. A node that was abandoned
+            // (cancelled at teardown, or whose body threw) never publishes: JoinPendingCompile
+            // drops it, so anything reachable here is either Complete or absent, and "absent"
+            // reads as the never-compiled defaults - COMPILE_STATUS false, empty info log,
+            // which is exactly what GL requires before the first glCompileShader.
+            static const ShaderCompileArtifacts& EmptyArtifacts() {
+                static const ShaderCompileArtifacts empty;
+                return empty;
+            }
+            const ShaderCompileArtifacts& Compiled() const {
+                EnsureCompileJoined();
+                return m_compiled ? m_compiled->artifacts : EmptyArtifacts();
+            }
+
+            void InvalidateCompiledState();
+
+            // ---- the ONLY two writers of m_compiled (P1 stage 6) ----
+            // Every adopter-count mutation lives in these two, which is what makes "exactly
+            // one AddAdopter per hold, exactly one ReleaseAdopter per hold" auditable rather
+            // than something review has to re-derive at each call site. DropCompileNode is
+            // null-guarded, so calling it on an object that already let go is a no-op and a
+            // double release is unrepresentable.
+            void AdoptCompileNode(SharedPtr<ShaderCompileTask> node) const;
+            void DropCompileNode() const;
+
+            // ---- P0b layer 1: per-object no-op recompile ----
+            // True iff `candidate` is byte-identical to the source that produced (or is
+            // producing) the compiled state this object currently holds.
+            Bool SourceMatchesCompiledState(const String& candidate) const;
+
+            // ---- GL-thread-owned state: never produced by a compile, so it never joins ----
+            const ShaderStage m_stage;
             const Uint m_externalIndex = 0;
-            const ShaderStage m_stage;
+            // The pre-glShaderSource state, shared by every untouched object rather than
-            String m_source;
+            // allocated per glCreateShader.
-            SharedPtr<glslang::TShader> m_shader;
+            static const SharedPtr<const String>& EmptySource() {
-            UnorderedMap<String, Uint> m_uniforms;
+                static const SharedPtr<const String> empty = MakeShared<const String>();
+                return empty;
+            }
+            // glShaderSource text, as an immutable snapshot. Never null. A job holds its own
+            // SharedPtr to the exact string it was given, so replacing the source under a
+            // running compile cannot race its storage - and the layer-1 memo collapses to a
+            // pointer comparison against the job's snapshot, because the setter only swaps
+            // the pointer when the text genuinely differs.
+            //
+            // Not necessarily unique to this object from stage 6 on: adopting a node also
+            // takes that node's source snapshot (see Compile()), so N shader objects sharing
+            // one compile share one copy of the text. The string is immutable and shared-
+            // owned, so that is invisible to every reader.
+            SharedPtr<const String> m_source = EmptySource();
+
+            // P0b layer 2: the owning context's cross-object memo, or null. Internally
+            // locked, because several workers hit it at once.
+            const SharedPtr<ShaderPreprocessCache> m_preprocessCache;
+            // P1 stage 6: the owning context's index of adoptable compile nodes, or null.
+            // Touched only from Compile(), i.e. only on the GL thread, so it carries no lock.
+            const SharedPtr<ShaderCompileAdoptionMap> m_adoptionMap;
 
-            String m_infoLog;
             Bool m_deleteStatus = false;
-            Bool m_compileStatus = false;
+
+            // ---- Compile OUTPUT ---- pending OR completed; reachable only through Compiled().
+            // Mutable because the join is a read-side operation: a const getter has to be
+            // able to settle an outstanding job before answering.
+            //
+            // SHARED from stage 6 on: several shader objects holding byte-identical source
+            // under the same CompileEnv point at one node. Every read below still goes
+            // through the same join gate, and a second joiner finds the node already
+            // terminal, so nothing about the read path changes - only the release path does
+            // (ReleaseCompileNode).
+            mutable SharedPtr<ShaderCompileTask> m_compiled;
+            // Exactly-once latch for the pull above. Armed with every new job node, set by
+            // the one join that consumes it.
+            mutable Bool m_compileJoined = false;
         };
     } // namespace MG_State::GLState
 } // namespace MobileGL

MobileGL/MG_State/GLState/ProgramState/ShaderPreprocessCache.cpp

@@ -0,0 +1,86 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderPreprocessCache.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "ShaderPreprocessCache.h"
+
+namespace MobileGL::MG_State::GLState {
+    ShaderPreprocessResultPtr ShaderPreprocessCache::Find(const ShaderStage stage, const Uint64 sourceHash,
+                                                         const String& source, const Uint64 envFingerprint) const {
+        const Key key{.stage = stage,
+                      .sourceHash = sourceHash,
+                      .sourceLength = source.length(),
+                      .envFingerprint = envFingerprint};
+
+        const std::lock_guard<std::mutex> lock(m_mutex);
+        const auto it = m_index.find(key);
+        if (it == m_index.end()) return nullptr;
+
+        // Never let correctness ride on a 64-bit hash: confirm the hit byte for byte.
+        // Lengths already matched (they are part of the key), so this is a plain memcmp.
+        const Entry& entry = *it->second;
+        if (entry.originalSource != source) return nullptr;
+
+        // A copy of the SharedPtr, taken under the lock: the payload now outlives any
+        // eviction the caller races with.
+        return entry.result;
+    }
+
+    void ShaderPreprocessCache::Insert(const ShaderStage stage, const Uint64 sourceHash, const String& source,
+                                       const Uint64 envFingerprint, ShaderPreprocessResultPtr result) {
+        if (!result) return;
+
+        const SizeT entryBytes = EntryBytes(source, *result);
+        // A single source bigger than the whole budget would evict every other entry and
+        // then itself; refuse it instead of thrashing the cache empty.
+        if (entryBytes > kMaxStoredSourceBytes) return;
+
+        const Key key{.stage = stage,
+                      .sourceHash = sourceHash,
+                      .sourceLength = source.length(),
+                      .envFingerprint = envFingerprint};
+
+        const std::lock_guard<std::mutex> lock(m_mutex);
+        if (const auto existing = m_index.find(key); existing != m_index.end()) {
+            // Either a re-insert of the same source (harmless) or a genuine hash collision
+            // with a different source. Both are resolved by letting the newcomer win: one
+            // entry per key keeps the index a plain map, and a collision is astronomically
+            // rare enough that the loser simply misses.
+            EraseEntryLocked(existing->second);
+        }
+
+        m_entries.push_back(Entry{.key = key, .originalSource = source, .result = Move(result)});
+        m_index[key] = std::prev(m_entries.end());
+        m_storedSourceBytes += entryBytes;
+
+        EvictUntilWithinBudgetLocked();
+    }
+
+    void ShaderPreprocessCache::Clear() {
+        const std::lock_guard<std::mutex> lock(m_mutex);
+        m_entries.clear();
+        m_index.clear();
+        m_storedSourceBytes = 0;
+    }
+
+    void ShaderPreprocessCache::EraseEntryLocked(const EntryList::iterator it) {
+        const SizeT bytes = EntryBytes(it->originalSource, *it->result);
+        m_storedSourceBytes = bytes > m_storedSourceBytes ? 0 : m_storedSourceBytes - bytes;
+        m_index.erase(it->key);
+        m_entries.erase(it);
+    }
+
+    void ShaderPreprocessCache::EvictUntilWithinBudgetLocked() {
+        // FIFO: the oldest insertion goes first. Insert() already refuses entries larger
+        // than the byte budget, so this loop always terminates with at least the entry
+        // that was just added still resident.
+        while (!m_entries.empty() &&
+               (m_entries.size() > kMaxEntries || m_storedSourceBytes > kMaxStoredSourceBytes)) {
+            EraseEntryLocked(m_entries.begin());
+        }
+    }
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ShaderPreprocessCache.h

@@ -0,0 +1,149 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderPreprocessCache.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+#include <list>
+#include <mutex>
+// Deliberately NOT ShaderObject.h: ShaderCompileTask.h needs this header, and ShaderObject.h
+// needs ShaderCompileTask.h. Only ShaderStage was ever used from there.
+#include <MG_State/GLState/ProgramState/ShaderStage.h>
+#include <MG_State/GLState/ProgramState/ShaderSourceKey.h>
+
+namespace MobileGL::MG_State::GLState {
+    // Where the shared, source-only half of ShaderObject::Compile() stopped. The two
+    // rejection verdicts are kept apart (rather than collapsed into "failed") so a hit
+    // reproduces the original diagnosis, not just the original info log.
+    enum class ShaderPreprocessOutcome : Uint8 {
+        // The source-only half ran clean; preprocessedSource and both maps are valid.
+        Preprocessed,
+        // ValidateComputeLocalSizeLimits rejected it (compute only).
+        ComputeLocalSizeRejected,
+        // FindReservedIdentifierViolation rejected it.
+        ReservedIdentifierRejected,
+        // The source-only half was clean but glslang rejected the preprocessed source.
+        // Memoizing this saves the parse itself on every later object with that source.
+        ParseFailed,
+    };
+
+    // Everything ShaderObject::Compile() derives from the source text alone, i.e.
+    // everything that is identical for two shader objects holding byte-identical source.
+    struct ShaderPreprocessResult {
+        ShaderPreprocessOutcome outcome = ShaderPreprocessOutcome::Preprocessed;
+        // Valid unless the preprocessor itself never ran; kept even for the rejection
+        // outcomes because that is the text the diagnostics refer to.
+        String preprocessedSource;
+        UnorderedMap<String, Int> explicitUniformLocations;
+        UnorderedMap<String, Uint> explicitOpaqueBindings;
+        // The compile info log to publish; empty when outcome == Preprocessed.
+        String infoLog;
+
+        Bool Preprocessed() const { return outcome == ShaderPreprocessOutcome::Preprocessed; }
+    };
+
+    // Cache hits hand out shared ownership, not a raw pointer into the entry list. That is
+    // what makes the cache safe once compiles run concurrently: a reader keeps its payload
+    // alive across any eviction, and a 107 KB preprocessedSource is never copied on a hit.
+    using ShaderPreprocessResultPtr = SharedPtr<const ShaderPreprocessResult>;
+
+    // P0b layer 2: a per-context, bounded memo of the source-only half of shader
+    // compilation, keyed by (stage, xxhash64(source), source length).
+    //
+    // Motivation: in the Iris shader-pack corpus ~21% of every glCompileShader in a trace
+    // is a *different* shader object holding byte-identical source (packs glue the same
+    // common/composite GLSL into many program stages), so the preprocess + reserved-
+    // identifier scan + explicit-location/binding extraction runs over the same megabytes
+    // again and again. Layer 1 (in ShaderObject) covers the same object recompiled with
+    // unchanged source; this covers the cross-object case.
+    //
+    // What is NOT cached: the glslang parse. glslang's TShader is consume-once (mapIO
+    // mutates the aliased intermediate at link), so every shader object still needs its
+    // own parse; only the text-processing half is shared.
+    //
+    // Correctness: the 64-bit hash is a lookup accelerator only. Every hit re-compares the
+    // full stored original source with memcmp before it is honored, so a hash collision
+    // degrades to a miss, never to a wrong answer. That is why the full original text is
+    // stored rather than a prefix/suffix digest - the cache is bounded, so the cost is.
+    //
+    // Eviction: FIFO (insertion order), bounded by BOTH an entry count and a stored-source
+    // byte budget, whichever binds first. FIFO rather than LRU because shader-pack loading
+    // is a burst of mostly-distinct sources whose reuse clusters around insertion time;
+    // LRU's extra list splice on every hit buys nothing measurable here, and FIFO keeps
+    // Find() a genuinely const, read-only operation.
+    class ShaderPreprocessCache {
+    public:
+        static constexpr SizeT kMaxEntries = 128;
+        static constexpr SizeT kMaxStoredSourceBytes = 8u * 1024u * 1024u;
+
+        // Returns the memoized result for this exact source under this exact compile
+        // environment, or null on a miss. The returned SharedPtr owns its payload, so it
+        // stays valid for as long as the caller holds it - across Insert(), Clear(), and
+        // across the destruction of the cache itself.
+        //
+        // envFingerprint joins the key because the source-only pipeline's compute
+        // local-size verdict is computed against CompileEnv's device limits: a memo must
+        // never outlive the environment it was computed against (memo-hazard rule).
+        ShaderPreprocessResultPtr Find(ShaderStage stage, Uint64 sourceHash, const String& source,
+                                       Uint64 envFingerprint) const;
+
+        // Memoizes `result` for this source. A source whose own storage cost already
+        // exceeds the byte budget is simply not cached (caching it would evict everything
+        // else and then itself).
+        void Insert(ShaderStage stage, Uint64 sourceHash, const String& source, Uint64 envFingerprint,
+                    ShaderPreprocessResultPtr result);
+
+        void Clear();
+
+        static Uint64 HashSource(const String& source) {
+            return static_cast<Uint64>(XXH64(source.data(), source.length(), 0));
+        }
+
+        SizeT GetEntryCount() const {
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            return m_entries.size();
+        }
+        SizeT GetStoredSourceBytes() const {
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            return m_storedSourceBytes;
+        }
+
+    private:
+        // Shared with ShaderCompileAdoptionMap so the two per-context memos cannot key
+        // themselves on different notions of "the same compile" - see ShaderSourceKey.h.
+        using Key = ShaderSourceKey;
+        using KeyHasher = ShaderSourceKeyHasher;
+
+        struct Entry {
+            Key key;
+            // The full original (pre-preprocess) source, kept so a hit can be confirmed by
+            // comparison instead of trusting the hash.
+            String originalSource;
+            ShaderPreprocessResultPtr result;
+        };
+
+        using EntryList = std::list<Entry>;
+
+        static SizeT EntryBytes(const String& source, const ShaderPreprocessResult& result) {
+            return source.length() + result.preprocessedSource.length();
+        }
+
+        void EvictUntilWithinBudgetLocked();
+
+        void EraseEntryLocked(EntryList::iterator it);
+
+        // P1: every public entry point takes this. The lock alone would NOT have been
+        // enough - the old Find() handed back a raw pointer into an entry that a
+        // concurrent Insert()'s FIFO eviction could erase while the caller was still
+        // reading it. Shared ownership of the payload is what closes that hole; the mutex
+        // only protects the containers below.
+        mutable std::mutex m_mutex;
+        EntryList m_entries;                                 // front = oldest (FIFO victim)
+        UnorderedMap<Key, EntryList::iterator, KeyHasher> m_index;
+        SizeT m_storedSourceBytes = 0;
+    };
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ShaderSourceKey.h

@@ -0,0 +1,51 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderSourceKey.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+#include <MG_State/GLState/ProgramState/ShaderStage.h>
+
+namespace MobileGL::MG_State::GLState {
+    // The identity of "one glCompileShader's worth of input" - the tuple that decides
+    // whether two compiles must produce byte-identical results. Shared by the two
+    // per-context memos keyed on it, so that neither can drift from the other:
+    //   * P0b's ShaderPreprocessCache, which memoizes the source-only half of a compile;
+    //   * P1 stage 6's ShaderCompileAdoptionMap, which shares the job NODE itself.
+    //
+    // The 64-bit source hash is a LOOKUP ACCELERATOR ONLY. Every user of this key confirms
+    // a candidate hit with a full byte comparison of the stored source before honoring it,
+    // so a hash collision degrades to a miss and never to a wrong answer. That rule is not
+    // negotiable - see the memo-hazard notes on ShaderPreprocessCache.
+    //
+    // envFingerprint is part of the identity because the pipeline's compute local-size
+    // verdict is computed against CompileEnv's device limits: a memo must never be handed
+    // back under an environment other than the one it was computed against.
+    struct ShaderSourceKey {
+        ShaderStage stage = ShaderStage::Unknown;
+        Uint64 sourceHash = 0;
+        SizeT sourceLength = 0;
+        Uint64 envFingerprint = 0;
+
+        Bool operator==(const ShaderSourceKey& other) const {
+            return stage == other.stage && sourceHash == other.sourceHash &&
+                   sourceLength == other.sourceLength && envFingerprint == other.envFingerprint;
+        }
+    };
+
+    struct ShaderSourceKeyHasher {
+        SizeT operator()(const ShaderSourceKey& key) const {
+            // The source hash already spreads well; fold the three discriminators in so
+            // that same-hash-different-stage/length/env keys land in different buckets.
+            Uint64 mixed = key.sourceHash;
+            mixed ^= static_cast<Uint64>(key.sourceLength) + 0x9e3779b97f4a7c15ull + (mixed << 6) + (mixed >> 2);
+            mixed ^= static_cast<Uint64>(static_cast<Int>(key.stage)) * 0xff51afd7ed558ccdull;
+            mixed ^= key.envFingerprint + 0x9e3779b97f4a7c15ull + (mixed << 6) + (mixed >> 2);
+            return static_cast<SizeT>(mixed);
+        }
+    };
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ShaderStage.h

@@ -0,0 +1,25 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ShaderStage.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+
+namespace MobileGL {
+    // Split out of ShaderObject.h so the compile pipeline's headers form a DAG:
+    // ShaderStage.h <- ShaderPreprocessCache.h <- ShaderCompileTask.h <- ShaderObject.h.
+    // Every existing includer of ShaderObject.h still sees this type unchanged.
+    enum class ShaderStage {
+        Vertex,
+        TessControl,
+        TessEval,
+        Geometry,
+        Fragment,
+        Compute,
+        ShaderStageCount,
+        Unknown = -1
+    };
+} // namespace MobileGL

MobileGL/MG_State/GLState/SamplerState/SamplerObject.cpp

@@ -8,6 +8,8 @@
 
 #include "SamplerObject.h"
 
+#include <MG_State/GLState/Core.h>
+
 #include <atomic>
 
 namespace MobileGL {
@@ -22,81 +24,91 @@ namespace MobileGL {
             SamplerObject::SamplerObject(Uint externalIndex)
                 : m_externalIndex(externalIndex), m_lifetimeId(AllocateLifetimeId()) {}
 
+            void SamplerObject::BumpVersion() {
+                ++m_version;
+                // Every setter early-outs on an unchanged value, so this only runs on a real
+                // parameter change. The generation is bumped for ALL parameters, not just filter
+                // ones that feed mipmap-completeness: a backend memo of the resolved per-unit
+                // bindings must never miss an invalidation, and over-invalidating on a wrap-mode
+                // write costs one re-resolve.
+                if (pGLContext) pGLContext->BumpSamplingResolutionGeneration();
+            }
+
             void SamplerObject::SetWrapS(SamplerWrapMode mode) {
                 if (mode == m_samplerParameters.wrapS) return;
 
                 m_samplerParameters.wrapS = mode;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetWrapT(SamplerWrapMode mode) {
                 if (mode == m_samplerParameters.wrapT) return;
 
                 m_samplerParameters.wrapT = mode;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetWrapR(SamplerWrapMode mode) {
                 if (mode == m_samplerParameters.wrapR) return;
 
                 m_samplerParameters.wrapR = mode;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetMinFilter(SamplerFilterMode mode) {
                 if (mode == m_samplerParameters.minFilter) return;
 
                 m_samplerParameters.minFilter = mode;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetMagFilter(SamplerFilterMode mode) {
                 if (mode == m_samplerParameters.magFilter) return;
 
                 m_samplerParameters.magFilter = mode;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetMipmapMode(SamplerMipmapMode mode) {
                 if (mode == m_samplerParameters.mipmapMode) return;
 
                 m_samplerParameters.mipmapMode = mode;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetLodRange(Float minLod, Float maxLod) {
                 if (minLod == m_samplerParameters.minLod && maxLod == m_samplerParameters.maxLod) return;
                 m_samplerParameters.minLod = minLod;
                 m_samplerParameters.maxLod = maxLod;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetLodBias(Float bias) {
                 if (bias == m_samplerParameters.lodBias) return;
 
                 m_samplerParameters.lodBias = bias;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetMaxAnisotropy(Float maxAnisotropy) {
                 if (maxAnisotropy == m_samplerParameters.maxAnisotropy) return;
 
                 m_samplerParameters.maxAnisotropy = maxAnisotropy;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetSamplerCompareFunc(SamplerCompareFunc func) {
                 if (func == m_samplerParameters.compareFunc) return;
 
                 m_samplerParameters.compareFunc = func;
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetCompareMode(SamplerCompareMode mode) {
                 if (mode == m_samplerParameters.compareMode) return;
 
                 m_samplerParameters.compareMode = mode;
-                ++m_version;
+                BumpVersion();
             }
 
             SamplerWrapMode SamplerObject::GetWrapS() const {
@@ -153,7 +165,7 @@ namespace MobileGL {
                 m_samplerParameters.borderColorUI =
                     UintVec4(static_cast<Uint32>(color.x()), static_cast<Uint32>(color.y()),
                              static_cast<Uint32>(color.z()), static_cast<Uint32>(color.w()));
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetBorderColorI(const IntVec4& color) {
@@ -166,7 +178,7 @@ namespace MobileGL {
                 m_samplerParameters.borderColor =
                     FloatVec4(static_cast<Float>(color.x()), static_cast<Float>(color.y()),
                               static_cast<Float>(color.z()), static_cast<Float>(color.w()));
-                ++m_version;
+                BumpVersion();
             }
 
             void SamplerObject::SetBorderColorUI(const UintVec4& color) {
@@ -179,7 +191,7 @@ namespace MobileGL {
                 m_samplerParameters.borderColor =
                     FloatVec4(static_cast<Float>(color.x()), static_cast<Float>(color.y()),
                               static_cast<Float>(color.z()), static_cast<Float>(color.w()));
-                ++m_version;
+                BumpVersion();
             }
 
             const FloatVec4& SamplerObject::GetBorderColor() const {

MobileGL/MG_State/GLState/SamplerState/SamplerObject.h

@@ -128,6 +128,12 @@ namespace MobileGL {
 
             private:
                 static Uint64 AllocateLifetimeId();
+                // The ONLY way m_version may move. Besides marking this object's parameters
+                // dirty for the backends it bumps the context-wide sampling-resolution
+                // generation: MIN_FILTER decides whether a lookup reads the mip chain, which
+                // decides whether a bound texture is mipmap-complete, which decides whether a
+                // backend binds it on its unit at all.
+                void BumpVersion();
 
                 const Uint m_externalIndex;
                 const Uint64 m_lifetimeId;

MobileGL/MG_State/GLState/TextureState/MipmapStorage.cpp

@@ -11,6 +11,23 @@
 namespace MobileGL {
     namespace MG_State {
         namespace GLState {
+            namespace {
+                // Overlapping OR abutting ([lo, hi) intervals meeting edge-to-edge) in
+                // every axis: merging abutting boxes keeps scanline/tile write patterns
+                // as one rect instead of a picket fence.
+                Bool RegionsTouch(const MipmapDirtyRegion& a, const MipmapDirtyRegion& b) {
+                    return a.lo.x() <= b.hi.x() && b.lo.x() <= a.hi.x() && a.lo.y() <= b.hi.y() &&
+                           b.lo.y() <= a.hi.y() && a.lo.z() <= b.hi.z() && b.lo.z() <= a.hi.z();
+                }
+
+                MipmapDirtyRegion RegionUnion(const MipmapDirtyRegion& a, const MipmapDirtyRegion& b) {
+                    return {IntVec3{std::min(a.lo.x(), b.lo.x()), std::min(a.lo.y(), b.lo.y()),
+                                    std::min(a.lo.z(), b.lo.z())},
+                            IntVec3{std::max(a.hi.x(), b.hi.x()), std::max(a.hi.y(), b.hi.y()),
+                                    std::max(a.hi.z(), b.hi.z())}};
+                }
+            } // namespace
+
             SizeT MipmapStorage::GetLevelCount() const {
                 return m_data.size();
             }
@@ -28,6 +45,7 @@ namespace MobileGL {
                     m_texelSizes.resize(requiredLevelCount);
                     m_isDirty.resize(requiredLevelCount, false);
                     m_dirtyRegions.resize(requiredLevelCount);
+                    m_dirtyRects.resize(requiredLevelCount);
                     m_compressedData.resize(requiredLevelCount);
                     m_compressedFormats.resize(requiredLevelCount, GL_NONE);
                 }
@@ -44,6 +62,11 @@ namespace MobileGL {
                                                         std::max(input.texelSize.z(), 1)}}
                             : MipmapDirtyRegion{};
                 }
+                // The rect list mirrors the union box's reset: whatever rects were
+                // pending measured the OLD extents. Empty list = union box tells all.
+                if (level < m_dirtyRects.size()) {
+                    m_dirtyRects[level].clear();
+                }
                 auto& data = m_data[level];
                 data.resize(input.byteSize, 0);
 
@@ -94,6 +117,7 @@ namespace MobileGL {
                 m_texelSizes.resize(levelCount);
                 m_isDirty.resize(levelCount);
                 m_dirtyRegions.resize(levelCount);
+                m_dirtyRects.resize(levelCount);
                 m_compressedData.resize(levelCount);
                 m_compressedFormats.resize(levelCount);
             }
@@ -142,6 +166,13 @@ namespace MobileGL {
                         m_dirtyRegions[level] = {};
                     }
                 }
+                // Both directions collapse the rect list to "just the union box": a
+                // whole-level dirty IS the union box, a clean level has nothing to say.
+                // clear() keeps the vector's capacity, so per-frame streaming levels
+                // allocate their slots once and reuse them.
+                if (level < m_dirtyRects.size()) {
+                    m_dirtyRects[level].clear();
+                }
             }
 
             bool MipmapStorage::IsDirty(Uint level) const {
@@ -158,6 +189,20 @@ namespace MobileGL {
                                std::min(offset.y() + size.y(), levelSize.y()),
                                std::min(offset.z() + std::max(size.z(), 1), std::max(levelSize.z(), 1))};
                 if (incoming.Empty()) return;
+                // Rect list first, while the union box still holds only the PREVIOUS
+                // writes: a level that is already dirty with an empty list is in the
+                // "union box tells all" resting state, so that box seeds the list
+                // before the incoming rect refines it.
+                if (level < m_dirtyRects.size()) {
+                    auto& rects = m_dirtyRects[level];
+                    if (!m_isDirty[level]) {
+                        rects.clear(); // stale-safety; MarkDirty(false) already cleared it
+                    } else if (rects.empty() && level < m_dirtyRegions.size() &&
+                               !m_dirtyRegions[level].Empty()) {
+                        rects.push_back(m_dirtyRegions[level]);
+                    }
+                    InsertDirtyRect(level, incoming);
+                }
                 if (level < m_dirtyRegions.size()) {
                     MipmapDirtyRegion& region = m_dirtyRegions[level];
                     if (m_isDirty[level] && !region.Empty()) {
@@ -174,10 +219,82 @@ namespace MobileGL {
                 m_isDirty[level] = true;
             }
 
+            void MipmapStorage::InsertDirtyRect(Uint level, MipmapDirtyRegion incoming) {
+                auto& rects = m_dirtyRects[level];
+                if (rects.capacity() < kMaxDirtyRects) {
+                    rects.reserve(kMaxDirtyRects);
+                }
+                // Cascade-merge: absorb every rect the incoming touches. The absorbed
+                // union can reach rects a smaller box did not, so rescan until stable;
+                // every merge shrinks the list, so this terminates. Swap-with-back keeps
+                // removal O(1) - the list is unordered by design.
+                Bool merged = true;
+                while (merged) {
+                    merged = false;
+                    for (SizeT i = 0; i < rects.size(); ++i) {
+                        if (RegionsTouch(rects[i], incoming)) {
+                            incoming = RegionUnion(rects[i], incoming);
+                            rects[i] = rects.back();
+                            rects.pop_back();
+                            merged = true;
+                            break;
+                        }
+                    }
+                }
+                if (rects.size() < kMaxDirtyRects) {
+                    rects.push_back(incoming);
+                    return;
+                }
+                // Full: fold the incoming rect into the neighbour whose box grows least
+                // (least new area dragged into the upload), then re-insert the grown
+                // box - it may now touch others. The removal above guarantees the
+                // recursion appends on the second pass at the latest.
+                SizeT best = 0;
+                SizeT bestGrowth = ~static_cast<SizeT>(0);
+                for (SizeT i = 0; i < rects.size(); ++i) {
+                    const SizeT growth = RegionUnion(rects[i], incoming).TexelCount() - rects[i].TexelCount();
+                    if (growth < bestGrowth) {
+                        bestGrowth = growth;
+                        best = i;
+                    }
+                }
+                incoming = RegionUnion(rects[best], incoming);
+                rects[best] = rects.back();
+                rects.pop_back();
+                InsertDirtyRect(level, incoming);
+            }
+
             MipmapDirtyRegion MipmapStorage::GetDirtyRegion(Uint level) const {
                 if (level >= m_dirtyRegions.size()) return {};
                 return m_dirtyRegions[level];
             }
+
+            SizeT MipmapStorage::GetDirtyRects(Uint level, MipmapDirtyRegion* outRects, SizeT maxRects) const {
+                if (outRects == nullptr || level >= m_dirtyRects.size() || level >= m_dirtyRegions.size()) {
+                    return 0;
+                }
+                const auto& rects = m_dirtyRects[level];
+                // 0 or 1 rects: the union box already says exactly this. More than the
+                // caller can take: never truncate - a dropped rect is a dropped write.
+                if (rects.size() < 2 || rects.size() > maxRects) {
+                    return 0;
+                }
+                // Total-bytes accounting: when the scattered rects add up to most of
+                // the union box anyway (>= 3/4), one driver call on the box beats many
+                // calls moving nearly the same bytes.
+                SizeT summedArea = 0;
+                for (const auto& rect : rects) {
+                    summedArea += rect.TexelCount();
+                }
+                const SizeT unionArea = m_dirtyRegions[level].TexelCount();
+                if (summedArea * 4 >= unionArea * 3) {
+                    return 0;
+                }
+                for (SizeT i = 0; i < rects.size(); ++i) {
+                    outRects[i] = rects[i];
+                }
+                return rects.size();
+            }
         } // namespace GLState
     } // namespace MG_State
 } // namespace MobileGL

MobileGL/MG_State/GLState/TextureState/MipmapStorage.h

@@ -31,6 +31,11 @@ namespace MobileGL {
                     return lo.x() <= 0 && lo.y() <= 0 && lo.z() <= 0 && hi.x() >= levelSize.x() &&
                            hi.y() >= levelSize.y() && hi.z() >= std::max(levelSize.z(), 1);
                 }
+                SizeT TexelCount() const {
+                    if (Empty()) return 0;
+                    return static_cast<SizeT>(hi.x() - lo.x()) * static_cast<SizeT>(hi.y() - lo.y()) *
+                           static_cast<SizeT>(hi.z() - lo.z());
+                }
             };
 
             class MipmapStorage {
@@ -54,6 +59,28 @@ namespace MobileGL {
                 // Meaningful only while IsDirty(level).
                 MipmapDirtyRegion GetDirtyRegion(Uint level) const;
 
+                // Behind the union box, the level keeps up to kMaxDirtyRects pairwise
+                // disjoint rects recording WHERE the writes actually landed. A frame of
+                // ~100 scattered sprite updates in a big atlas has a union box that
+                // covers nearly the whole level while the touched texels are ~5% of it;
+                // the union box stays the source of truth (every write funnels through
+                // MarkDirty/MarkDirtyRegion into BOTH representations), backends OPT IN
+                // to the list purely as an upload-size refinement. 96 slots because the
+                // pattern this exists for is Minecraft's ~100 sprites/frame: a 16-slot
+                // list forced into far-apart merges was measured at >90% of the union
+                // box's area on exactly that pattern, i.e. worthless. Inserts merge any
+                // touching/overlapping rect (cascading, so the list stays disjoint);
+                // when full, the incoming rect folds into the neighbour whose box grows
+                // least and the list degrades gracefully toward the union box.
+                static constexpr SizeT kMaxDirtyRects = 96;
+                // Copies the level's dirty rects into outRects and returns how many were
+                // written. 0 means "upload the union box instead" and covers every
+                // reason at once: tracking unavailable, a single rect (identical to the
+                // union box by construction), more rects than maxRects, or a summed
+                // area so close to the union box's that one big upload beats many small
+                // ones (fewer driver calls wins when the bytes are nearly equal).
+                SizeT GetDirtyRects(Uint level, MipmapDirtyRegion* outRects, SizeT maxRects) const;
+
                 // The bytes an application handed to glCompressedTexImage*, kept verbatim beside the
                 // (uncompressed) texel shadow rather than in place of it. GL 4.6 core 8.11 requires
                 // glGetCompressedTexImage to return the image *as stored*, and no backend here has a
@@ -70,10 +97,22 @@ namespace MobileGL {
                 const void* MapCompressedData(Uint level) const;
 
             protected:
+                // Insert one clamped, non-empty write box, keeping the list disjoint
+                // and bounded (see kMaxDirtyRects).
+                void InsertDirtyRect(Uint level, MipmapDirtyRegion incoming);
+
                 Vector<IntVec3> m_texelSizes;
                 Vector<Vector<Uint8>> m_data;
                 Vector<bool> m_isDirty;
                 Vector<MipmapDirtyRegion> m_dirtyRegions;
+                // Per level, the disjoint rect list behind m_dirtyRegions' union box.
+                // An EMPTY list is the common resting state and always means "the union
+                // box is the whole story" - clean levels, whole-level dirties and
+                // respecifies all just clear it, so plain full-level uploads never pay
+                // a heap allocation; the first scattered MarkDirtyRegion on an
+                // already-dirty level seeds the list from the union box accumulated so
+                // far and refines from there.
+                Vector<Vector<MipmapDirtyRegion>> m_dirtyRects;
                 Vector<Vector<Uint8>> m_compressedData;
                 Vector<GLenum> m_compressedFormats;
             };

MobileGL/MG_State/GLState/TextureState/MipmapUploadTargetArray.h

@@ -84,6 +84,12 @@ namespace MobileGL {
                     return m_storage[targetIndex].GetDirtyRegion(level);
                 }
 
+                SizeT GetDirtyRects(Uint targetIndex, Uint level, MipmapDirtyRegion* outRects,
+                                    SizeT maxRects) const {
+                    MOBILEGL_ASSERT(targetIndex < TargetCount, "GetDirtyRects: target invalid");
+                    return m_storage[targetIndex].GetDirtyRects(level, outRects, maxRects);
+                }
+
                 void SetCompressedImage(Uint targetIndex, Uint level, GLenum internalFormat, const void* data,
                                         SizeT size) {
                     MOBILEGL_ASSERT(targetIndex < TargetCount, "SetCompressedImage: target invalid");

MobileGL/MG_State/GLState/TextureState/TextureObject.cpp

@@ -25,6 +25,18 @@ namespace MobileGL {
                 return s_nextTextureLifetimeId.fetch_add(1, std::memory_order_relaxed);
             }
 
+            void TextureObjectBase::BumpShapeVersion() {
+                ++m_shapeVersion;
+                // Shape is what mipmap-completeness is computed from, and completeness decides
+                // whether a backend binds this texture on its unit at all. Nothing else tells a
+                // backend memo of the resolved per-unit bindings that the answer moved - no bind
+                // changed and the texel content may be untouched. Proxy textures (used only to
+                // answer PROXY queries) are never bound, so their shape churn costs a memo
+                // invalidation for nothing; that is accepted rather than filtered, because a
+                // missed bump renders wrong pixels while a spare bump only costs one re-resolve.
+                if (pGLContext) pGLContext->BumpSamplingResolutionGeneration();
+            }
+
             TextureObjectBase::TextureObjectBase(TextureTarget target, Uint externalIndex)
                 : m_externalIndex(externalIndex), m_lifetimeId(AllocateLifetimeId()), m_target(target) {
                 m_sampler = MakeShared<SamplerObject>(0);
@@ -85,7 +97,7 @@ namespace MobileGL {
                 }
 
                 m_internalFormat = format;
-                ++m_shapeVersion;
+                BumpShapeVersion();
                 ++m_textureParamsVersion;
             }
 
@@ -197,7 +209,7 @@ namespace MobileGL {
                     m_levelRange.y() = m_levelRange.x();
                 }
                 ++m_textureParamsVersion;
-                ++m_shapeVersion;
+                BumpShapeVersion();
             }
 
             void TextureObjectBase::SetMaxLevel(Uint maxLevel) {
@@ -208,7 +220,7 @@ namespace MobileGL {
 
                 m_levelRange.y() = maxLevel;
                 ++m_textureParamsVersion;
-                ++m_shapeVersion;
+                BumpShapeVersion();
             }
 
             Bool TextureObjectBase::IsImmutable() const {
@@ -291,12 +303,12 @@ namespace MobileGL {
 
             void TextureObjectWithOneMipmap::AllocateStorage(TextureUploadTarget uploadTarget, Uint mipmapLevel,
                                                              MipmapInput input) {
-                ++m_shapeVersion;
+                BumpShapeVersion();
                 m_textureStorage.AllocateLevel(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel, input);
             }
 
             void TextureObjectWithOneMipmap::TruncateMipmapLevels(TextureUploadTarget uploadTarget, Uint levelCount) {
-                ++m_shapeVersion;
+                BumpShapeVersion();
                 m_textureStorage.TruncateToLevelCount(GetIndexOfTextureUploadTarget(uploadTarget), levelCount);
             }
 
@@ -333,6 +345,13 @@ namespace MobileGL {
                 return m_textureStorage.GetDirtyRegion(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel);
             }
 
+            SizeT TextureObjectWithOneMipmap::GetStorageDirtyRects(TextureUploadTarget uploadTarget, Uint mipmapLevel,
+                                                                   MipmapDirtyRegion* outRects,
+                                                                   SizeT maxRects) const {
+                return m_textureStorage.GetDirtyRects(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel,
+                                                      outRects, maxRects);
+            }
+
             void TextureObjectWithOneMipmap::SetMipmapCompressedImage(TextureUploadTarget uploadTarget, Uint mipmapLevel,
                                                               GLenum internalFormat, const void* data, SizeT size) {
                 m_textureStorage.SetCompressedImage(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel,

MobileGL/MG_State/GLState/TextureState/TextureObject.h

@@ -121,6 +121,12 @@ namespace MobileGL::MG_State::GLState {
 
     protected:
         static Uint64 AllocateLifetimeId();
+        // The ONLY way m_shapeVersion may move. Besides invalidating this object's own
+        // completeness memo it bumps the context-wide sampling-resolution generation, which is
+        // what a backend memo of the resolved per-unit bindings watches: completeness decides
+        // whether a bound texture reaches its native target at all, and a shape change is
+        // otherwise invisible to such a memo (no bind moved).
+        void BumpShapeVersion();
 
         const Uint m_externalIndex;
         const Uint64 m_lifetimeId;
@@ -180,6 +186,20 @@ namespace MobileGL::MG_State::GLState {
             const IntVec3 size = GetMipmapTexelSize(uploadTarget, mipmapLevel);
             return {IntVec3{0, 0, 0}, IntVec3{size.x(), size.y(), std::max(size.z(), 1)}};
         }
+        // Scatter detail behind GetStorageDirtyRegion: up to maxRects disjoint rects
+        // that together cover every dirty texel, so ~100 sprite writes in a big atlas
+        // need not be uploaded as one atlas-sized box. Returns how many rects were
+        // written to outRects; 0 means "no list, upload the union box" and is always a
+        // safe answer - this base fallback keeps whole-level semantics for storage
+        // classes that do not track rects, and backends OPT IN by calling this.
+        virtual SizeT GetStorageDirtyRects(TextureUploadTarget uploadTarget, Uint mipmapLevel,
+                                           MipmapDirtyRegion* outRects, SizeT maxRects) const {
+            (void)uploadTarget;
+            (void)mipmapLevel;
+            (void)outRects;
+            (void)maxRects;
+            return 0;
+        }
 
         // The compressed image a glCompressedTexImage* call shadowed for this level, kept verbatim
         // next to the texel data rather than instead of it - see MipmapStorage. The texel shadow
@@ -251,6 +271,8 @@ namespace MobileGL::MG_State::GLState {
         void MarkStorageDirtyRegion(TextureUploadTarget uploadTarget, Uint mipmapLevel, IntVec3 offset,
                                     IntVec3 size) override;
         MipmapDirtyRegion GetStorageDirtyRegion(TextureUploadTarget uploadTarget, Uint mipmapLevel) const override;
+        SizeT GetStorageDirtyRects(TextureUploadTarget uploadTarget, Uint mipmapLevel, MipmapDirtyRegion* outRects,
+                                   SizeT maxRects) const override;
         void SetMipmapCompressedImage(TextureUploadTarget uploadTarget, Uint mipmapLevel, GLenum internalFormat,
                                       const void* data, SizeT size) override;
         GLenum GetMipmapCompressedFormat(TextureUploadTarget uploadTarget, Uint mipmapLevel) const override;

MobileGL/MG_State/GLState/TextureState/TextureObject2DCube.cpp

@@ -28,12 +28,12 @@ namespace MobileGL {
 
             void TextureObject2DCube::AllocateStorage(TextureUploadTarget uploadTarget, Uint mipmapLevel,
                                                       MipmapInput input) {
-                ++m_shapeVersion;
+                BumpShapeVersion();
                 m_textureStorage.AllocateLevel(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel, input);
             }
 
             void TextureObject2DCube::TruncateMipmapLevels(TextureUploadTarget uploadTarget, Uint levelCount) {
-                ++m_shapeVersion;
+                BumpShapeVersion();
                 m_textureStorage.TruncateToLevelCount(GetIndexOfTextureUploadTarget(uploadTarget), levelCount);
             }
 
@@ -69,6 +69,12 @@ namespace MobileGL {
                 return m_textureStorage.GetDirtyRegion(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel);
             }
 
+            SizeT TextureObject2DCube::GetStorageDirtyRects(TextureUploadTarget uploadTarget, Uint mipmapLevel,
+                                                            MipmapDirtyRegion* outRects, SizeT maxRects) const {
+                return m_textureStorage.GetDirtyRects(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel,
+                                                      outRects, maxRects);
+            }
+
             void TextureObject2DCube::SetMipmapCompressedImage(TextureUploadTarget uploadTarget, Uint mipmapLevel,
                                                               GLenum internalFormat, const void* data, SizeT size) {
                 m_textureStorage.SetCompressedImage(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel,

MobileGL/MG_State/GLState/TextureState/TextureObject2DCube.h

@@ -31,6 +31,8 @@ namespace MobileGL {
                                             IntVec3 size) override;
                 MipmapDirtyRegion GetStorageDirtyRegion(TextureUploadTarget uploadTarget,
                                                         Uint mipmapLevel) const override;
+                SizeT GetStorageDirtyRects(TextureUploadTarget uploadTarget, Uint mipmapLevel,
+                                           MipmapDirtyRegion* outRects, SizeT maxRects) const override;
                 void SetMipmapCompressedImage(TextureUploadTarget uploadTarget, Uint mipmapLevel,
                                               GLenum internalFormat, const void* data, SizeT size) override;
                 GLenum GetMipmapCompressedFormat(TextureUploadTarget uploadTarget, Uint mipmapLevel) const override;

MobileGL/MG_State/GLState/TextureState/TextureState.cpp

@@ -7,6 +7,8 @@
 // End of Source File Header
 
 #include "TextureState.h"
+
+#include <atomic>
 #include "Defines.h"
 #include "TextureEnum.h"
 #include "TextureObject.h"
@@ -18,6 +20,12 @@
 #include "TextureObjectStubs.h"
 
 namespace MobileGL::MG_State::GLState {
+    static std::atomic<Uint64> s_nextTextureStateContextId = 1;
+
+    Uint64 TextureState::AllocateContextId() {
+        return s_nextTextureStateContextId.fetch_add(1, std::memory_order_relaxed);
+    }
+
     static SharedPtr<ITextureObject> MakeTextureObjectForTarget(Uint index, TextureTarget target) {
         switch (target) {
         case TextureTarget::Texture1D:
@@ -50,7 +58,7 @@ namespace MobileGL::MG_State::GLState {
         }
     }
 
-    TextureState::TextureState() : m_indexGenerator(1024, 1) {
+    TextureState::TextureState() : m_contextId(AllocateContextId()), m_indexGenerator(1024, 1) {
         // GL 3.3 core 3.8: each target owns one default texture object (name 0) per context,
         // shared across all texture units, and it is the initial binding of every unit/target
         // slot. It is created outside m_textureObjects so name-based paths (glIsTexture,

MobileGL/MG_State/GLState/TextureState/TextureState.h

@@ -84,8 +84,33 @@ namespace MobileGL::MG_State::GLState {
         Uint64 GetTextureBindGeneration() const { return m_textureBindGeneration; }
         void BumpTextureBindGeneration() { ++m_textureBindGeneration; }
 
+        // Sibling of the bind generation for everything that changes WHICH native texture a
+        // backend ends up putting on a unit WITHOUT any binding moving. Two families feed it:
+        // a texture's SHAPE (internal format, stored level set, level range - all that
+        // mipmap-completeness is computed from) and any sampler object's parameters (MIN_FILTER
+        // decides whether the mip chain is read at all, and an incomplete-for-the-filter texture
+        // is deliberately left unbound so it samples as (0,0,0,1)). Deliberately coarse - ANY
+        // texture, ANY sampler - so that no mutation can slip past a per-unit binding memo; the
+        // setters that feed it all early-out when the value is unchanged, so the redundant
+        // glTexParameteri calls applications issue every frame do not churn it. Kept separate
+        // from the bind generation because the sampled texture SET is unaffected by these, and
+        // the Vulkan backend's set memo keys on that one.
+        Uint64 GetSamplingResolutionGeneration() const { return m_samplingResolutionGeneration; }
+        void BumpSamplingResolutionGeneration() { ++m_samplingResolutionGeneration; }
+
+        // Globally-unique, never-reused id of THIS texture state, i.e. of the context that owns
+        // it. Both generations above restart at 0 with a new context, so a backend memo keyed on
+        // them alone would accept a destroyed-and-recreated context whose counters happen to line
+        // up - and the heap address is no help, since a context freed and remade lands on it
+        // again (the unit tests do exactly that between cases).
+        Uint64 GetContextId() const { return m_contextId; }
+
     private:
+        static Uint64 AllocateContextId();
+
+        const Uint64 m_contextId;
         Uint64 m_textureBindGeneration = 0;
+        Uint64 m_samplingResolutionGeneration = 0;
         Int m_maxTouchedUnit = -1;
         Int m_activeTextureUnit = 0;
         Array<TextureUnit, MAX_TEXTURE_IMAGE_UNITS> m_textureUnits;

MobileGL/MG_State/GLState/TextureState/TextureUnit.cpp

@@ -8,6 +8,8 @@
 
 #include "TextureUnit.h"
 
+#include <MG_State/GLState/Core.h>
+
 namespace MobileGL::MG_State::GLState {
     TextureUnit::TextureUnit() : m_sampler(nullptr) {
         for (int i = 0; i < (int)TextureTarget::TextureTargetCount; ++i) {
@@ -24,7 +26,17 @@ namespace MobileGL::MG_State::GLState {
     }
 
     void TextureUnit::SetSamplerObject(const SharedPtr<SamplerObject>& sampler) {
+        if (m_sampler == sampler) return;
+
         m_sampler = sampler;
+        // Which sampler object a unit carries is part of "what is bound at this unit": it
+        // overrides the texture's own sampler state, so it selects the filter that decides a
+        // bound texture's mipmap-completeness. glBindSampler already bumps the generation
+        // through NoteUnitTouched, but glDeleteSamplers unbinds the deleted object from every
+        // unit straight through here (GLContext::MarkSamplerObjectForDeletion) and would
+        // otherwise leave a backend memo of the resolved per-unit bindings replaying the
+        // deleted sampler.
+        if (pGLContext) pGLContext->BumpTextureBindGeneration();
     }
 
     const SharedPtr<SamplerObject>& TextureUnit::GetSamplerObject() const {

MobileGL/MG_State/GLState/VertexArrayState/VertexArrayObject.cpp

@@ -8,7 +8,18 @@
 
 #include "VertexArrayObject.h"
 
+#include <atomic>
+
 namespace MobileGL::MG_State::GLState {
+    // Starts at 1 so a zero-initialized memo slot can never carry a live object's id.
+    // Atomic because VAOs are GL-thread-only today but the counter costs nothing to
+    // make safe, and a duplicate id would resurrect exactly the bug it exists to kill.
+    static std::atomic<Uint64> s_nextVertexArrayLifetimeId{1};
+
+    Uint64 VertexArrayObject::AllocateLifetimeId() {
+        return s_nextVertexArrayLifetimeId.fetch_add(1, std::memory_order_relaxed);
+    }
+
     VertexArrayObject::VertexArrayObject(Uint externIndex) : m_externalIndex(externIndex) {
         for (int index = 0; index < MAX_VERTEX_ATTRIBS; ++index) {
             auto& attr = m_attributes[index];