[Fix] (git) Added the ignored item .gradle

Initialize the desktop replay Request from the three iterationRP environment flags before loading MobileGL. Without this, the Request defaults caused the replay core to unset CI's exported flags, leaving all repairs disabled despite the workflow configuration.

.github/workflows/apk.yml

@@ -209,7 +209,7 @@ jobs:
       - name: Load trace cases
         id: trace-cases
         run: |
-          echo "android=$(python3 tools/trace_replay/trace_cases.py --ci --format github-apk)" >> "$GITHUB_OUTPUT"
+          echo "android=$(python3 tools/trace_replay/trace_cases.py --ci --format github-apk-matrix)" >> "$GITHUB_OUTPUT"
           echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
 
   trace-fixtures:
@@ -337,13 +337,7 @@ jobs:
     strategy:
       fail-fast: false
       max-parallel: 4
-      matrix:
+      matrix: ${{ fromJSON(needs.trace-cases.outputs.android) }}
-        backend:
-          - name: DirectGLES
-            gpu: software
-          - name: DirectVulkan
-            gpu: lavapipe
-        case: ${{ fromJSON(needs.trace-cases.outputs.android) }}
     steps:
       - name: Set Swap Space
         uses: pierotofy/set-swap-space@v1.0
@@ -426,6 +420,9 @@ jobs:
           MOBILEGL_USE_ANGLE: ${{ matrix.backend.name == 'DirectGLES' && '1' || '0' }}
           MOBILEGL_TRACE_ANGLE_VARIANT: ${{ matrix.case.name == 'minecraft-1.21.4-fabric-iris-bliss-in-world' && '90a62123d794' || 'ec889e6ea831' }}
           MOBILEGL_MAGMA_R11G11B10F_FALLBACK: ${{ matrix.backend.name == 'DirectVulkan' && '1' || '0' }}
+          MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH: ${{ matrix.backend.name == 'DirectVulkan' && matrix.case.name == 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' && '1' || '0' }}
+          MOBILEGL_DERIVE_NUM_SUBGROUPS: ${{ matrix.backend.name == 'DirectVulkan' && matrix.case.name == 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' && '1' || '0' }}
+          MOBILEGL_ITERATIONRP_FIX_BARRIER: ${{ matrix.backend.name == 'DirectVulkan' && matrix.case.name == 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' && '1' || '0' }}
         run: |
           apk_file="android-retrace-apks/MobileGL-plugin-trace-release-${GITHUB_SHA}.apk"
           test -f "${apk_file}"

.github/workflows/test.yml

@@ -265,6 +265,9 @@ jobs:
         # crash stack without burning a CI round on an in-workflow debugger.
         env:
           MOBILEGL_ITEST_REQUIRE_GPU: "1"
+          MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH: "1"
+          MOBILEGL_DERIVE_NUM_SUBGROUPS: "1"
+          MOBILEGL_ITERATIONRP_FIX_BARRIER: "1"
         run: |
           ulimit -c unlimited
           sudo sysctl -w kernel.core_pattern='/tmp/core.%e.%p'
@@ -491,6 +494,7 @@ jobs:
       - benchmark
       - integration
     outputs:
+      matrix: ${{ steps.trace-cases.outputs.matrix }}
       names: ${{ steps.trace-cases.outputs.names }}
     steps:
       - name: Checkout repo
@@ -498,7 +502,9 @@ jobs:
 
       - name: Load trace cases
         id: trace-cases
-        run: echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
+        run: |
+          echo "matrix=$(python3 tools/trace_replay/trace_cases.py --ci --format github-test-matrix)" >> "$GITHUB_OUTPUT"
+          echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
 
   trace-fixtures:
     name: trace fixture (${{ matrix.case }})
@@ -577,11 +583,7 @@ jobs:
     strategy:
       fail-fast: false
       max-parallel: 4
-      matrix:
+      matrix: ${{ fromJSON(needs.trace-cases.outputs.matrix) }}
-        backend:
-          - DirectGLES
-          - DirectVulkan
-        case: ${{ fromJSON(needs.trace-cases.outputs.names) }}
 
     steps:
       - name: Set Swap Space
@@ -640,6 +642,12 @@ jobs:
           if [ '${{ matrix.backend }}' = 'DirectVulkan' ]; then
             export MOBILEGL_MAGMA_R11G11B10F_FALLBACK=1
           fi
+          if [ '${{ matrix.backend }}' = 'DirectVulkan' ] \
+             && [ '${{ matrix.case }}' = 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' ]; then
+            export MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH=1
+            export MOBILEGL_DERIVE_NUM_SUBGROUPS=1
+            export MOBILEGL_ITERATIONRP_FIX_BARRIER=1
+          fi
           # The blended depth-write quirk auto-enables only on Qualcomm, which no CI
           # runner has, so force it on for the OIT case it exists to fix. ForceOn
           # bypasses only the vendor gate, so this exercises the real strip on

.gitignore

@@ -27,3 +27,4 @@ MobileGL/MG*/cmake-build*
 tools/trace_replay/work/
 __pycache__/
 *.py[cod]
+/.gradle

CMakeLists.txt

@@ -182,6 +182,7 @@ set(ENABLE_SPVREMAPPER         OFF CACHE BOOL "Enable SPVRemapper" FORCE)
 set(ENABLE_OPT                 ON  CACHE BOOL "Enable SPIRV-Tools opt usage in glslang" FORCE)
 set(BUILD_EXTERNAL             ON  CACHE BOOL "Build external deps in External/" FORCE)
 set(ENABLE_GLSLANG_INSTALL     OFF CACHE BOOL "Install glslang targets" FORCE)
+set(SPIRV_SKIP_EXECUTABLES     ON  CACHE BOOL "Skip building SPIRV-Tools executables" FORCE)
 
 set(SPIRV_CROSS_C_API          ON  CACHE BOOL "Enable C API" FORCE)
 set(SPIRV_CROSS_ENABLE_GLSL    ON  CACHE BOOL "Enable GLSL backend" FORCE)
@@ -283,7 +284,13 @@ set(SOURCE_FILES
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/SplitArrayVertexInputsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RebaseInstanceIndexPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/ZeroBaseVertexPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DeriveNumSubgroupsPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/FixIterationRPBarrierPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/FixIterationRPSubgroupScratchPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateSubgroupsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/NormalizeRectCoordinatesPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/Lower1DArrayImagesPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/BakeImageFormatsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/PrivateToEntryLocalPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripUniformLocationsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripUboMemberRelaxedPrecisionPass.cpp
@@ -295,6 +302,7 @@ set(SOURCE_FILES
     MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.cpp
 
     MobileGL/MG_Util/SelfTest/DriverPost.cpp
+    MobileGL/MG_Util/SelfTest/DriverPostIterationRPWitness.cpp
 
     MobileGL/MG_Util/Texture/PixelStoreProcessor.cpp
     MobileGL/MG_Util/Texture/TextureFormatProcessor.cpp
@@ -455,7 +463,7 @@ set(MOBILEGL_INCLUDE_DIR
         # 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
+        ${CMAKE_SOURCE_DIR}/3rdparty/asio/include
 )
 
 add_library(${CMAKE_PROJECT_NAME} SHARED
@@ -667,3 +675,10 @@ if (NOT ANDROID)
         add_subdirectory(tools/trace_replay)
     endif()
 endif()
+
+# The integration binary is also useful as a standalone adb-shell executable.
+# Android cannot use the desktop-only MobileGL_s target, so its CMake module
+# links libMobileGL.so and creates an AImageReader-backed window instead.
+if (ANDROID AND MOBILEGL_BUILD_INTEGRATION_TEST)
+    add_subdirectory(MobileGL/MG_IntegrationTest)
+endif()

MobileGL/Config.h

@@ -66,22 +66,53 @@ 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_VALIDATE_SPIRV: test suites like SpirvPassTest exercise
-    //     ShaderCompiler without ever running MobileGL::Initialize(), and every
-    //     Initialize() re-runs MG_ConfigLoader::Init, which would clobber a
-    //     programmatic override stored here (see ShaderCompiler.cpp,
-    //     SpirvValidationEnabled).
     struct FeaturesTable {
         // MOBILEGL_DISABLE_TIMERQUERY: do not advertise or use GPU timer queries.
         Bool DisableTimerQuery = false;
+        // MOBILEGL_ENABLE_SPIRV_VALIDATION: validate generated and transformed SPIR-V.
+        // Disabled by default because validation is a diagnostics-only cost.
+        Bool EnableSpirvValidation = false;
         // MOBILEGL_USE_ANGLE: load ANGLE EGL/GLES libraries.
         Bool UseAngle = false;
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS)
         // MOBILEGL_TRACE_ANGLE_VARIANT: signed trace-APK ANGLE build short hash.
         String TraceAngleVariant;
 #endif
-        // MOBILEGL_DISABLE_SUBGROUP: force-disable Vulkan shader subgroup support.
+        // MOBILEGL_DISABLE_SUBGROUP: force-disable Vulkan shader subgroup support,
+        // including the opt-in emulated compute path below.
         Bool DisableSubgroup = false;
+        // MOBILEGL_MAGMA_EMULATE_SUBGROUP: implement GL_KHR_shader_subgroup's compute
+        // stage on a 32-lane VIRTUAL subgroup lowered to workgroup-shared memory
+        // (ShaderTranspiler::EmulateSubgroupsPass). Strictly a last resort: it only ever
+        // engages when this flag is set AND the device has no native subgroup support at
+        // all - a device with real subgroup operations always uses them natively,
+        // whatever their width (the known iterationRP defect is patched by
+        // FixIterationRPSubgroupScratch below instead). Off by default.
+        Bool MagmaEmulateSubgroup = false;
+        // MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH: patch iterationRP's own bug - the
+        // pack declares `shared vec2 prefixSumCache[32]` for a 512-invocation exposure
+        // reduction and indexes it by gl_SubgroupID, so any device with sub-16-lane
+        // subgroups (8-lane lavapipe -> 64 subgroups) writes shared memory out of
+        // bounds. The pass grows that one array to what the device's topology needs and
+        // touches nothing else; it only rewrites modules positively matching the pack's
+        // reduction fingerprint (ShaderTranspiler::FixIterationRPSubgroupScratchPass),
+        // so every other shader passes through byte-identical - as does iterationRP
+        // itself on >= 16-lane devices. Auto is ON; ForceOff replays the pack's bug
+        // verbatim.
+        QuirkOverride FixIterationRPSubgroupScratch = QuirkOverride::Auto;
+        // MOBILEGL_ITERATIONRP_FIX_BARRIER: repair Program 203's missing workgroup
+        // rendezvous between its two reductions over prefixSumCache. Off by default and
+        // fingerprint-gated by FixIterationRPBarrierPass when enabled.
+        Bool IterationRPFixBarrier = false;
+        // MOBILEGL_DERIVE_NUM_SUBGROUPS: replace compute gl_NumSubgroups loads with
+        // ceil(workgroup invocations / gl_SubgroupSize) on the NATIVE subgroup path
+        // (ShaderTranspiler::DeriveNumSubgroupsPass). Auto is ON: GL requires
+        // gl_SubgroupID < gl_NumSubgroups, Adreno's builtin reports 1 while the same
+        // dispatch emits IDs 0..7, and the derived value is the one Vulkan guarantees
+        // whenever the pipeline can request REQUIRE_FULL_SUBGROUPS (which the renderer
+        // does whenever local_size_x is a multiple of the native width). ForceOff returns
+        // to the raw driver builtin.
+        QuirkOverride DeriveNumSubgroups = QuirkOverride::Auto;
         // MOBILEGL_ADVERTISE_FP64: add GL_ARB_gpu_shader_fp64 to the advertised extension
         // string. `double` in a shader always WORKS - it is narrowed to 32 bits before any
         // module reaches a backend (ShaderTranspiler::DemoteFloat64Pass) - but the extension
@@ -117,16 +148,19 @@ namespace MobileGL::MG_Config {
         // per-draw glBufferSubData path instead of the persistent-mapped ring allocator
         // (negative control / driver-bug escape hatch).
         Bool DisableUboRing = false;
+        // MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION: make DirectGLES skip the native ES
+        // depth/stencil reads and always go through the shader-sampling emulation. Core GL
+        // ES has no depth or stencil readback, but some drivers accept it anyway (Mesa does,
+        // Adreno does not), which means the emulation is dead code on exactly the stack the
+        // headless suite runs on. This forces it live so the scenarios and the CTS can
+        // exercise the path, and gives the device an A/B lever over the same choice.
+        Bool EsprytForceDepthStencilReadbackEmulation = false;
         // MOBILEGL_RELAXED_SEMANTICS: relax strict core-profile rules (e.g. VAO-0 draws,
         // texture-name reuse after delete) even on contexts that explicitly requested a core
         // profile. Without it, relaxed semantics still apply to every context that did not
         // explicitly request a core profile via EGL_CONTEXT_OPENGL_PROFILE_MASK / a >=3.1
         // version request.
         Bool RelaxedSemantics = false;
-        // MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN: overrides the shader-source quirk that
-        // rewrites the recognized workgroup prefix-scan template on Qualcomm devices with
-        // subgroups wider than 32 lanes (see ShaderSourceProcessor's quirk registry).
-        QuirkOverride SubgroupPrefixScanQuirk = QuirkOverride::Auto;
         // MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE: overrides the DirectVulkan quirk that
         // strips depth writes from accumulation-blended pipelines (MIN/MAX or additive
         // ONE+ONE - the multi-pass depth-equality signature) on drivers without

MobileGL/ConfigLoader.cpp

@@ -162,11 +162,17 @@ namespace MobileGL::MG_ConfigLoader {
     inline void InitFeatures() {
         auto& features = MG_Config::Features;
         features.DisableTimerQuery = QueryEnvFlag("MOBILEGL_DISABLE_TIMERQUERY");
+        features.EnableSpirvValidation = QueryEnvFlag("MOBILEGL_ENABLE_SPIRV_VALIDATION");
         features.UseAngle = QueryEnvFlag("MOBILEGL_USE_ANGLE");
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS)
         QueryEnvVariable("MOBILEGL_TRACE_ANGLE_VARIANT", features.TraceAngleVariant, "");
 #endif
         features.DisableSubgroup = QueryEnvFlag("MOBILEGL_DISABLE_SUBGROUP");
+        features.MagmaEmulateSubgroup = QueryEnvFlag("MOBILEGL_MAGMA_EMULATE_SUBGROUP");
+        features.FixIterationRPSubgroupScratch =
+            QueryEnvQuirkOverride("MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH");
+        features.IterationRPFixBarrier = QueryEnvFlag("MOBILEGL_ITERATIONRP_FIX_BARRIER");
+        features.DeriveNumSubgroups = QueryEnvQuirkOverride("MOBILEGL_DERIVE_NUM_SUBGROUPS");
         features.AdvertiseFp64 = QueryEnvFlag("MOBILEGL_ADVERTISE_FP64");
         features.MagmaR11G11B10FFallback = QueryEnvFlag("MOBILEGL_MAGMA_R11G11B10F_FALLBACK");
         features.MagmaFramesInFlight = QueryEnvUint32("MOBILEGL_MAGMA_FRAMESINFLIGHT", 3, 1, 64);
@@ -176,8 +182,9 @@ namespace MobileGL::MG_ConfigLoader {
         features.CoherentAsFlush = QueryEnvFlag("MOBILEGL_COHERENT_AS_FLUSH");
         features.TraceSkipAutodestroy = QueryEnvFlag("MOBILEGL_TRACE_SKIP_AUTODESTROY");
         features.DisableUboRing = QueryEnvFlag("MOBILEGL_DISABLE_UBO_RING");
+        features.EsprytForceDepthStencilReadbackEmulation =
+            QueryEnvFlag("MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION");
         features.RelaxedSemantics = QueryEnvFlag("MOBILEGL_RELAXED_SEMANTICS");
-        features.SubgroupPrefixScanQuirk = QueryEnvQuirkOverride("MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN");
         features.MagmaDisableBlendedDepthWriteQuirk =
             QueryEnvQuirkOverride("MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE");
         features.DisableRobustBufferAccess = QueryEnvFlag("MOBILEGL_DISABLE_ROBUST_BUFFER_ACCESS");

MobileGL/Defines.h

@@ -52,11 +52,15 @@
 // that includes Defines.h without Log.h both tokens would silently evaluate to 0 in the
 // preprocessor conditional - enabling the assert in exactly the INFO-level builds it is
 // documented to be compiled out of. Log.h redefines them identically, which is legal.
+//
+// Severity order, ascending: DEBUG < INFO < WARN < ERROR < FATAL. MOBILEGL_LOG_ACTIVE_LEVEL
+// names the lowest severity compiled in, so the production default INFO keeps I/W/E/F and
+// drops only D. Any edit here must be mirrored in Log.h.
 #ifndef MOBILEGL_LOG_LEVEL_DEBUG
 #define MOBILEGL_LOG_LEVEL_DEBUG 0
-#define MOBILEGL_LOG_LEVEL_WARN 1
+#define MOBILEGL_LOG_LEVEL_INFO 1
-#define MOBILEGL_LOG_LEVEL_ERROR 2
+#define MOBILEGL_LOG_LEVEL_WARN 2
-#define MOBILEGL_LOG_LEVEL_INFO 3
+#define MOBILEGL_LOG_LEVEL_ERROR 3
 #define MOBILEGL_LOG_LEVEL_FATAL 4
 #endif
 
@@ -91,6 +95,12 @@
 #endif
 
 // =============================== Utils ================================ //
+// Asserts are live in exactly the builds where MGLOG_D is live, i.e. DEBUG builds only;
+// an INFO build (the production default) compiles them out. DEBUG is the lowest severity
+// in the ordering above, so "ACTIVE <= DEBUG" is true only for ACTIVE == DEBUG - the same
+// gate MGLOG_D uses in Log.h. That equivalence is what makes this gate survive the
+// 2026-08-13 renumbering unchanged; the contract is and stays
+// "INFO builds: asserts OFF; DEBUG builds: asserts ON".
 #if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_DEBUG
     #define MOBILEGL_ASSERT(condition, ...)                                                                                \
         do {                                                                                                               \

MobileGL/Init.cpp

@@ -15,6 +15,7 @@
 #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_Impl/GLImpl/Query/GL_Query.h>
 #include <MG_Util/Async/ShaderCompilePool.h>
 #include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
 
@@ -51,6 +52,11 @@ namespace MobileGL {
             // before a re-initialized library could pair them with the wrong
             // backend's DeleteSync).
             MG_Impl::GLImpl::DestroyAllSyncObjects();
+            // Queries die with their contexts for the same reason, and their registry
+            // is the same shape of process-global map: drain it here too, while the
+            // function table can still pair each backend handle with the backend that
+            // minted it.
+            MG_Impl::GLImpl::DestroyAllQueryObjects();
             MG_Backend::pActiveBackendObject.reset();
             MG_State::pGLContext.reset();
             MG_State::pEGLContext.reset();

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -712,9 +712,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     {
                         .TargetGLVersion = {4, 0, 0},   // GL target version
                         .TargetGLSLVersion = {4, 6, 0}, // Target Shading Language Version
-                        // Baseline advertisement (no timer queries / anisotropy yet); reconciled
+                        // Baseline advertisement (no runtime capabilities yet); reconciled once
-                        // once the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
+                        // the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
-                        .Extensions = BuildAdvertisedExtensions(false, false),
+                        .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                         .IsCompatibilityProfile = false // Is Compatibility Profile
                     },
                 .StaticBackendCapability = {.AllowVSOnlyPrograms = false} // Backend Capability
@@ -734,9 +734,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // thread can only observe the extension string after the
         // advertisement for its context has settled; rebuilding the whole
         // list keeps the re-run after a context recreation idempotent.
-        void UpdateAdvertisedCapabilityExtensions(Bool anisotropicFilteringSupported) {
+        void UpdateAdvertisedCapabilityExtensions(const MG_External::GLESCapabilities& capabilities) {
-            MutableRendererInfo().RendererGLInfo.Extensions =
+            MutableRendererInfo().RendererGLInfo.Extensions = BuildAdvertisedExtensions(
-                BuildAdvertisedExtensions(AreTimerQueriesSupported(), anisotropicFilteringSupported);
+                AreTimerQueriesSupported(), capabilities.SupportsTextureFilterAnisotropy,
+                capabilities.SupportsDrawIndirect,
+                capabilities.SupportsDrawIndirect && capabilities.SupportsBaseInstance);
         }
     } // namespace
 
@@ -779,11 +781,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return false;
         }
         DirectGLES::SetGLESCapabilities(m_GLESCapabilities);
-        // Now that g_GLESCapabilities knows about GL_EXT_disjoint_timer_query and
+        // Now that g_GLESCapabilities knows the host extensions, entry points, and ES version,
-        // GL_EXT_texture_filter_anisotropic, reconcile the advertisement (see the comment on
+        // reconcile every runtime-gated advertisement (see the comment on
-        // UpdateAdvertisedCapabilityExtensions for why it cannot happen when the extension
+        // UpdateAdvertisedCapabilityExtensions for why this cannot happen when the list is first
-        // list is first built).
+        // built).
-        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities.SupportsTextureFilterAnisotropy);
+        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities);
         UpdateDynamicBackendParameters();
         PopulateFormatCapabilities(m_GLESFunctions, m_GLESCapabilities, MutableFormatCapabilities());
         PrintFormatCapabilities(GetFormatCapabilities());
@@ -924,11 +926,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
         return MutableRendererInfo();
     }
 
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported) {
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
-            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_EXT_framebuffer_object,
+            E_GL_ARB_clear_buffer_object, E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_EXT_framebuffer_object,
             E_GL_ARB_depth_texture, E_GL_ARB_buffer_storage, E_GL_ARB_texture_storage,
             E_GL_ARB_texture_storage_multisample, E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access,
             E_GL_ARB_multi_draw_indirect, E_GL_ARB_indirect_parameters, E_GL_ARB_shader_draw_parameters,
@@ -940,10 +944,34 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // picks a whole different shader for draw_buffers without
             // explicit_attrib_location. DirectVulkan advertises both.
             E_GL_ARB_explicit_attrib_location, E_GL_ARB_texture_multisample, E_GL_ARB_shader_image_size,
+            // Core since GL 3.1 and implemented for every version advertised here. The string
+            // matters because applications gate the ENTRY POINTS on it rather than on the
+            // version: a caller that finds the extension missing never resolves
+            // glGetUniformBlockIndex / glUniformBlockBinding, and one that then uses uniform
+            // blocks anyway calls through a null pointer.
+            E_GL_ARB_uniform_buffer_object,
+            // Sampling the stencil aspect through DEPTH_STENCIL_TEXTURE_MODE. Core from 4.3,
+            // so on a 4.0 context the string is the only way to reach it. The host ES driver
+            // has had the same texture parameter since ES 3.1, which every device MobileGL
+            // runs on provides.
+            E_GL_ARB_stencil_texturing,
             // Advertised with GL_NUM_PROGRAM_BINARY_FORMATS = 0, which the
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Minecraft 26.3 checks this prerequisite before it even considers
+        // GL_ARB_multi_draw_indirect. ES 3.1 supplies both single-draw entry points; the loader
+        // folds the version and pointer checks into SupportsDrawIndirect.
+        if (drawIndirectSupported) {
+            extensions.push_back(E_GL_ARB_draw_indirect);
+        }
+        // ARB_base_instance also defines the last word of an indirect command. Direct calls are
+        // emulated on every Espryt device, but without host GL_EXT_base_instance a native indirect
+        // draw cannot shift divisor attributes by a GPU-authored non-zero value, so do not promise
+        // that incomplete case.
+        if (drawIndirectSupported && nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         // 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

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.h

@@ -67,9 +67,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
     const RendererInfo& GetRendererIdentity();
 
     // The full OpenGL extension list Espryt advertises (glGetString(GL_EXTENSIONS))
-    // for a device whose timer queries / anisotropic filtering are (or are not) usable.
+    // for a device whose timer queries / anisotropic filtering / native indirect draws /
+    // non-zero indirect baseInstance semantics are (or are not) usable.
     // The MOBILEGL_DISABLE_TIMERQUERY escape hatch is applied inside.
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported);
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <OpenGL ES Renderer>, OpenGL ES <Major>.<Minor> — the exact string an
     // initialized backend returns from GetBackendAPIVersionString (and that ends up

MobileGL/MG_Backend/DirectGLES/Managers.h

@@ -286,6 +286,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // context loss.
             Bool persistentMapped = false;
             void* persistentPtr = nullptr;
+            // The GL store behind `id` was created with glBufferStorageEXT and is
+            // therefore IMMUTABLE - glBufferData cannot respecify it and it must never be
+            // recycled through the size-keyed buffer pool. Tracked separately from
+            // persistentMapped because the two come apart: a glMapBufferRange that fails
+            // after its glBufferStorageEXT succeeded leaves immutable storage behind with
+            // no map, and a respecification then has to retire the id rather than hand it
+            // to glBufferData, which the driver would silently refuse.
+            Bool immutableStorage = false;
         };
 
         // Registered as the frontend's BufferBackendOps at backend init and on
@@ -338,6 +346,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // client-attribute staging buffers): scrub every buffer-binding shadow that
         // could false-skip when the name is recycled.
         void NoteBufferIdDeleted(Uint id);
+        // Bumped whenever a live GLESBufferResource's driver id is retired and re-minted
+        // while its frontend buffer stays alive (persistent-map adoption, immutable-store
+        // retire). The VAO twins' baked glVertexAttribPointer / element-array bindings
+        // key on FRONTEND versions, which a backend-side re-mint does not move - without
+        // this generation the driver VAO would keep fetching through the deleted id (or
+        // its retained store) forever. Compared and stamped by
+        // BackendVertexArrayObject::SyncToBackend.
+        extern Uint64 g_bufferBackendIdGeneration;
         // Redundant-bind cache for INDEXED buffer bindings (glBindBufferBase/Range on
         // GL_UNIFORM_BUFFER / GL_SHADER_STORAGE_BUFFER): skips the GL call when the
         // (id, range) already at that index matches, like the array-buffer/texture/
@@ -457,6 +473,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Array<Uint, MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS> m_clientAttributeBufferIds;
             Bool m_isInitialized = false;
             Uint16 m_syncedIndexBufferVersion = 0;
+            // Identity of the buffer the version above was stamped against. Raw and never
+            // dereferenced: the slot version is a wrapping Uint16 (see the ResolvedDrawBuffers
+            // IBO memo and the packed_pixels postmortem at BindCurrentFBO), so the version
+            // alone would read a wrapped-back count with a different buffer bound as clean.
+            const MG_State::GLState::BufferObject* m_syncedIndexBufferObject = nullptr;
             // 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
@@ -472,6 +493,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // Kept here because it describes what was last EMITTED, which is what the next sync
             // has to correct.
             Uint32 m_syncedFetchBaseInstance = 0;
+            // BufferImpl::g_bufferBackendIdGeneration as of this twin's last emit. A
+            // mismatch means some live buffer's driver id was re-minted since; the ids
+            // baked into the driver VAO's attribute/element bindings may be dead even
+            // though every frontend version matches, so the next sync re-emits them all.
+            Uint64 m_syncedBufferIdGeneration = 0;
         };
 
         extern StateBackendObjectRegistry<MG_State::GLState::VertexArrayObject, BackendVertexArrayObject>
@@ -691,8 +717,18 @@ namespace MobileGL::MG_Backend::DirectGLES {
             FloatVec4 m_cacheBorderColor = {0.0f, 0.0f, 0.0f, 0.0f};
             Vec4<TextureSwizzleParam> m_cacheSwizzleParams = {TextureSwizzleParam::Red, TextureSwizzleParam::Green,
                                                               TextureSwizzleParam::Blue, TextureSwizzleParam::Alpha};
+            // GL_DEPTH_STENCIL_TEXTURE_MODE. GL_DEPTH_COMPONENT is the GL and ES default, so a
+            // texture that never asks for the stencil aspect never emits the call. The
+            // depth/stencil readback and replicate-blit emulations also write this parameter
+            // raw, but only ever on their own scratch textures (never on an application
+            // texture), so they cannot desynchronise this cache.
+            GLenum m_cacheDepthStencilTextureMode = GL_DEPTH_COMPONENT;
             Uint16 m_syncedSamplerVersion = 0;
             Uint16 m_syncedTextureParamsVersion = 0;
+            // Set when the driver texture underneath was regenerated and has therefore lost every
+            // parameter already pushed onto it: the params-version early-out has to be overridden
+            // once, or an unchanged version would skip the re-push forever.
+            Bool m_forceTextureParamsResync = false;
         };
 
         void ActivateTextureUnit(Uint unit);
@@ -781,6 +817,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
             using FramebufferObject = MG_State::GLState::FramebufferObject;
             FramebufferObject::FramebufferAttachmentVersionArray m_syncedFrontendAttachmentVersions = {0};
+            // g_attachmentBackendIdGeneration as of this twin's last attachment walk. A
+            // mismatch means some backend texture id was re-minted since, and any of this
+            // twin's attachment points may still hold the dead id even though the frontend
+            // attachment versions match - so the walk re-attaches everything first.
+            Uint64 m_syncedBackendIdGeneration = 0;
         };
 
         extern StateBackendObjectRegistry<MG_State::GLState::FramebufferObject, BackendFramebufferObject>
@@ -870,6 +911,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
         extern Array<MG_State::GLState::FramebufferObject*, SizeT(FramebufferTarget::FramebufferTargetCount)>
             g_fboSyncedObjects;
 
+        // Bumped whenever a live backend texture's driver id is re-minted while its
+        // frontend texture may still be attached to application FBOs
+        // (BackendTextureObject::RecreateBackendTexture - e.g. a respecify of a texture
+        // whose backend storage went immutable). The FBO twins' attachment memos key on
+        // FRONTEND attachment versions, which a backend-side re-mint does not move, so
+        // the driver FBO would keep the deleted texture name attached forever. The
+        // SyncCurrentFBO gate compares this generation (below) to re-enter the sync,
+        // and each twin re-arms its per-attachment memo on a mismatch (SyncToBackend).
+        extern Uint64 g_attachmentBackendIdGeneration;
+        // What g_attachmentBackendIdGeneration was when SyncCurrentFBO last stamped each
+        // target; part of the synced tuple above.
+        extern Array<Uint64, SizeT(FramebufferTarget::FramebufferTargetCount)> g_fboSyncedBackendIdGenerations;
+
         // Driver-level READ/DRAW framebuffer-binding shadow. Every backend
         // glBindFramebuffer routes through BindFramebufferId so scoped helpers can
         // save/restore the current binding without a glGetIntegerv round-trip (that
@@ -1104,6 +1158,28 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // stale as one built before a relink - while the sampler half, which really is
             // re-issued per draw, needs nothing of the sort.
             Uint32 GetSyncedImageUnitVersion() const { return m_syncedImageUnitVersion; }
+            // Whether the (unit, bound format) pairs this program's FORMAT-LESS image uniforms
+            // resolve to are still the ones its ESSL was generated against.
+            //
+            // A fourth condition of the same family as the three above, and the only one that
+            // reads live state rather than a program-side counter, because that is where the
+            // dependency actually is. GLSL ES requires a format layout qualifier on every image
+            // where desktop GLSL lets a writeonly declaration omit one, and the only correct
+            // qualifier is whatever glBindImageTexture named - so a declaration with no format
+            // is compiled against the BINDING, and a rebind to a different format makes the
+            // built program wrong. Keyed on the units the program's own images address (cached
+            // at sync, since a unit can only move by glUniform1i, which bumps the image-unit
+            // version above and forces a re-sync anyway), so the cost on a program with no
+            // format-less image - which is all but a handful - is one empty-vector test.
+            //
+            // Deliberately NOT reached from glBindImageTexture: that entry point must never
+            // trigger a build (same constraint as glShaderStorageBlockBinding). It moves the
+            // state and this comparison notices at the next Prepare, which is also what makes
+            // an image first bound AFTER link work.
+            Bool ImageUnitFormatsStillMatch() const;
+            // The value ImageUnitFormatsStillMatch() compares against, recomputed from live
+            // image-unit state. 0 when the program has no format-less image uniform.
+            Uint64 ComputeImageUnitFormatSignature() const;
 
         private:
             void CacheResourceLocations(const SharedPtr<MG_State::GLState::ProgramObject>& stateProgramObject);
@@ -1137,6 +1213,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
             BufferImpl::UboRingAllocation m_globalUboRingAllocation;
             Uint32 m_syncedLinkVersion = ~0u;
             Uint32 m_syncedImageUnitVersion = ~0u;
+            // Image units addressed by the program's FORMAT-LESS image uniforms, and the digest
+            // of the (unit, format) pairs the generated ESSL baked. Empty/0 for every program
+            // that declares a format on all of its images, which is the overwhelming majority -
+            // and what keeps the per-draw comparison free for them.
+            Vector<Int> m_formatlessImageUnits;
+            Uint64 m_imageUnitFormatSignature = 0;
             SamplerPassMemo m_samplerPassMemo;
         };
 
@@ -1180,6 +1262,40 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // already has costs nothing. 0 when nothing was ever rebound.
         Uint64 ComputeShaderStorageBlockBindingSignature(
             const MG_State::GLState::ProgramObject& stateProgramObject);
+
+        // Everything the image-format bake needs from one walk of a program's uniform
+        // reflection. GLSL ES requires a format layout qualifier on every image uniform;
+        // desktop GLSL lets a writeonly (or readonly) declaration omit one, and the only
+        // format that is CORRECT to substitute is whatever glBindImageTexture named for the
+        // unit that uniform addresses - so the transpile bakes it in and the build is keyed
+        // on it.
+        struct ImageFormatBakeInputs {
+            // Uniform name (SPIR-V spelling, i.e. an array named once, unsubscripted) to the GL
+            // internal format to bake. Holds only uniforms that DECLARED no format; a declared
+            // one is authoritative and is never overridden.
+            UnorderedMap<String, Uint> glFormatByUniformName;
+            // The same uniforms whose format SPIRV-Cross REFUSES to print for ESSL (it throws on
+            // its desktop-only set, which loses the stage), paired with the ESSL spelling to
+            // write into the emitted declaration instead. Disjoint from the map above by
+            // construction: a format is baked into the module or completed in the text, never
+            // both. r8ui - the stencil half of the packed_depth_stencil case - lands here.
+            UnorderedMap<String, String> esslFormatQualifierByUniformName;
+            // Units those uniforms address, kept so the draw path can re-read their formats
+            // without walking the reflection again.
+            Vector<Int> units;
+            // Digest of the (unit, format) pairs above. 0 when the program has no format-less
+            // image uniform, which is all but a handful.
+            Uint64 signature = 0;
+            // Array uniforms whose elements resolved to units holding DIFFERENT formats: one
+            // declaration carries one qualifier, so there is nothing correct to bake and they
+            // are dropped from the map above. Kept for diagnostics.
+            Vector<String> conflictedNames;
+            // Some format in play - declared or baked - is outside the GLSL ES core image
+            // format set, so the emitted ESSL needs the GL_NV_image_formats directive.
+            Bool needsExtendedImageFormats = false;
+        };
+        ImageFormatBakeInputs CollectImageFormatBakeInputs(
+            const MG_State::GLState::ProgramObject& stateProgramObject);
     } // namespace PrgramImpl
 
     namespace SamplerImpl {

MobileGL/MG_Backend/DirectGLES/MultiDraw.cpp

@@ -252,7 +252,7 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
             g_resolvedTier =
                 ResolveTier(g_GLESCapabilities, g_GLESFuncs, MG_Config::Features.EsprytMultiDrawMode,
                             &g_tierResolution);
-            MGLOG_I("DirectGLES multi-draw: %s", g_tierResolution.c_str());
+            MGLOG_D("DirectGLES multi-draw: %s", g_tierResolution.c_str());
         }
 
         // Which tiers have already announced themselves, one bit per GLESMultiDrawMode.
@@ -267,7 +267,7 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
             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));
+            MGLOG_D("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
@@ -490,7 +490,7 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
                 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 "
+                    MGLOG_E_ONCE("DirectGLES multi-draw (drawelements tier): sub-draw %d reads outside the bound index "
                             "buffer; skipping the batch",
                             i);
                     return false;
@@ -596,7 +596,7 @@ void main() {
 
             const GLuint shader = g_GLESFuncs.glCreateShader(GL_COMPUTE_SHADER);
             if (shader == 0) {
-                MGLOG_E("DirectGLES multi-draw (compute tier): glCreateShader(GL_COMPUTE_SHADER) failed");
+                MGLOG_E_ONCE("DirectGLES multi-draw (compute tier): glCreateShader(GL_COMPUTE_SHADER) failed");
                 return false;
             }
             const char* source = kFlattenComputeSource;
@@ -607,14 +607,14 @@ void main() {
             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);
+                MGLOG_E_ONCE("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");
+                MGLOG_E_ONCE("DirectGLES multi-draw (compute tier): glCreateProgram failed");
                 g_GLESFuncs.glDeleteShader(shader);
                 return false;
             }
@@ -625,7 +625,7 @@ void main() {
             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);
+                MGLOG_E_ONCE("DirectGLES multi-draw (compute tier): index-flattening program failed to link: %s", log);
                 g_GLESFuncs.glDeleteProgram(program);
                 return false;
             }
@@ -635,7 +635,7 @@ void main() {
             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);
+            MGLOG_D("DirectGLES multi-draw: index-flattening compute program ready (id %u)", program);
             return true;
         }
 
@@ -920,7 +920,7 @@ void main() {
                                            feedBaseVertex);
         }
         if (!drawn) {
-            MGLOG_E("DirectGLES multi-draw: no usable tier for a %d sub-draw batch (mode 0x%x, type 0x%x); "
+            MGLOG_E_ONCE("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);
         }

MobileGL/MG_Backend/DirectGLES/Utils.cpp

@@ -417,10 +417,27 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 result = std::regex_replace(result, pattern, "$1flat $2");
             };
 
+            // Every stage that has an integer interface at all, on BOTH sides. Interpolation is
+            // only ever consumed at a fragment input, so the qualifier is semantically inert on
+            // a tessellation or geometry interface - but an ES linker still compares the two
+            // sides of every interface and rejects a program whose producer says `flat` and
+            // whose consumer does not. Covering only the stages that "need" it left exactly two
+            // holes, and a program that used tessellation fell into both:
+            //   vertex `flat out uint` -> tess-control `in uint`   (producer flat, consumer not)
+            //   tess-eval `out uint`   -> geometry `flat in uint`  (consumer flat, producer not)
+            // Adreno answers "output ... interpolation mismatch with other stage" and the whole
+            // program fails to link, which is a draw that silently paints nothing.
+            //
+            // Adding rather than stripping, because a fragment input's `flat` is load-bearing
+            // (ESSL forbids an interpolated integer) and would have to be put back for the last
+            // stage before the fragment shader anyway - so "everything integer is flat" is the
+            // one rule that is consistent no matter which stages a program happens to have.
             switch (shaderType) {
             case GL_VERTEX_SHADER:
                 addFlatQualifier("out");
                 break;
+            case GL_TESS_CONTROL_SHADER:
+            case GL_TESS_EVALUATION_SHADER:
             case GL_GEOMETRY_SHADER:
                 addFlatQualifier("in");
                 addFlatQualifier("out");
@@ -518,6 +535,92 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return glslCode;
         }
 
+        String RequestExtendedImageFormats(String glslCode, Bool needed) {
+#ifdef TRACY_ENABLE
+            ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
+#endif
+            // GLSL ES core has thirteen image formats; GL has forty. SPIRV-Cross prints whatever
+            // format the OpTypeImage carries and asks for no extension for it, so an r8ui or
+            // rg16f image - declared as such, or baked from the bound one - reaches the driver as
+            // a format its core language does not know. GL_NV_image_formats is the only thing
+            // that adds them, and it has to be requested by name.
+            //
+            // The caller decides `needed`: it knows which formats are in play (from the uniform
+            // reflection and the image-unit bindings) and whether the driver advertises the
+            // extension at all - `#extension` on an unadvertised name is itself a hard error, so
+            // this must never be emitted speculatively.
+            static constexpr const char* kDirective = "#extension GL_NV_image_formats : require\n";
+            static constexpr const char* kExtName = "GL_NV_image_formats";
+            if (!needed || glslCode.find(kExtName) != String::npos) {
+                return glslCode;
+            }
+            // After the #version line, which must stay first. Everything else about the header is
+            // order-insensitive, and ForceSupporterOutput's scan for the LAST #extension
+            // directive still finds whichever one that is.
+            const SizeT versionPos = glslCode.find("#version");
+            if (versionPos == String::npos) {
+                return kDirective + glslCode;
+            }
+            const SizeT lineEnd = glslCode.find('\n', versionPos);
+            if (lineEnd == String::npos) {
+                return glslCode + "\n" + kDirective;
+            }
+            glslCode.insert(lineEnd + 1, kDirective);
+            return glslCode;
+        }
+
+        String BakeImageFormatQualifiers(String glslCode,
+                                         const UnorderedMap<String, String>& esslFormatByUniformName) {
+#ifdef TRACY_ENABLE
+            ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
+#endif
+            if (esslFormatByUniformName.empty() || glslCode.find("image") == String::npos) {
+                return glslCode;
+            }
+            // Same declaration shape RebindImageUniformsToFrontendUnits matches, and for the same
+            // reason: one line, one image uniform, the name in group 3.
+            static const std::regex imageDeclRegex(
+                R"((layout\s*\(([^)]*)\)\s*)?uniform\s+(?:(?:readonly|writeonly|coherent|volatile|restrict|highp|mediump|lowp)\s+)*[iu]?image[A-Za-z0-9]+\s+([A-Za-z_][A-Za-z0-9_]*)\s*(\[[^\]]*\])?\s*;)");
+            // Every image format spelling GLSL has, so a declaration that already carries one is
+            // recognised whatever it says - the caller's map is consulted only for declarations
+            // with NO format, never to override a written one.
+            static const std::regex existingFormatRegex(
+                R"(\b(rgba32f|rgba16f|rg32f|rg16f|r11f_g11f_b10f|r32f|r16f|rgba16|rgb10_a2|rg16|rg8|r16|r8|rgba16_snorm|rgba8_snorm|rg16_snorm|rg8_snorm|r16_snorm|r8_snorm|rgba32i|rgba16i|rgba8i|rg32i|rg16i|rg8i|r32i|r16i|r8i|rgba32ui|rgba16ui|rgba8ui|rgb10_a2ui|rg32ui|rg16ui|rg8ui|r32ui|r16ui|r8ui)\b)");
+
+            String result;
+            result.reserve(glslCode.size());
+            SizeT lineStart = 0;
+            while (lineStart <= glslCode.size()) {
+                const SizeT lineEnd = glslCode.find('\n', lineStart);
+                const Bool lastLine = lineEnd == String::npos;
+                String line = glslCode.substr(lineStart, lastLine ? String::npos : lineEnd - lineStart);
+
+                std::smatch match;
+                if (std::regex_search(line, match, imageDeclRegex)) {
+                    const String name = match[3].str();
+                    const auto formatIt = esslFormatByUniformName.find(name);
+                    const String layoutContents = match[2].matched ? match[2].str() : String();
+                    if (formatIt != esslFormatByUniformName.end() && !formatIt->second.empty() &&
+                        !std::regex_search(layoutContents, existingFormatRegex)) {
+                        if (match[1].matched) {
+                            const SizeT layoutOpen = line.find('(', match.position(1));
+                            line.insert(layoutOpen + 1, formatIt->second + ", ");
+                        } else {
+                            line.insert(match.position(0), "layout(" + formatIt->second + ") ");
+                        }
+                    }
+                }
+
+                result += line;
+                if (lastLine) {
+                    break;
+                }
+                result += '\n';
+                lineStart = lineEnd + 1;
+            }
+            return result;
+        }
+
         String RemoveLayoutBinding(const String& glslCode) {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
@@ -1073,7 +1176,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             for (GLenum err = g_GLESFuncs.glGetError(); err != GL_NO_ERROR; err = g_GLESFuncs.glGetError()) {
-                MGLOG_E("-> GLES Error: %s", MG_Util::ConvertGLEnumToString(err).c_str());
+                MGLOG_D("-> GLES Error: %s", MG_Util::ConvertGLEnumToString(err).c_str());
             }
         }
 
@@ -1493,26 +1596,18 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return (rowBytes + align - 1) / align * align;
         }
 
-    // Repacks wide RGBA(_INTEGER) rows into the client's (format, type) layout, honoring the
+        // Walks the client-side destination the PACK parameters describe and hands each row to
-        // client-side PACK parameters and the bound pixel-pack buffer. `wide` holds
+        // `fillRow(slice, row, dstRow)`, which writes width * dstPixelBytes bytes of finished client
-        // `sliceHeight * sliceCount` rows of `width` texels (slice-major, tightly stacked),
+        // texels. Shared by the converting and the raw-word stores so both address the destination -
-        // 4 components x GetReadbackComponentSize(wideType) bytes each.
+        // and feed the bound pixel-pack buffer - identically.
         // applyPackImageParams: GL_PACK_IMAGE_HEIGHT / GL_PACK_SKIP_IMAGES apply only to GetTexImage
         // of 3D/array images; ReadPixels and 2D GetTexImage ignore them (GL 3.3 sections 4.3.1, 6.1.4).
         // Per the GL addressing rules, slice k row j lands at
         // SKIP_IMAGES*imageStride + SKIP_ROWS*rowStride + SKIP_PIXELS*pixelBytes
         //   + k*imageStride + j*rowStride, with imageStride = max(IMAGE_HEIGHT, sliceHeight)*rowStride.
-        Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
+        template <typename FillRow>
-                                   GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
+        static Bool StoreClientRows(SizeT dstPixelBytes, SizeT swapGroupSize, GLsizei width, GLsizei sliceHeight,
-                                   void* pixels, Bool applyPackImageParams) {
+                                    GLsizei sliceCount, void* pixels, Bool applyPackImageParams, FillRow&& fillRow) {
-        const SizeT dstPixelBytes = GetReadbackDstPixelSize(mapping, type);
-        if (dstPixelBytes == 0) {
-            return false;
-        }
-        PackedReadbackLayout packedLayout{};
-        const Bool isPackedType = GetPackedReadbackLayout(type, packedLayout);
-        const SizeT dstComponentSize = GetReadbackComponentSize(type);
-
             const auto& pixelPackBufferObject =
                 MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelPack).GetBoundObject();
 
@@ -1539,29 +1634,20 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                         static_cast<SizeT>(sliceCount - 1) * dstImageStride +
                                         static_cast<SizeT>(sliceHeight - 1) * dstRowStride + dstRowBytes;
                 if (requiredSize > pixelPackBufferObject->GetSize()) {
-                MGLOG_E("Readback conversion: pixel pack buffer is too small");
+                    MGLOG_E_ONCE("Readback conversion: pixel pack buffer is too small");
                     return true;
                 }
             }
 
-        const SizeT srcComponentSize = GetReadbackComponentSize(wideType);
-        const SizeT srcPixelBytes = 4 * srcComponentSize;
             Vector<Uint8> convertedRow(dstRowBytes);
 
             for (GLsizei slice = 0; slice < sliceCount; ++slice) {
                 for (GLsizei row = 0; row < sliceHeight; ++row) {
-                const SizeT flatRow = static_cast<SizeT>(slice) * static_cast<SizeT>(sliceHeight) +
+                    fillRow(slice, row, convertedRow.data());
-                                   static_cast<SizeT>(row);
-                const Uint8* srcRow = wide + flatRow * static_cast<SizeT>(width) * srcPixelBytes;
-                ConvertWideReadbackRow(srcRow, convertedRow.data(), static_cast<SizeT>(width), wideType,
-                                                  mapping, type);
 
-                if (packParams.SwapBytes) {
+                    if (packParams.SwapBytes && swapGroupSize > 1) {
-                    const SizeT groupSize = isPackedType ? packedLayout.byteSize : dstComponentSize;
+                        for (SizeT offset = 0; offset + swapGroupSize <= dstRowBytes; offset += swapGroupSize) {
-                    if (groupSize > 1) {
+                            std::reverse(convertedRow.data() + offset, convertedRow.data() + offset + swapGroupSize);
-                        for (SizeT offset = 0; offset + groupSize <= dstRowBytes; offset += groupSize) {
-                            std::reverse(convertedRow.data() + offset, convertedRow.data() + offset + groupSize);
-                        }
                         }
                     }
 
@@ -1582,5 +1668,52 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
             return true;
         }
+
+        // Repacks wide RGBA(_INTEGER) rows into the client's (format, type) layout, honoring the
+        // client-side PACK parameters and the bound pixel-pack buffer. `wide` holds
+        // `sliceHeight * sliceCount` rows of `width` texels (slice-major, tightly stacked),
+        // 4 components x GetReadbackComponentSize(wideType) bytes each.
+        Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
+                                   GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
+                                   void* pixels, Bool applyPackImageParams) {
+            const SizeT dstPixelBytes = GetReadbackDstPixelSize(mapping, type);
+            if (dstPixelBytes == 0) {
+                return false;
+            }
+            PackedReadbackLayout packedLayout{};
+            const Bool isPackedType = GetPackedReadbackLayout(type, packedLayout);
+            const SizeT swapGroupSize = isPackedType ? packedLayout.byteSize : GetReadbackComponentSize(type);
+            const SizeT srcPixelBytes = 4 * GetReadbackComponentSize(wideType);
+
+            return StoreClientRows(dstPixelBytes, swapGroupSize, width, sliceHeight, sliceCount, pixels,
+                                   applyPackImageParams,
+                                   [&](GLsizei slice, GLsizei row, Uint8* dstRow) {
+                                       const SizeT flatRow = static_cast<SizeT>(slice) *
+                                                                 static_cast<SizeT>(sliceHeight) +
+                                                             static_cast<SizeT>(row);
+                                       const Uint8* srcRow =
+                                           wide + flatRow * static_cast<SizeT>(width) * srcPixelBytes;
+                                       ConvertWideReadbackRow(srcRow, dstRow, static_cast<SizeT>(width), wideType,
+                                                              mapping, type);
+                                   });
+        }
+
+        Bool StorePackedWordsToClient(const Uint8* srcWords, GLsizei width, GLsizei sliceHeight, GLsizei sliceCount,
+                                      GLenum type, void* pixels, Bool applyPackImageParams) {
+            PackedReadbackLayout packedLayout{};
+            if (!GetPackedReadbackLayout(type, packedLayout) || packedLayout.byteSize != 4) {
+                return false;
+            }
+            const SizeT srcRowBytes = static_cast<SizeT>(width) * 4;
+
+            return StoreClientRows(4, packedLayout.byteSize, width, sliceHeight, sliceCount, pixels,
+                                   applyPackImageParams,
+                                   [&](GLsizei slice, GLsizei row, Uint8* dstRow) {
+                                       const SizeT flatRow = static_cast<SizeT>(slice) *
+                                                                 static_cast<SizeT>(sliceHeight) +
+                                                             static_cast<SizeT>(row);
+                                       Memcpy(dstRow, srcWords + flatRow * srcRowBytes, srcRowBytes);
+                                   });
+        }
     } // namespace ReadbackImpl
 } // namespace MobileGL::MG_Backend::DirectGLES

MobileGL/MG_Backend/DirectGLES/Utils.h

@@ -115,6 +115,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
         Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
                                    GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
                                    void* pixels, Bool applyPackImageParams);
+
+        // Stores packed 32-bit source words verbatim, with the same destination addressing, PACK
+        // parameters and pixel-pack-buffer handling as StoreWideRowsToClient. For the sources whose
+        // storage word already IS the client word (MG_Util::IsRawPackedPixelTransfer): routing those
+        // through the wide float intermediate re-encodes them, and the RGB9_E5 encoder canonicalizes
+        // the shared exponent, so glGetTexImage would answer with different bits than were stored.
+        // `srcWords` holds sliceHeight * sliceCount tightly stacked rows of `width` 32-bit words.
+        // False when `type` is not a 4-byte packed type.
+        Bool StorePackedWordsToClient(const Uint8* srcWords, GLsizei width, GLsizei sliceHeight, GLsizei sliceCount,
+                                      GLenum type, void* pixels, Bool applyPackImageParams);
     } // namespace ReadbackImpl
 
     namespace PrgramImpl {
@@ -137,6 +147,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // ES 3.2 needs no directive at all and an EXT driver already has the right one.
         String RetargetTextureBufferExtension(String glslCode,
                                               MG_External::GLESCapabilities::TextureBufferTier tier);
+        // Adds `#extension GL_NV_image_formats : require` when the shader carries an image
+        // format qualifier GLSL ES has no core spelling for. SPIRV-Cross prints the format and
+        // asks for nothing, so the request has to be made here. `needed` is the caller's answer,
+        // because only it knows which formats are in play AND whether the driver advertises the
+        // extension - requesting an unadvertised extension is itself a compile error, so this is
+        // never emitted speculatively. A no-op when not needed or already present.
+        String RequestExtendedImageFormats(String glslCode, Bool needed);
+        // Writes a format layout qualifier into the image declarations named in
+        // `esslFormatByUniformName` that still have none. The completion half of the image-format
+        // bake, and ONLY that: the SPIR-V pass (BakeImageFormatsPass) is what normally puts the
+        // format in, but SPIRV-Cross throws rather than printing the formats it calls
+        // desktop-only when it targets ESSL - r8ui among them, which is what the stencil half of
+        // KHR-GL4x.packed_depth_stencil.stencil_texturing binds - and a throw loses the whole
+        // stage. So those formats stay out of the module and are spelled here instead, on the
+        // emitted text, where nothing can refuse them.
+        //
+        // Declarations that already carry a format are left exactly as they are, whoever wrote
+        // it. Must run before RemoveLayoutBinding, which is where an image's layout qualifier
+        // stops being safe to edit by hand.
+        String BakeImageFormatQualifiers(String glslCode, const UnorderedMap<String, String>& esslFormatByUniformName);
         String RemoveLayoutBinding(const String& glslCode);
         // Prefix of the writeonly half a read+write image uniform is split into (see
         // SplitReadWriteImageUniforms); the suffix is the image's own name.

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -9,7 +9,9 @@
 #include "BackendObject_DirectVulkan.h"
 #include "MG_Backend/BackendObject.h"
 #include "DirectVulkan.h"
+#include "SubgroupSupportPolicy.h"
 #include "MG_State/GLState/FramebufferState/FramebufferObject.h"
+#include "MG_State/GLState/Core.h"
 #include "MG_State/GLState/TextureState/TextureState.h"
 #include "MG_Util/Classifiers/TextureEnumClassifier.h"
 #include "MG_Util/Converters/MGToGL/TextureEnumConverter.h"
@@ -383,6 +385,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         UpdateDynamicBackendParameters();
         UpdateAdvertisedExtensions();
+        if (MG_State::pGLContext) {
+            MG_State::pGLContext->InvalidateCompileEnv();
+        }
         PopulateFormatCapabilities(physicalDevice.handle, vkGetPhysicalDeviceFormatProperties, m_vulkanCaps,
                                    MutableFormatCapabilities());
         PrintFormatCapabilities(GetFormatCapabilities());
@@ -497,30 +502,48 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             .ExtraVendor = Nullopt,
             .RendererGLInfo = {.TargetGLVersion = {4, 0, 0},
                                .TargetGLSLVersion = {4, 6, 0},
-                               // Baseline advertisement (no shader subgroup, no timer queries); a
+                               // Baseline advertisement (no runtime-gated capabilities); a live
-                               // live backend reconciles its copy in UpdateAdvertisedExtensions.
+                               // backend reconciles its copy in UpdateAdvertisedExtensions.
-                               .Extensions = BuildAdvertisedExtensions(false, false, false),
+                               .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                                .IsCompatibilityProfile = false},
             .StaticBackendCapability = {.AllowVSOnlyPrograms = false}};
         return rendererInfo;
     }
 
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported) {
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
-            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_multi_draw_indirect,
+            E_GL_ARB_clear_buffer_object, E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_draw_indirect,
+            E_GL_ARB_multi_draw_indirect,
             E_GL_ARB_indirect_parameters, E_GL_EXT_framebuffer_object, E_GL_ARB_depth_texture, E_GL_ARB_buffer_storage,
             E_GL_ARB_texture_storage, E_GL_ARB_texture_storage_multisample, E_GL_ARB_texture_multisample,
             E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access, E_GL_ARB_shader_draw_parameters,
             E_GL_ARB_gpu_shader_int64, E_GL_KHR_debug, E_GL_ARB_gpu_shader5, E_GL_ARB_multi_bind,
             E_GL_ARB_shading_language_420pack, E_GL_ARB_vertex_attrib_binding, E_GL_ARB_shader_image_size,
             E_GL_ARB_explicit_attrib_location,
+            // Core since GL 3.1 and implemented for every version advertised here. The string
+            // matters because applications gate the ENTRY POINTS on it rather than on the
+            // version: a caller that finds the extension missing never resolves
+            // glGetUniformBlockIndex / glUniformBlockBinding, and one that then uses uniform
+            // blocks anyway calls through a null pointer.
+            E_GL_ARB_uniform_buffer_object,
+            // Sampling the stencil aspect through DEPTH_STENCIL_TEXTURE_MODE. Core from 4.3,
+            // so on a 4.0 context the string is the only way to reach it.
+            E_GL_ARB_stencil_texturing,
             // Advertised with GL_NUM_PROGRAM_BINARY_FORMATS = 0, which the
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Vulkan's drawIndirectFirstInstance feature is optional. Direct base-instance calls work
+        // without it, but ARB_base_instance also promises non-zero firstInstance in GPU indirect
+        // commands; the renderer supplies true only when that word is legal and gl_InstanceID can
+        // be rebased to OpenGL's zero-based semantics.
+        if (nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         if (shaderSubgroupSupported && !MG_Config::Features.DisableSubgroup) {
             extensions.push_back(E_GL_KHR_shader_subgroup);
         }
@@ -669,6 +692,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_vulkanCaps = capabilities;
         UpdateDynamicBackendParameters();
         UpdateAdvertisedExtensions();
+        if (MG_State::pGLContext) {
+            MG_State::pGLContext->InvalidateCompileEnv();
+        }
         MutableFormatCapabilities().Clear();
     }
 
@@ -679,9 +705,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // real device timestamp support. ApplyVulkanCapabilitiesForTesting may
         // run without a renderer; no timer query is advertised then. Rebuilding
         // the whole list keeps re-runs idempotent.
+        // The opt-in emulated compute path (SubgroupSupportPolicy.h) carries the
+        // extension by itself on devices with no native subgroup support at all; a
+        // device with native subgroups always advertises - and uses - those.
+        const Bool subgroupSupportAdvertised =
+            m_vulkanCaps.SupportsShaderSubgroup ||
+            ShouldEmulateSubgroups(m_vulkanCaps.SupportsShaderSubgroup);
         m_rendererInfo.RendererGLInfo.Extensions = BuildAdvertisedExtensions(
-            m_vulkanCaps.SupportsShaderSubgroup, pVulkanRenderer && pVulkanRenderer->IsTimerQuerySupported(),
+            subgroupSupportAdvertised, pVulkanRenderer && pVulkanRenderer->IsTimerQuerySupported(),
-            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported());
+            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported(),
+            pVulkanRenderer && pVulkanRenderer->IsNonZeroIndirectBaseInstanceSupported());
     }
 
     void BackendObject_DirectVulkan::UpdateDynamicBackendParameters() {
@@ -915,6 +948,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_dynamicParameters.SubgroupSupportedFeatures =
                 mapSubgroupFeatures(m_vulkanCaps.SubgroupSupportedOperations);
             m_dynamicParameters.SubgroupQuadOperationsInAllStages = m_vulkanCaps.SubgroupQuadOperationsInAllStages;
+        } else if (ShouldEmulateSubgroups(m_vulkanCaps.SupportsShaderSubgroup)) {
+            // MOBILEGL_MAGMA_EMULATE_SUBGROUP on a device with no native subgroups: the
+            // advertised values describe the 32-lane virtual subgroup the compute
+            // lowering implements (SubgroupSupportPolicy.h / EmulateSubgroupsPass).
+            // GL requires the advertisement and the execution to agree, and on this
+            // path the emulation is what executes; only the compute stage is offered.
+            m_dynamicParameters.SubgroupSize = kEmulatedSubgroupSize;
+            m_dynamicParameters.SubgroupSupportedStages = kEmulatedSubgroupStages;
+            m_dynamicParameters.SubgroupSupportedFeatures = kEmulatedSubgroupFeatures;
+            m_dynamicParameters.SubgroupQuadOperationsInAllStages = false;
+            MGLOG_I("DirectVulkan: emulating 32-lane compute subgroups "
+                    "(MOBILEGL_MAGMA_EMULATE_SUBGROUP, no native subgroup support)");
         } else {
             m_dynamicParameters.SubgroupSize = 0;
             m_dynamicParameters.SubgroupSupportedStages = 0;

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.h

@@ -62,8 +62,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // POST screen shows.
 
     // Static identity of the Magma renderer (renderer/backend names, target GL/GLSL
-    // versions, ExtraVendor) with the baseline extension advertisement (no shader
+    // versions, ExtraVendor) with the baseline extension advertisement (no runtime-gated
-    // subgroup, no timer queries). A live backend copies this in its constructor and
+    // capabilities). A live backend copies this in its constructor and
     // reconciles the Extensions in UpdateAdvertisedExtensions once real capabilities
     // exist; callers that need the advertised list for a known capability set must
     // use BuildAdvertisedExtensions instead.
@@ -74,7 +74,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // MOBILEGL_DISABLE_TIMERQUERY escape hatches are applied inside, so callers pass
     // the detected device support (passing an already-gated value is harmless).
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported);
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <GPU Name>, Vulkan <Vulkan Version>, Driver <Driver Version> — the exact
     // string an initialized backend returns from GetBackendAPIVersionString (and that

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp

@@ -69,6 +69,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // slot's ownership unambiguous.
             Uint64 programLifetimeId = 0;
             Uint32 backendStateVersion = 0;
+            // glShaderStorageBlockBinding deliberately does NOT bump the backend state
+            // version, and the pipeline composite is unnamed so the in-place patch in
+            // DirectVulkan::ShaderStorageBlockBinding can never reach its slot - the
+            // mirror replay bumps only the program's block-binding version. Without this
+            // key the composite's slot kept serving the pre-rebind block.binding.
+            Uint32 blockBindingVersion = 0;
             Vector<StorageBlockResource> storageBlocks;
             Vector<BufferVariableResource> bufferVariables;
             GLint computeWorkGroupSize[3] = {1, 1, 1};
@@ -156,18 +162,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             auto& cache = g_programResourceCaches[program.GetExternalIndex()];
             const Uint64 programLifetimeId = program.GetLifetimeId();
             const Uint32 backendStateVersion = program.GetBackendStateVersion();
+            const Uint32 blockBindingVersion = program.GetBlockBindingVersion();
             // The lifetime id must match too: a new program that reuses a deleted
             // program's name and happens to land on the same backendStateVersion (both
             // count from zero) would otherwise be served the dead program's reflection.
             if (cache.programLifetimeId == programLifetimeId &&
                 cache.backendStateVersion == backendStateVersion &&
                 (!cache.storageBlocks.empty() || !cache.bufferVariables.empty())) {
+                if (cache.blockBindingVersion != blockBindingVersion) {
+                    // Only the block bindings moved (glShaderStorageBlockBinding, or the
+                    // pipeline composite's mirror replay - neither touches the backend
+                    // state version): the reflection itself is unchanged, so re-apply the
+                    // overrides by name instead of re-running spirv-reflect. Overrides
+                    // only ever accumulate, so a block without one still holds its
+                    // declared binding.
+                    for (auto& block : cache.storageBlocks) {
+                        const Int rebound = program.GetShaderStorageBlockBindingOverride(block.name);
+                        if (rebound >= 0) block.binding = static_cast<Uint32>(rebound);
+                    }
+                    cache.blockBindingVersion = blockBindingVersion;
+                }
                 return cache;
             }
 
             cache = {};
             cache.programLifetimeId = programLifetimeId;
             cache.backendStateVersion = backendStateVersion;
+            cache.blockBindingVersion = blockBindingVersion;
 
             Vector<SpvReflectShaderModule> modules;
             Vector<Bool> validModules;
@@ -269,14 +290,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 drawBuffer->SyncPersistentMappedRange();
                 const SizeT commandOffset = reinterpret_cast<SizeT>(indirect);
                 if (drawBuffer->MappedData() == nullptr || commandOffset + requiredBytes > drawBuffer->GetSize()) {
-                    MGLOG_E("%s skipped: invalid GL_DRAW_INDIRECT_BUFFER binding or range", label);
+                    MGLOG_E_ONCE("%s skipped: invalid GL_DRAW_INDIRECT_BUFFER binding or range", label);
                     return nullptr;
                 }
                 return drawBuffer->MappedData() + commandOffset;
             }
 
             if (!indirect) {
-                MGLOG_E("%s skipped: indirect pointer is null", label);
+                MGLOG_E_ONCE("%s skipped: indirect pointer is null", label);
                 return nullptr;
             }
 
@@ -398,7 +419,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             stride = sizeof(DrawArraysIndirectCommand);
         }
         if (stride < static_cast<GLsizei>(sizeof(DrawArraysIndirectCommand))) {
-            MGLOG_E("MultiDrawArraysIndirect skipped: stride %d is smaller than command size %zu",
+            MGLOG_E_ONCE("MultiDrawArraysIndirect skipped: stride %d is smaller than command size %zu",
                     stride, sizeof(DrawArraysIndirectCommand));
             return;
         }
@@ -446,20 +467,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             stride = sizeof(DrawArraysIndirectCommand);
         }
         if (stride < static_cast<GLsizei>(sizeof(DrawArraysIndirectCommand))) {
-            MGLOG_E("MultiDrawArraysIndirectCount skipped: stride %d is smaller than command size %zu",
+            MGLOG_E_ONCE("MultiDrawArraysIndirectCount skipped: stride %d is smaller than command size %zu",
                     stride, sizeof(DrawArraysIndirectCommand));
             return;
         }
 
         auto parameterBuffer = MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::Parameter).GetBoundObject();
         if (!parameterBuffer || drawcount < 0 || static_cast<SizeT>(drawcount) + sizeof(Uint32) > parameterBuffer->GetSize()) {
-            MGLOG_E("MultiDrawArraysIndirectCount skipped: invalid GL_PARAMETER_BUFFER binding or range");
+            MGLOG_E_ONCE("MultiDrawArraysIndirectCount skipped: invalid GL_PARAMETER_BUFFER binding or range");
             return;
         }
 
         parameterBuffer->SyncPersistentMappedRange();
         if (parameterBuffer->MappedData() == nullptr) {
-            MGLOG_E("MultiDrawArraysIndirectCount skipped: CPU fallback cannot read parameter buffer");
+            MGLOG_E_ONCE("MultiDrawArraysIndirectCount skipped: CPU fallback cannot read parameter buffer");
             return;
         }
 
@@ -513,7 +534,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const SizeT indexSize = MG_Util::GetGLTypeSize(type);
         if (indexSize == 0) {
-            MGLOG_E("DrawElementsIndirect skipped: unsupported index type 0x%x", type);
+            MGLOG_E_ONCE("DrawElementsIndirect skipped: unsupported index type 0x%x", type);
             return;
         }
 
@@ -1009,7 +1030,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // 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);
+            MGLOG_E_ONCE("MultiDrawElements skipped: unsupported index type 0x%x", type);
             return;
         }
         const Uint32 indexSizeShift = static_cast<Uint32>(std::countr_zero(indexSize));

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.cpp

@@ -205,7 +205,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // commands away. The device is gone on that path anyway - stay silent-safe
         // rather than trade a lost device for a barrier into a closed buffer.
         if (frame.hasCommandBufferRecorded) {
-            MGLOG_E("TransitionToPresent: command buffer already closed; skipping the present barrier");
+            MGLOG_E_ONCE("TransitionToPresent: command buffer already closed; skipping the present barrier");
             return false;
         }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -206,6 +206,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         XXHASH_VERIFY(
             XXH64_update(m_hashState, &payload.primitiveRestartEnable, sizeof(payload.primitiveRestartEnable)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.patchControlPoints, sizeof(payload.patchControlPoints)));
+        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.viewportCount, sizeof(payload.viewportCount)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.polygonMode, sizeof(payload.polygonMode)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.cullMode, sizeof(payload.cullMode)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.frontFace, sizeof(payload.frontFace)));
@@ -259,7 +260,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // is the correct price for a broken pipeline and is bounded by the draw itself being
         // skipped.
         if (pipeline == VK_NULL_HANDLE) {
-            MGLOG_I("PipelineFactory::GetOrCreatePipeline: creation failed for hash=0x%llx "
+            // Unlatched, like the CreatePipeline report it accompanies: a pipeline MobileGL
+            // assembled and the driver refused is a broken invariant, not an expected failure,
+            // so it stays loud for as long as it is reachable. Raised from MGLOG_I once the
+            // Log.h ordering fix made MGLOG_E live in INFO builds.
+            MGLOG_E("PipelineFactory::GetOrCreatePipeline: creation failed for hash=0x%llx "
                     "programHash=0x%llx; not caching the failure",
                     static_cast<unsigned long long>(hash),
                     static_cast<unsigned long long>(payload.programHash));
@@ -402,8 +407,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         tessellation.patchControlPoints = payload.patchControlPoints;
 
         VkPipelineViewportStateCreateInfo vpci{VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO};
-        vpci.viewportCount = 1;
+        // Both counts move together: GL has one scissor rectangle per viewport, and Vulkan
-        vpci.scissorCount = 1;
+        // requires viewportCount == scissorCount whenever both are dynamic
+        // (VUID-VkPipelineViewportStateCreateInfo-scissorCount-04136). The caller has already
+        // clamped this to the device's multiViewport capability.
+        vpci.viewportCount = std::max<Uint32>(payload.viewportCount, 1u);
+        vpci.scissorCount = vpci.viewportCount;
 
         VkPipelineRasterizationStateCreateInfo raster{VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO};
         raster.polygonMode = payload.polygonMode;
@@ -471,9 +480,56 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         blend.attachmentCount = payload.colorAttachmentCount;
         blend.pAttachments = colorAttachments.empty() ? nullptr : colorAttachments.data();
 
+        // A GL program may have a tessellation EVALUATION stage and no CONTROL stage: GL 4.6 core
+        // 11.2.2 gives it a fixed-function pass-through instead. Vulkan has no such stage, and
+        // VUID-VkGraphicsPipelineCreateInfo-pStages-00730 requires both tessellation stages or
+        // neither - so the renderer synthesizes the pass-through GL describes and hands it in
+        // here (see ProgramFactory::GetOrCreatePassthroughTessControlStage).
+        //
+        // The refusal below is what keeps the half-tessellated shape away from the driver when
+        // there is no synthesized stage to add - because Mali does not reject it, it dereferences
+        // null INSIDE vkCreateGraphicsPipelines and takes the process down (SIGSEGV, fault addr
+        // 0x34, on Mali-G715/r54p2 and Mali-G925/r49p1 alike; Adreno and lavapipe merely render
+        // wrong). Returning VK_NULL_HANDLE routes this through the same path a driver rejection
+        // takes: the draw is skipped, nothing is memoised, and the process survives.
+        const Vector<VkPipelineShaderStageCreateInfo>* effectiveStages = payload.stages;
+        Vector<VkPipelineShaderStageCreateInfo> stagesWithPassthrough;
+        if (payload.passthroughTessControlStage.module != VK_NULL_HANDLE) {
+            stagesWithPassthrough = *payload.stages;
+            stagesWithPassthrough.push_back(payload.passthroughTessControlStage);
+            effectiveStages = &stagesWithPassthrough;
+        }
+        {
+            VkShaderStageFlags stagesPresent = 0;
+            for (const auto& stageInfo : *effectiveStages) {
+                stagesPresent |= stageInfo.stage;
+            }
+            const Bool hasTessControl = (stagesPresent & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT) != 0;
+            const Bool hasTessEval = (stagesPresent & VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT) != 0;
+            if (hasTessControl != hasTessEval) {
+                // Latched, and the latch is the point: a failed creation is deliberately never
+                // memoised (see GetOrCreatePipeline), so a program in this state re-enters here
+                // once per draw, every frame - and a refusal diagnostic that repeats per draw is
+                // noise, not a diagnostic. One line names the program; the draws it explains are
+                // all the same draw.
+                static Bool s_warnedHalfTessellatedPipeline = false;
+                if (!s_warnedHalfTessellatedPipeline) {
+                    s_warnedHalfTessellatedPipeline = true;
+                    MGLOG_E_ONCE("PipelineFactory::CreatePipeline: refusing a pipeline with %s tessellation stage and "
+                            "no %s stage (VUID-VkGraphicsPipelineCreateInfo-pStages-00730). programHash=0x%llx "
+                            "patchControlPoints=%u. Its draws are skipped; logged once.",
+                            hasTessEval ? "an evaluation" : "a control",
+                            hasTessEval ? "control" : "evaluation",
+                            static_cast<unsigned long long>(payload.programHash),
+                            payload.patchControlPoints);
+                }
+                return VK_NULL_HANDLE;
+            }
+        }
+
         VkGraphicsPipelineCreateInfo gpi{VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO};
-        gpi.stageCount = static_cast<Uint32>(payload.stages->size());
+        gpi.stageCount = static_cast<Uint32>(effectiveStages->size());
-        gpi.pStages = payload.stages->data();
+        gpi.pStages = effectiveStages->data();
         gpi.pVertexInputState = payload.vertexInputState;
         gpi.pInputAssemblyState = &ia;
         gpi.pTessellationState =
@@ -490,6 +546,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkPipeline pipeline = VK_NULL_HANDLE;
         const VkResult result = vkCreateGraphicsPipelines(m_device, m_pipelineCache, 1, &gpi, nullptr, &pipeline);
+        // Loud, at MGLOG_F, and deliberately NOT latched. vkCreateGraphicsPipelines refusing a
+        // pipeline MobileGL assembled is a should-never-happen state, and the driver's own
+        // answer is VK_ERROR_UNKNOWN - no information at all - so this dump is the entire
+        // diagnosis. It is not an expected failure mode, so the one-shot rule that quiets W/E
+        // does not apply: while this is reachable it should keep saying so on every draw.
+        // GetOrCreatePipeline deliberately does not cache the failure, which is what makes that
+        // repetition happen; if the repetition ever needs to stop, fix the pipeline, not the log.
         if (result != VK_SUCCESS) {
             MGLOG_F("PipelineFactory::CreatePipeline failed: result=%s (%d) programHash=0x%llx vertexInputHash=0x%llx stageCount=%u topology=%s(%d) colorAttachmentCount=%u samples=%s(%d) subpass=%u",
                     VkResultToString(result),
@@ -522,8 +585,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     payload.vertexInputState->vertexAttributeDescriptionCount);
             // The driver's own answer is VK_ERROR_UNKNOWN, i.e. no information at all, so the only
             // way to work out WHICH shader it choked on (the open sampler-array-in-struct
-            // investigation) is to name the modules. MGLOG_I, not _D/_E: this must survive in the
+            // investigation) is to name the modules. MGLOG_I, not _D: this is part of a
-            // INFO-level builds that CTS actually runs against.
+            // should-never-happen report and must survive in the INFO-level builds that CTS
+            // actually runs against, alongside the MGLOG_F lines above.
             if (payload.stageSpirvDigests) {
                 for (SizeT i = 0; i < payload.stageSpirvDigests->size(); ++i) {
                     const auto& digest = (*payload.stageSpirvDigests)[i];

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -42,6 +42,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool primitiveRestartEnable = false;
             // GL_PATCH_VERTICES; only read for a PATCH_LIST topology.
             Uint32 patchControlPoints = 3;
+            // How many of ARB_viewport_array's viewports this pipeline rasterizes into. 1 for
+            // every program that never assigns gl_ViewportIndex, which is all of them outside the
+            // conformance suite - the wide shape costs a longer vkCmdSetViewport/Scissor per state
+            // change and can cost hardware fast paths, so it is opt-in per program. Baked into the
+            // pipeline (viewportCount is not dynamic without VK_EXT_extended_dynamic_state) and
+            // therefore hashed; the DYNAMIC viewport/scissor arrays the draw pushes must have
+            // exactly this many elements (VUID-vkCmdDraw-viewportCount-03417/-03418).
+            Uint32 viewportCount = 1;
             VkPolygonMode polygonMode = VK_POLYGON_MODE_FILL;
             VkCullModeFlags cullMode = VK_CULL_MODE_BACK_BIT;
             VkFrontFace frontFace = VK_FRONT_FACE_CLOCKWISE;
@@ -71,6 +79,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool fragmentReplacesDepth = false;
             Array<VkPipelineColorBlendAttachmentState, kMaxColorAttachments> colorBlendAttachments{};
             const Vector<VkPipelineShaderStageCreateInfo>* stages = nullptr;
+            // The tessellation control stage this renderer synthesized for a program that has
+            // an evaluation stage and none of its own (GL 4.6 core 11.2.2 gives such a program a
+            // fixed-function pass-through; Vulkan has no such thing and
+            // VUID-VkGraphicsPipelineCreateInfo-pStages-00730 forbids the half-tessellated
+            // pipeline outright). Appended to `stages` at creation. A null module means the
+            // renderer could not build one, and CreatePipeline refuses the pipeline - the same
+            // refusal it applies when `stages` itself is half-tessellated.
+            //
+            // NOT hashed: it is a pure function of the program and of patchControlPoints, both
+            // of which ComputeHash already mixes in.
+            VkPipelineShaderStageCreateInfo passthroughTessControlStage{};
             const VkPipelineVertexInputStateCreateInfo* vertexInputState = nullptr;
             // Diagnostic only; may be null. Read solely from the pipeline-creation failure path.
             const Vector<ShaderStageSpirvDigest>* stageSpirvDigests = nullptr;

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -33,6 +33,32 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         using SpvcSession = MG_Util::ShaderTranspiler::SpvcSession;
         using SessionUsageBit = MG_Util::ShaderTranspiler::SessionUsageBit;
 
+        // Local size of a compute module, read from OpExecutionMode LocalSize; all-zero
+        // when absent. The compile chain pins SPIR-V 1.3, where a literal local size
+        // always reaches the module as this execution mode (LocalSizeId does not exist
+        // yet).
+        struct ComputeLocalSize {
+            Uint32 x = 0;
+            Uint32 y = 0;
+            Uint32 z = 0;
+            Uint64 Total() const { return static_cast<Uint64>(x) * y * z; }
+        };
+        ComputeLocalSize TryGetComputeLocalSize(const Vector<Uint>& spirv) {
+            constexpr SizeT kHeaderWords = 5;
+            constexpr Uint32 kOpExecutionMode = 16;
+            constexpr Uint32 kModeLocalSize = 17;
+            for (SizeT offset = kHeaderWords; offset < spirv.size();) {
+                const Uint32 wordCount = spirv[offset] >> 16u;
+                const Uint32 opcode = spirv[offset] & 0xffffu;
+                if (wordCount == 0 || offset + wordCount > spirv.size()) break;
+                if (opcode == kOpExecutionMode && wordCount >= 6 && spirv[offset + 2] == kModeLocalSize) {
+                    return {spirv[offset + 3], spirv[offset + 4], spirv[offset + 5]};
+                }
+                offset += wordCount;
+            }
+            return {};
+        }
+
         struct DescriptorKey {
             ProgramFactory::DescriptorBindingKind kind = ProgramFactory::DescriptorBindingKind::None;
             String name;
@@ -376,12 +402,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             spv_diagnostic diagnostic = nullptr;
             const spv_result_t result = spvValidateWithOptions(context, options, &binary, &diagnostic);
             if (result != SPV_SUCCESS) {
-                // MGLOG_I, not E: at the INFO compile level of the CI/test lanes that arm
+                // MGLOG_E, unlatched: reaching here already requires the validation switch to
-                // the validation switch, MGLOG_E is compiled out (Log.h orders
+                // be armed, which bounds the volume, and each VUID names a different defect.
-                // DEBUG < WARN < ERROR < INFO) and the VUID would never reach a log. The
+                // (Parked at MGLOG_I until the Log.h level ordering was fixed, when E was
-                // latch is what a test harness asserts on.
+                // compiled out of every INFO build.) The latch is what a test harness asserts on.
                 MG_Util::ShaderTranspiler::ShaderCompiler::NoteSpirvValidationFailure();
-                MGLOG_I(
+                MGLOG_E(
                     "ProgramFactory::ValidateTransformedSpirv: validation failed for stage=%d program=%u result=%d index=%zu msg=%s",
                     static_cast<Int>(shaderStage),
                     programExternalIndex,
@@ -1266,7 +1292,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     for (SizeT i = 1; i < group.offsets.size(); ++i) {
                         if (group.elementBytes == 0 ||
                             group.offsets[i] != group.offsets[i - 1] + group.elementBytes) {
-                            MGLOG_I("XfbCaptureDecoratePass: block member %u of type %%%u is captured with a "
+                            MGLOG_D("XfbCaptureDecoratePass: block member %u of type %%%u is captured with a "
                                     "non-contiguous element set; the capture layout will differ from GL's",
                                     key.second, key.first);
                             break;
@@ -1721,6 +1747,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 return ProgramFactory::DescriptorBindingKind::CombinedImageSampler;
             case SPV_REFLECT_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
                 return ProgramFactory::DescriptorBindingKind::UniformTexelBuffer;
+            case SPV_REFLECT_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+                return ProgramFactory::DescriptorBindingKind::StorageTexelBuffer;
             case SPV_REFLECT_DESCRIPTOR_TYPE_STORAGE_BUFFER:
                 return ProgramFactory::DescriptorBindingKind::StorageBuffer;
             case SPV_REFLECT_DESCRIPTOR_TYPE_STORAGE_IMAGE:
@@ -1750,6 +1778,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
             if (kind == ProgramFactory::DescriptorBindingKind::CombinedImageSampler ||
                 kind == ProgramFactory::DescriptorBindingKind::UniformTexelBuffer ||
+                kind == ProgramFactory::DescriptorBindingKind::StorageTexelBuffer ||
                 kind == ProgramFactory::DescriptorBindingKind::StorageImage) {
                 const auto arraySuffix = name.find("[0]");
                 if (arraySuffix != String::npos) {
@@ -1839,8 +1868,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     // UniformManager::BindProgramUniformBuffers: UBO instance arrays
                     // (uniform Block {...} b[N];), storage-block instance arrays, image uniform
                     // arrays, and combined-image-sampler arrays (uniform sampler2D s[N];).
-                    // Anything else - a uniform TEXEL buffer array is the one remaining kind -
+                    // Anything else - the two TEXEL buffer kinds are what remain, samplerBuffer[N]
-                    // must fail program creation cleanly rather than continue with corrupt state.
+                    // and imageBuffer[N] - must fail program creation cleanly rather than continue
+                    // with corrupt state. Their per-draw path writes pTexelBufferView as the
+                    // address of a vector element sized for one descriptor per binding, so an
+                    // array would not merely be unresolved, it would dangle.
                     //
                     // Getting listed here is not cosmetic: a kind that is rejected leaves
                     // GetOrCreateProgram's MOBILEGL_ASSERT(remapOk) as the only complaint, and
@@ -1849,16 +1881,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     // unification and the set->0 normalisation this function exists to do. A
                     // program with an image array plus any second descriptor got aliased
                     // bindings out of that, and a DEBUG build trapped on the same program.
-                    // Which is also why the message below is MGLOG_I: MGLOG_E is compiled out
+                    // The refusal below is MGLOG_E and per-program-compile, so it reports every
-                    // of an INFO build, so a refusal that only said MGLOG_E said nothing at all
+                    // program it declines. It spent time at MGLOG_I because the old level
-                    // in the builds that ship.
+                    // ordering compiled E out of the builds that ship.
                     const Bool arraySupportedForKind =
                         kind == ProgramFactory::DescriptorBindingKind::UniformBufferDynamic ||
                         kind == ProgramFactory::DescriptorBindingKind::StorageBuffer ||
                         kind == ProgramFactory::DescriptorBindingKind::StorageImage ||
                         kind == ProgramFactory::DescriptorBindingKind::CombinedImageSampler;
                     if (binding->count != 1 && !arraySupportedForKind) {
-                        MGLOG_I("ProgramFactory: descriptor arrays are unsupported for this descriptor "
+                        MGLOG_E("ProgramFactory: descriptor arrays are unsupported for this descriptor "
                                 "kind (name='%s' count=%u type=%d)",
                                 binding->name ? binding->name : "<null>", binding->count,
                                 static_cast<Int>(binding->descriptor_type));
@@ -1991,6 +2023,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ReflectedDeclaresInputBuiltin(reflectModule, SpvBuiltInBaseVertex);
     }
 
+    // gl_ViewportIndex on the last pre-rasterization stage. glslang emits it natively for Vulkan
+    // (BuiltIn ViewportIndex plus OpCapability MultiViewport), and nothing in the SpirvPasses
+    // chain touches it, so a plain reflection of the declared output builtins is the whole test.
+    Bool ProgramFactory::ReflectedWritesViewportIndexBuiltin(const SpvReflectShaderModule& reflectModule) {
+        return ReflectedDeclaresOutputBuiltin(reflectModule, SpvBuiltInViewportIndex);
+    }
+
+    Bool ProgramFactory::ReflectedDeclaresOutputBuiltin(const SpvReflectShaderModule& reflectModule,
+                                                       SpvBuiltIn builtin) {
+        for (Uint32 entryIndex = 0; entryIndex < reflectModule.entry_point_count; ++entryIndex) {
+            const SpvReflectEntryPoint& entryPoint = reflectModule.entry_points[entryIndex];
+            for (Uint32 variableIndex = 0; variableIndex < entryPoint.output_variable_count; ++variableIndex) {
+                const SpvReflectInterfaceVariable* variable = entryPoint.output_variables[variableIndex];
+                if (variable != nullptr &&
+                    (variable->decoration_flags & SPV_REFLECT_DECORATION_BUILT_IN) != 0 &&
+                    variable->built_in == builtin) {
+                    return true;
+                }
+            }
+        }
+        return false;
+    }
+
     Bool ProgramFactory::ReflectedDeclaresInputBuiltin(const SpvReflectShaderModule& reflectModule,
                                                        SpvBuiltIn builtin) {
         for (Uint32 entryIndex = 0; entryIndex < reflectModule.entry_point_count; ++entryIndex) {
@@ -2333,6 +2388,46 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    // Which pre-rasterization stage assigns gl_ViewportIndex is not fixed: GL 4.1 allows only the
+    // geometry stage, ARB_shader_viewport_layer_array/GL 4.6 also the vertex and tessellation
+    // evaluation stages. Rather than guess which one is last, every non-fragment, non-compute
+    // module is asked - one writer anywhere means this program's draws need a multi-viewport
+    // pipeline, and a false positive costs only a wider viewportCount.
+    void ProgramFactory::ReflectViewportIndexUsage(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                                   const Vector<Vector<Uint>>& spirv,
+                                                   VkProgramObject& entry) const {
+        entry.writesViewportIndexBuiltin = false;
+
+        for (SizeT moduleIndex = 0; moduleIndex < shaders.size() && moduleIndex < spirv.size(); ++moduleIndex) {
+            if (!shaders[moduleIndex]) continue;
+            const ShaderStage stage = shaders[moduleIndex]->GetShaderStage();
+            if (stage == ShaderStage::Fragment || stage == ShaderStage::Compute) continue;
+
+            const auto& module = spirv[moduleIndex];
+            if (module.empty()) continue;
+
+            SpvReflectShaderModule reflectModule{};
+            const SpvReflectResult createResult =
+                spvReflectCreateShaderModule(module.size() * sizeof(Uint), module.data(), &reflectModule);
+            if (createResult != SPV_REFLECT_RESULT_SUCCESS) {
+                // Fail toward the wide pipeline. Missing a real gl_ViewportIndex writer would
+                // silently collapse every viewport onto 0 (the exact bug this reflection exists
+                // to fix); over-declaring costs one extra viewport slot on a program that never
+                // uses it.
+                MGLOG_E_ONCE("ProgramFactory::ReflectViewportIndexUsage: reflection failed (result=%d); assuming the "
+                             "program writes gl_ViewportIndex",
+                             static_cast<Int>(createResult));
+                entry.writesViewportIndexBuiltin = true;
+                continue;
+            }
+
+            if (ReflectedWritesViewportIndexBuiltin(reflectModule)) {
+                entry.writesViewportIndexBuiltin = true;
+            }
+            spvReflectDestroyShaderModule(&reflectModule);
+        }
+    }
+
     void ProgramFactory::ReflectFragmentOutputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                                                 const Vector<Vector<Uint>>& spirv,
                                                 VkProgramObject& entry) const {
@@ -2462,7 +2557,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // inert; a device whose binding cap is smaller than a shader's array is not a
             // configuration MobileGL can serve at all. Needs a >maxBindings-element array to
             // reach (256 on desktop, ~16 on mobile).
-            MGLOG_I("ProgramFactory::ReflectLayout: %s array '%s' at binding %u has %u elements, past the %u "
+            MGLOG_D("ProgramFactory::ReflectLayout: %s array '%s' at binding %u has %u elements, past the %u "
                     "this device can describe - declining the program",
                     kindLabel, uniformName.c_str(), binding, count, maxBindings);
             outDeclined = true;
@@ -2470,7 +2565,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         if (baseLocation < 0 ||
             !program.UniformLocationsAliasSameUniform(baseLocation, baseLocation + static_cast<Int>(count - 1u))) {
-            MGLOG_I("ProgramFactory::ReflectLayout: %s array '%s' at binding %u spans %u descriptors but the "
+            MGLOG_D("ProgramFactory::ReflectLayout: %s array '%s' at binding %u spans %u descriptors but the "
                     "reflection reserved fewer uniform locations for it (base=%d) - a multi-dimensional array "
                     "is the usual cause, and MobileGL declines it rather than resolve elements onto a "
                     "neighbouring uniform",
@@ -2661,6 +2756,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const auto descriptorKind = ReflectDescriptorTypeToBindingKind(sampler->descriptor_type);
                 if (descriptorKind != DescriptorBindingKind::CombinedImageSampler &&
                     descriptorKind != DescriptorBindingKind::UniformTexelBuffer &&
+                    descriptorKind != DescriptorBindingKind::StorageTexelBuffer &&
                     descriptorKind != DescriptorBindingKind::StorageImage &&
                     descriptorKind != DescriptorBindingKind::StorageBuffer) {
                     continue;
@@ -2706,7 +2802,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     // a Uint16 on the way, where 65536 would silently become 0.
                     const Uint32 storageArrayCount = std::max<Uint32>(1u, sampler->count);
                     if (storageArrayCount > m_maxBindings) {
-                        MGLOG_I("ProgramFactory::ReflectLayout: storage block array '%s' at binding %u has %u "
+                        MGLOG_D("ProgramFactory::ReflectLayout: storage block array '%s' at binding %u has %u "
                                 "elements, past the %u this device can describe - declining the program",
                                 uniformName.c_str(), binding, storageArrayCount, m_maxBindings);
                         entry.declinedDescriptors = true;
@@ -2729,7 +2825,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     // so at a level that survives a release build, because dropping the binding
                     // leaves the shader reading a descriptor the layout never declared.
                     if (sampler->count > 1) {
-                        MGLOG_I("ProgramFactory::ReflectLayout: declining '%s' at binding %u - a %u-element "
+                        MGLOG_E("ProgramFactory::ReflectLayout: declining '%s' at binding %u - a %u-element "
                                 "descriptor array with no frontend uniform location (a multi-dimensional array "
                                 "of samplers or images is the known cause)",
                                 uniformName.c_str(), binding, sampler->count);
@@ -2790,6 +2886,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     }
                 }
 
+                if (descriptorKind == DescriptorBindingKind::StorageTexelBuffer) {
+                    // Only the declared format is recorded, and only so the per-draw resolve can
+                    // prefer it over the one glBindImageTexture named. Everything the StorageImage
+                    // branch above does about ARRAYS is deliberately absent: an imageBuffer array
+                    // is refused outright by the array gate in RemapDescriptorBindingsForVulkan,
+                    // exactly as a samplerBuffer array is, so bindingDescriptorCounts stays at the
+                    // default 1 and the descriptor write below may take the address of a vector
+                    // element without reserving room for extra elements.
+                    const VkFormat reflectedFormat =
+                        ConvertSpirvImageFormatToVkFormat(sampler->image.image_format);
+                    VkFormat& existingFormat = entry.storageImageFormatByBinding[binding];
+                    MOBILEGL_ASSERT(existingFormat == VK_FORMAT_UNDEFINED ||
+                                        reflectedFormat == VK_FORMAT_UNDEFINED ||
+                                        existingFormat == reflectedFormat,
+                                    "ProgramFactory::ReflectLayout: storage texel buffer binding %u ('%s') "
+                                    "has conflicting reflected formats (%d vs %d)",
+                                    binding, uniformName.c_str(), static_cast<Int>(existingFormat),
+                                    static_cast<Int>(reflectedFormat));
+                    if (existingFormat == VK_FORMAT_UNDEFINED) {
+                        existingFormat = reflectedFormat;
+                    }
+                }
+
                 const TextureTarget target = UniformTypeToTextureTarget(uniformType);
                 MOBILEGL_ASSERT(target != TextureTarget::Unknown,
                                 "ProgramFactory::ReflectLayout: failed to resolve texture target for '%s'",
@@ -2867,6 +2986,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 entry.dynamicBindings.push_back(binding);
             } else if (kind == DescriptorBindingKind::UniformTexelBuffer) {
                 layoutBinding.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER;
+            } else if (kind == DescriptorBindingKind::StorageTexelBuffer) {
+                layoutBinding.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
             } else if (kind == DescriptorBindingKind::StorageBuffer) {
                 layoutBinding.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
             } else if (kind == DescriptorBindingKind::StorageImage) {
@@ -2878,8 +2999,73 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             bindings.push_back(layoutBinding);
         }
 
+        // UPDATE_AFTER_BIND is strictly an optional per-layout acceleration. The GL
+        // descriptor model still resolves every sampler uniform element independently
+        // (including its texture-unit sampler-object override); selecting this path
+        // changes neither that resolution nor the set versioning in UniformManager.
+        // A conservative count keeps a layout on ordinary descriptors whenever any
+        // relevant update-after-bind limit is not large enough, rather than asking a
+        // driver to reject it during vkCreateDescriptorSetLayout.
+        Uint32 updateAfterBindSamplers = 0;
+        Uint32 updateAfterBindUniformBuffers = 0;
+        Uint32 updateAfterBindStorageBuffers = 0;
+        Uint32 updateAfterBindSampledImages = 0;
+        Uint32 updateAfterBindStorageImages = 0;
+        for (Uint32 binding = 0; binding < m_maxBindings; ++binding) {
+            const Uint32 count = entry.bindingDescriptorCounts[binding];
+            switch (entry.bindingKinds[binding]) {
+                case DescriptorBindingKind::UniformBufferDynamic:
+                    updateAfterBindUniformBuffers += count;
+                    break;
+                case DescriptorBindingKind::CombinedImageSampler:
+                    updateAfterBindSamplers += count;
+                    updateAfterBindSampledImages += count;
+                    break;
+                case DescriptorBindingKind::UniformTexelBuffer:
+                    updateAfterBindSampledImages += count;
+                    break;
+                case DescriptorBindingKind::StorageBuffer:
+                case DescriptorBindingKind::StorageTexelBuffer:
+                    updateAfterBindStorageBuffers += count;
+                    break;
+                case DescriptorBindingKind::StorageImage:
+                    updateAfterBindStorageImages += count;
+                    break;
+                case DescriptorBindingKind::None:
+                    break;
+            }
+        }
+        const Uint32 updateAfterBindResources = updateAfterBindUniformBuffers + updateAfterBindStorageBuffers +
+                                                updateAfterBindSampledImages + updateAfterBindStorageImages;
+        const auto& uab = m_updateAfterBindLimits;
+        entry.usesUpdateAfterBind =
+            uab.enabled && updateAfterBindSamplers <= uab.maxPerStageSamplers &&
+            updateAfterBindUniformBuffers <= uab.maxPerStageUniformBuffers &&
+            updateAfterBindStorageBuffers <= uab.maxPerStageStorageBuffers &&
+            updateAfterBindSampledImages <= uab.maxPerStageSampledImages &&
+            updateAfterBindStorageImages <= uab.maxPerStageStorageImages &&
+            updateAfterBindResources <= uab.maxPerStageResources &&
+            updateAfterBindSamplers <= uab.maxSetSamplers &&
+            updateAfterBindUniformBuffers <= uab.maxSetUniformBuffers &&
+            updateAfterBindUniformBuffers <= uab.maxSetUniformBuffersDynamic &&
+            updateAfterBindStorageBuffers <= uab.maxSetStorageBuffers &&
+            updateAfterBindStorageBuffers <= uab.maxSetStorageBuffersDynamic &&
+            updateAfterBindSampledImages <= uab.maxSetSampledImages &&
+            updateAfterBindStorageImages <= uab.maxSetStorageImages;
+
+        Vector<VkDescriptorBindingFlags> bindingFlags;
+        VkDescriptorSetLayoutBindingFlagsCreateInfo bindingFlagsInfo{};
+        if (entry.usesUpdateAfterBind) {
+            bindingFlags.assign(bindings.size(), VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT);
+            bindingFlagsInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO;
+            bindingFlagsInfo.bindingCount = static_cast<Uint32>(bindingFlags.size());
+            bindingFlagsInfo.pBindingFlags = bindingFlags.data();
+        }
+
         VkDescriptorSetLayoutCreateInfo setLayoutInfo{};
         setLayoutInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
+        setLayoutInfo.flags = entry.usesUpdateAfterBind ? VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT : 0;
+        setLayoutInfo.pNext = entry.usesUpdateAfterBind ? &bindingFlagsInfo : nullptr;
         setLayoutInfo.bindingCount = static_cast<Uint32>(bindings.size());
         setLayoutInfo.pBindings = bindings.data();
         VK_VERIFY(vkCreateDescriptorSetLayout(m_device, &setLayoutInfo, nullptr, &entry.descriptorSetLayout),
@@ -2959,6 +3145,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto& shaders = program.GetAttachedShaders();
         auto& spirv = program.GetGeneratedSpirv();
         Vector<Vector<Uint>> moduleSpirvs(spirv.size());
+        const Bool enableSpirvValidation = program.GetSpirvValidationEnabled();
+        if (enableSpirvValidation) {
+            MG_Util::ShaderTranspiler::ShaderCompiler::PrepareSpirvValidation();
+        }
 
         const ShaderStage fixupStage = PickClipFixupStage(shaders);
 
@@ -2999,12 +3189,81 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
             }
 
+            // GL_KHR_shader_subgroup handling (SubgroupSupportPolicy.h). Native subgroup
+            // operations execute natively; module repairs keep the GL contract intact
+            // around them. The opt-in emulation path replaces them only on devices with no
+            // subgroup support at all (MOBILEGL_MAGMA_EMULATE_SUBGROUP).
+            if (shaders[i] && shaders[i]->GetShaderStage() == ShaderStage::Compute) {
+                // Program 203 broadcasts the first reduction through
+                // prefixSumCache[0], then lets the second reduction overwrite that
+                // scratch without first rendezvousing all readers. Patch that exact
+                // fingerprint before either native or emulated subgroup lowering.
+                if (m_subgroupPolicy.fixIterationRPBarrier) {
+                    Vector<Uint> patchedSpirv;
+                    if (MG_Util::ShaderTranspiler::ShaderCompiler::FixIterationRPBarrierForVulkan(
+                            moduleSpirvs[i], patchedSpirv, enableSpirvValidation)) {
+                        moduleSpirvs[i] = std::move(patchedSpirv);
+                    } else {
+                        MGLOG_E("ProgramFactory: iterationRP barrier patch failed for program %u; "
+                                "Program 203 keeps its shared-scratch race",
+                                program.GetExternalIndex());
+                    }
+                }
+                if (m_subgroupPolicy.emulateSubgroups) {
+                    Vector<Uint> emulatedSpirv;
+                    if (MG_Util::ShaderTranspiler::ShaderCompiler::EmulateSubgroupsForVulkan(
+                            moduleSpirvs[i], emulatedSpirv,
+                            m_subgroupPolicy.maxComputeSharedMemoryBytes, enableSpirvValidation)) {
+                        moduleSpirvs[i] = std::move(emulatedSpirv);
+                    } else {
+                        MGLOG_E("ProgramFactory: subgroup emulation failed for program %u; the "
+                                "module keeps subgroup operations the device cannot execute",
+                                program.GetExternalIndex());
+                    }
+                } else {
+                    // iterationRP under-declares its cross-subgroup scratch
+                    // (prefixSumCache[32] for 512 invocations); on a sub-16-lane device
+                    // grow that one fingerprinted array to what the topology needs.
+                    if (m_subgroupPolicy.fixIterationRPSubgroupScratch) {
+                        Vector<Uint> patchedSpirv;
+                        if (MG_Util::ShaderTranspiler::ShaderCompiler::FixIterationRPSubgroupScratchForVulkan(
+                                moduleSpirvs[i], patchedSpirv, m_subgroupPolicy.nativeSubgroupSize,
+                                m_subgroupPolicy.maxComputeSharedMemoryBytes,
+                                enableSpirvValidation)) {
+                            moduleSpirvs[i] = std::move(patchedSpirv);
+                        } else {
+                            MGLOG_E("ProgramFactory: iterationRP subgroup scratch patch failed for "
+                                    "program %u; the pack's declared array sizes stay in effect",
+                                    program.GetExternalIndex());
+                        }
+                    }
+                    // gl_NumSubgroups must agree with the gl_SubgroupID range GL promises;
+                    // derive it from the workgroup dimensions and gl_SubgroupSize instead of
+                    // trusting a driver builtin that can disagree with the topology the same
+                    // dispatch emits (Adreno reports 1 while emitting IDs 0..7 for a
+                    // 512-invocation, 64-wide workgroup). The ceil() partition this derives
+                    // is pinned by REQUIRE_FULL_SUBGROUPS at pipeline creation whenever the
+                    // workgroup shape makes that flag legal (see the stage setup below).
+                    if (m_subgroupPolicy.deriveNumSubgroups) {
+                        Vector<Uint> derivedNumSubgroupsSpirv;
+                        if (MG_Util::ShaderTranspiler::ShaderCompiler::DeriveNumSubgroupsForVulkan(
+                                moduleSpirvs[i], derivedNumSubgroupsSpirv, enableSpirvValidation)) {
+                            moduleSpirvs[i] = std::move(derivedNumSubgroupsSpirv);
+                        } else {
+                            MGLOG_E("ProgramFactory: failed to derive gl_NumSubgroups for program %u; "
+                                    "compute shaders may observe a driver-inconsistent subgroup count",
+                                    program.GetExternalIndex());
+                        }
+                    }
+                }
+            }
+
             // Vulkan's SPIR-V environment has no rectangle image dimension, so a
             // GL_TEXTURE_RECTANGLE lookup has to become the 2D one the texture is really
             // stored as - which addresses [0,1] where the application addressed texels.
             {
                 Vector<Uint> rectLoweredSpirv;
-                if (MG_Util::ShaderTranspiler::ShaderCompiler::LowerRectImages(moduleSpirvs[i], rectLoweredSpirv) &&
+                if (MG_Util::ShaderTranspiler::ShaderCompiler::LowerRectImages(moduleSpirvs[i], rectLoweredSpirv, enableSpirvValidation) &&
                     !rectLoweredSpirv.empty()) {
                     moduleSpirvs[i] = Move(rectLoweredSpirv);
                 }
@@ -3017,7 +3276,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             {
                 Vector<Uint> invariantSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::DecoratePositionInvariantForVulkan(
-                        moduleSpirvs[i], invariantSpirv)) {
+                        moduleSpirvs[i], invariantSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(invariantSpirv);
                 } else {
                     // The pass round-trips through SPIRV-Tools IR, so an unparseable module
@@ -3041,7 +3300,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 m_shaderDrawParametersEnabled) {
                 Vector<Uint> rebasedSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::RebaseInstanceIndexForVulkan(moduleSpirvs[i],
-                                                                                            rebasedSpirv)) {
+                                                                                            rebasedSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(rebasedSpirv);
                 } else {
                     MGLOG_E("ProgramFactory: failed to rebase gl_InstanceID for program %u; "
@@ -3059,7 +3318,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 (flags & CompileOptionBit::ZeroBaseVertex)) {
                 Vector<Uint> zeroedSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::ZeroBaseVertexForVulkan(moduleSpirvs[i],
-                                                                                       zeroedSpirv)) {
+                                                                                       zeroedSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(zeroedSpirv);
                 } else {
                     // Failing open keeps the native builtin, which is the pre-fix behavior:
@@ -3082,7 +3341,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (shaders[i] && shaders[i]->GetShaderStage() == ShaderStage::Vertex) {
                 Vector<Uint> packedSpirv;
                 const Bool packOk = MG_Util::ShaderTranspiler::ShaderCompiler::PackDoubleVertexInputsForVulkan(
-                    moduleSpirvs[i], packedSpirv);
+                    moduleSpirvs[i], packedSpirv, enableSpirvValidation);
                 MOBILEGL_ASSERT(packOk,
                                 "ProgramFactory: 64-bit vertex input packing failed for program %u; the "
                                 "vertex-input format and the shader input type now disagree",
@@ -3106,7 +3365,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (m_unformattedFloatStorageImagesEnabled) {
                 Vector<Uint> unformattedSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(
-                        moduleSpirvs[i], unformattedSpirv)) {
+                        moduleSpirvs[i], unformattedSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(unformattedSpirv);
                 } else {
                     MGLOG_E("ProgramFactory: failed to make float storage images unformatted for program %u",
@@ -3127,7 +3386,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 #else
             // Final module the driver receives; also checked in the INFO-level CI/test
             // lanes, where the DEBUG gate above is compiled out.
-            if (MG_Util::ShaderTranspiler::ShaderCompiler::SpirvValidationEnabled()) {
+            if (enableSpirvValidation) {
                 ValidateTransformedSpirv(moduleSpv, shaders[i]->GetShaderStage(), program.GetExternalIndex());
             }
 #endif
@@ -3144,6 +3403,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             stage.stage = ToVkStage(shaderStage);
             stage.module = module;
             stage.pName = "main";
+            // Pin the full-subgroup launch the derived gl_NumSubgroups assumes. Legal
+            // exactly when the computeFullSubgroups feature is enabled and local_size_x is
+            // a multiple of the subgroup size (VUID-VkPipelineShaderStageCreateInfo-
+            // flags-02759/-02785), and only worth requesting while the resulting subgroup
+            // count fits the device's maxComputeWorkgroupSubgroups (lavapipe caps it at
+            // 32, below a 512-invocation dispatch's 64). With the bit set, "Full
+            // Subgroups" guarantees every subgroup launches with all invocations active,
+            // making the subgroup count exactly invocations / size. Shapes the flag
+            // cannot cover (e.g. 32x16 on a 64-wide device) fall back to the driver's
+            // own - spec-encouraged - tight partitioning, which the DriverPost witness
+            // verifies per device.
+            if (shaderStage == ShaderStage::Compute && m_subgroupPolicy.requireFullSubgroups &&
+                !m_subgroupPolicy.emulateSubgroups && m_subgroupPolicy.nativeSubgroupSize != 0) {
+                const ComputeLocalSize localSize = TryGetComputeLocalSize(moduleSpv);
+                const Uint64 fullSubgroupCount =
+                    localSize.Total() / m_subgroupPolicy.nativeSubgroupSize;
+                if (localSize.x != 0 && localSize.x % m_subgroupPolicy.nativeSubgroupSize == 0 &&
+                    fullSubgroupCount <= m_subgroupPolicy.maxComputeWorkgroupSubgroups) {
+                    stage.flags |= VK_PIPELINE_SHADER_STAGE_CREATE_REQUIRE_FULL_SUBGROUPS_BIT;
+                }
+            }
 
             entry.modules.push_back(module);
             entry.stages.push_back(stage);
@@ -3157,7 +3437,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ValidateRasterizationStageInterface(shaders, moduleSpirvs, entry, program.GetExternalIndex());
 #endif
         ReflectVertexInputs(shaders, moduleSpirvs, entry);
+        ReflectViewportIndexUsage(shaders, moduleSpirvs, entry);
         ReflectFragmentOutputs(shaders, moduleSpirvs, entry);
+        ReflectPassthroughTessControlNeed(shaders, moduleSpirvs, entry);
         ReflectLayout(program, moduleSpirvs, entry);
         // A failed remap means the modules kept glslang's per-stage auto-mapped binding numbers -
         // no cross-stage unification, no set->0 normalisation - so the bindings this layout
@@ -3168,7 +3450,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // "the layout and the shader disagree", so route it through that. Set AFTER ReflectLayout,
         // which clears the flag.
         if (!remapOk) {
-            MGLOG_I("ProgramFactory::GetOrCreateProgram: declining program %u - its descriptor bindings could not "
+            MGLOG_E("ProgramFactory::GetOrCreateProgram: declining program %u - its descriptor bindings could not "
                     "be remapped, so the layout does not describe what the shader reads",
                     program.GetExternalIndex());
             entry.declinedDescriptors = true;
@@ -3202,17 +3484,249 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const VkDescriptorSetLayout descriptorSetLayout = it->second.descriptorSetLayout;
                 MGLOG_D("ProgramFactory::OnFrameBoundary: evicting idle program entry hash=0x%llx",
                         static_cast<unsigned long long>(hash));
-                // erase runs ~VkProgramObject (modules/layouts destroyed); notify after
+                // The observer destroys dependent pipelines and frees descriptor sets while
-                // so an observer never observes a half-destroyed entry through a lookup.
+                // this entry still owns its layout. Vulkan requires every descriptor set to be
-                // Observers only need the handle values to purge their keyed caches.
+                // freed before its VkDescriptorSetLayout is destroyed.
-                ++m_cacheStructureEpoch; // erase moves/kills entries: memoised pointers die
-                it = m_cache.erase(it);
                 if (m_evictionObserver != nullptr) {
                     m_evictionObserver->OnProgramEvicted(hash, descriptorSetLayout);
                 }
+                ++m_cacheStructureEpoch; // erase moves/kills entries: memoised pointers die
+                it = m_cache.erase(it);
             } else {
                 ++it;
             }
         }
     }
+
+    ProgramFactory::~ProgramFactory() {
+        for (auto& entry : m_passthroughTessControlStages) {
+            if (entry.second.module != VK_NULL_HANDLE) {
+                vkDestroyShaderModule(m_device, entry.second.module, nullptr);
+            }
+        }
+    }
+
+    String ProgramFactory::BuildPassthroughTessControlSource(Uint32 patchVertices) {
+        // The stage GL 4.6 core 11.2.2 describes when a program has an evaluation shader and no
+        // control shader: "the input patch is passed through unmodified", the output patch has
+        // as many vertices as the input one (PATCH_VERTICES), and the levels come from the
+        // PATCH_DEFAULT_OUTER_LEVEL / PATCH_DEFAULT_INNER_LEVEL state.
+        //
+        // Those two levels default to 1.0 and are baked here as literals because
+        // glPatchParameterfv - their only setter - is not implemented in this frontend (it is a
+        // stub in MG_Impl/GLImpl/Exporting/Definitions.cpp). Implementing that entry point means
+        // making the levels a parameter of this source AND of the cache key in
+        // GetOrCreatePassthroughTessControlStage; the two must move together, so they are named
+        // together here.
+        //
+        // gl_out carries gl_Position and nothing else on purpose. The evaluation stage that
+        // reads it was linked against the VERTEX stage directly, so its input gl_PerVertex holds
+        // exactly the built-ins that stage used, and its user-defined inputs (if any) come
+        // straight off the vertex stage's outputs - which a control stage sitting in between
+        // would leave unwritten. ReflectPassthroughTessControlNeed refuses those programs rather
+        // than let this write a partial interface.
+        //
+        // All four outer levels and both inner levels are written unconditionally: writing a
+        // level the evaluation stage's domain does not use is legal and ignored, and it saves
+        // this from having to know the domain.
+        String source = "#version 450 core\n";
+        source += "layout(vertices = " + std::to_string(patchVertices) + ") out;\n";
+        // gl_in and gl_out are redeclared to the exact gl_PerVertex the FRONTEND's linked programs
+        // carry - gl_Position, gl_PointSize, gl_ClipDistance[1], in that order - because Vulkan
+        // matches built-in interface blocks by their whole shape, and the two obvious spellings
+        // are both wrong:
+        //   * narrowing the block to gl_Position alone makes the evaluation stage read a patch of
+        //     zeroes (degenerate triangles, nothing rasterized), and
+        //   * taking glslang's DEFAULT block for a standalone control stage yields FOUR members -
+        //     it appends gl_CullDistance - where a linked vertex+evaluation program has three.
+        // PassthroughTessControlTest.MatchesTheFrontendPerVertexBlock is the latch: it links a
+        // vertex+evaluation program through this same compiler and fails if the two shapes ever
+        // stop agreeing, rather than letting the mismatch show up as a black frame.
+        //
+        // Only gl_Position is written. gl_PointSize is declared but left alone deliberately:
+        // writing it from a tessellation stage requires the shaderTessellationAndGeometryPointSize
+        // feature, which this renderer does not enable, so a program whose evaluation stage reads
+        // gl_in[].gl_PointSize gets an undefined point size instead of the vertex stage's - a gap
+        // this trades for not making every tessellated pipeline depend on an optional feature.
+        source += "in gl_PerVertex {\n"
+                  "    vec4 gl_Position;\n"
+                  "    float gl_PointSize;\n"
+                  "    float gl_ClipDistance[1];\n"
+                  "} gl_in[gl_MaxPatchVertices];\n";
+        source += "out gl_PerVertex {\n"
+                  "    vec4 gl_Position;\n"
+                  "    float gl_PointSize;\n"
+                  "    float gl_ClipDistance[1];\n"
+                  "} gl_out[];\n";
+        source += "void main() {\n";
+        source += "    gl_out[gl_InvocationID].gl_Position = gl_in[gl_InvocationID].gl_Position;\n";
+        source += "    gl_TessLevelOuter[0] = 1.0;\n";
+        source += "    gl_TessLevelOuter[1] = 1.0;\n";
+        source += "    gl_TessLevelOuter[2] = 1.0;\n";
+        source += "    gl_TessLevelOuter[3] = 1.0;\n";
+        source += "    gl_TessLevelInner[0] = 1.0;\n";
+        source += "    gl_TessLevelInner[1] = 1.0;\n";
+        source += "}\n";
+        return source;
+    }
+
+    VkPipelineShaderStageCreateInfo ProgramFactory::GetOrCreatePassthroughTessControlStage(Uint32 patchVertices) {
+        // A cached VK_NULL_HANDLE is a remembered failure, not a miss: returning it keeps a
+        // generator that cannot compile from re-running glslang on every draw.
+        const auto cached = m_passthroughTessControlStages.find(patchVertices);
+        if (cached != m_passthroughTessControlStages.end()) {
+            return cached->second;
+        }
+
+        VkPipelineShaderStageCreateInfo stage{VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO};
+        stage.stage = VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT;
+        stage.module = VK_NULL_HANDLE;
+        stage.pName = "main";
+
+        using namespace MG_Util::ShaderTranspiler;
+        const String source = BuildPassthroughTessControlSource(patchVertices);
+        // Same compile configuration as every other stage of every other program: this runs on
+        // the GL thread (the draw path), so the live compile env is the right one, and flags=0
+        // is the Vulkan-targeting form (CompileForOpenGL is what the GLES backend adds).
+        const SharedPtr<const CompileEnv>& env = GetCurrentCompileEnv();
+        ShaderAttrib shaderAttrib{.shaderType = GL_TESS_CONTROL_SHADER,
+                                  .sourceStr = source,
+                                  .flags = 0,
+                                  .env = env.get()};
+        auto compiled = ShaderCompiler::CompileShader(shaderAttrib);
+        if (!compiled) {
+            MGLOG_E("ProgramFactory: could not compile the pass-through tessellation control stage for "
+                    "patchVertices=%u; a program with an evaluation stage and no control stage cannot draw. %s",
+                    patchVertices, compiled.error().log.c_str());
+            m_passthroughTessControlStages.emplace(patchVertices, stage);
+            return stage;
+        }
+
+        ProgramAttrib programAttrib{};
+        programAttrib.shaders.push_back(compiled.value());
+        auto linked = ShaderCompiler::LinkProgram(programAttrib);
+        if (!linked) {
+            MGLOG_E("ProgramFactory: could not link the pass-through tessellation control stage for "
+                    "patchVertices=%u. %s", patchVertices, linked.error().log.c_str());
+            m_passthroughTessControlStages.emplace(patchVertices, stage);
+            return stage;
+        }
+
+        ProgramBinaryAttrib binaryAttrib{.shaderTypes = {GL_TESS_CONTROL_SHADER}, .program = *linked.value()};
+        auto binary = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+        if (!binary || binary.value().empty() || binary.value().front().empty()) {
+            MGLOG_E("ProgramFactory: could not generate SPIR-V for the pass-through tessellation control stage "
+                    "for patchVertices=%u", patchVertices);
+            m_passthroughTessControlStages.emplace(patchVertices, stage);
+            return stage;
+        }
+
+        const Vector<Uint>& spirv = binary.value().front();
+#if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_DEBUG
+        ValidateTransformedSpirv(spirv, ShaderStage::TessControl, 0);
+#else
+        if (m_enableSpirvValidation) {
+            MG_Util::ShaderTranspiler::ShaderCompiler::PrepareSpirvValidation();
+            ValidateTransformedSpirv(spirv, ShaderStage::TessControl, 0);
+        }
+#endif
+
+        VkShaderModuleCreateInfo smci{VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO};
+        smci.codeSize = spirv.size() * sizeof(Uint);
+        smci.pCode = spirv.data();
+        VkShaderModule module = VK_NULL_HANDLE;
+        const VkResult result = vkCreateShaderModule(m_device, &smci, nullptr, &module);
+        if (result != VK_SUCCESS) {
+            MGLOG_E("ProgramFactory: vkCreateShaderModule failed (%d) for the pass-through tessellation control "
+                    "stage for patchVertices=%u", static_cast<Int>(result), patchVertices);
+            m_passthroughTessControlStages.emplace(patchVertices, stage);
+            return stage;
+        }
+
+        stage.module = module;
+        MGLOG_D("ProgramFactory: built the pass-through tessellation control stage for patchVertices=%u "
+                "(GL 4.6 11.2.2; Vulkan has no fixed-function equivalent)", patchVertices);
+        m_passthroughTessControlStages.emplace(patchVertices, stage);
+        return stage;
+    }
+
+    void ProgramFactory::ReflectPassthroughTessControlNeed(
+        const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+        const Vector<Vector<Uint>>& spirv,
+        VkProgramObject& entry) const {
+        entry.needsPassthroughTessControl = false;
+        entry.passthroughTessControlEmulatable = false;
+
+        Bool hasTessEval = false;
+        Bool hasTessControl = false;
+        SizeT tessEvalModuleIndex = 0;
+        for (SizeT i = 0; i < shaders.size(); ++i) {
+            if (!shaders[i]) continue;
+            const auto stage = shaders[i]->GetShaderStage();
+            if (stage == ShaderStage::TessControl) hasTessControl = true;
+            if (stage == ShaderStage::TessEval) {
+                hasTessEval = true;
+                tessEvalModuleIndex = i;
+            }
+        }
+        if (!hasTessEval || hasTessControl) return;
+
+        entry.needsPassthroughTessControl = true;
+
+        if (tessEvalModuleIndex >= spirv.size() || spirv[tessEvalModuleIndex].empty()) return;
+        const auto& module = spirv[tessEvalModuleIndex];
+
+        SpvReflectShaderModule reflectModule{};
+        const SpvReflectResult createResult =
+            spvReflectCreateShaderModule(module.size() * sizeof(Uint), module.data(), &reflectModule);
+        if (createResult != SPV_REFLECT_RESULT_SUCCESS) {
+            MGLOG_E("ProgramFactory::ReflectPassthroughTessControlNeed: reflection failed (result=%d); the "
+                    "evaluation stage's inputs are unknown, so the pass-through is not offered",
+                    static_cast<Int>(createResult));
+            return;
+        }
+
+        uint32_t inputCount = 0;
+        SpvReflectResult reflectResult = spvReflectEnumerateInputVariables(&reflectModule, &inputCount, nullptr);
+        Vector<SpvReflectInterfaceVariable*> inputs(inputCount);
+        if (reflectResult == SPV_REFLECT_RESULT_SUCCESS && inputCount > 0) {
+            reflectResult = spvReflectEnumerateInputVariables(&reflectModule, &inputCount, inputs.data());
+        }
+        if (reflectResult != SPV_REFLECT_RESULT_SUCCESS) {
+            spvReflectDestroyShaderModule(&reflectModule);
+            return;
+        }
+
+        // The question is only ever "does this stage read anything a control stage would have to
+        // forward", and the answer is: does it have a LOCATION. A located input is a user-defined
+        // varying (or a per-patch input), which the vertex stage writes today and would stop
+        // reaching once a control stage sits in between - the pass-through carries gl_Position and
+        // nothing else, so such a program is declined instead of being handed undefined values.
+        // Everything without a location is a built-in: gl_in, gl_TessCoord, gl_PatchVerticesIn,
+        // gl_PrimitiveID, gl_TessLevel*, all either forwarded or generated for the evaluation
+        // stage by the tessellator itself.
+        //
+        // This deliberately does NOT judge on SpvReflectInterfaceVariable::built_in. gl_in is an
+        // array of interface blocks, and for those SPIRV-Reflect reports built_in == -1 on the
+        // block AND leaves every member's built_in at 0 - which is SpvBuiltInPosition, so a
+        // member walk reads "Position, Position, Position" for a {Position, PointSize,
+        // ClipDistance} block and would accept anything on the strength of parse garbage. The
+        // location, by contrast, is decorated on the OpVariable and is what SPIRV-Reflect reads
+        // straight through.
+        constexpr Uint32 kNoLocation = 0xFFFFFFFFu;
+        Bool emulatable = true;
+        for (auto* input : inputs) {
+            if (input == nullptr) continue;
+            if (input->location == kNoLocation) continue;
+            MGLOG_E("ProgramFactory: a tessellation evaluation stage with no control stage reads the "
+                    "user-defined input '%s' at location=%u; a synthesized control stage cannot forward it, so "
+                    "this program's draws are declined rather than fed an undefined varying",
+                    input->name != nullptr ? input->name : "<null>", input->location);
+            emulatable = false;
+            break;
+        }
+
+        spvReflectDestroyShaderModule(&reflectModule);
+        entry.passthroughTessControlEmulatable = emulatable;
+    }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -33,7 +33,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             CombinedImageSampler,
             UniformTexelBuffer,
             StorageBuffer,
-            StorageImage
+            StorageImage,
+            // GLSL `imageBuffer` - a buffer texture reached through an IMAGE unit rather than a
+            // texture unit. Vulkan spells it VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER, which is a
+            // VkBufferView like UniformTexelBuffer and not a VkImageView like StorageImage: it is
+            // the one image uniform whose descriptor is a buffer. Appended, never inserted -
+            // DescriptorKeyHash mixes the enumerator's value.
+            StorageTexelBuffer
         };
 
         enum class CompileOptionBit : Uint {
@@ -70,6 +76,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         using CompileOptionFlags = Flags<CompileOptionBit>;
         using HashType = Uint64;
 
+        struct UpdateAfterBindLimits {
+            Bool enabled = false;
+            Uint32 maxPerStageSamplers = 0;
+            Uint32 maxPerStageUniformBuffers = 0;
+            Uint32 maxPerStageStorageBuffers = 0;
+            Uint32 maxPerStageSampledImages = 0;
+            Uint32 maxPerStageStorageImages = 0;
+            Uint32 maxPerStageResources = 0;
+            Uint32 maxSetSamplers = 0;
+            Uint32 maxSetUniformBuffers = 0;
+            Uint32 maxSetUniformBuffersDynamic = 0;
+            Uint32 maxSetStorageBuffers = 0;
+            Uint32 maxSetStorageBuffersDynamic = 0;
+            Uint32 maxSetSampledImages = 0;
+            Uint32 maxSetStorageImages = 0;
+        };
+
         struct VkProgramObject {
             static constexpr Uint32 kMaxVertexInputLocations = 32;
 
@@ -82,6 +105,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             // Layout data (previously in separate VkProgramLayout)
             VkDescriptorSetLayout descriptorSetLayout = VK_NULL_HANDLE;
+            // True only when this layout passed every descriptor-indexing feature and
+            // update-after-bind limit gate at reflection time. It controls both the
+            // layout/binding flags and the pool class used by UniformManager.
+            Bool usesUpdateAfterBind = false;
             VkPipelineLayout pipelineLayout = VK_NULL_HANDLE;
             Vector<DescriptorBindingKind> bindingKinds;
             // The bindings this program actually declares, ascending. bindingKinds is sized to the
@@ -103,6 +130,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<Int> samplerUniformLocationByBinding;
             Vector<TextureTarget> samplerTextureTargetByBinding;
             Vector<SamplerNumericDomain> samplerNumericDomainByBinding;
+            // Shared by StorageImage and StorageTexelBuffer bindings: a binding is one kind or
+            // the other, never both, and both need exactly the same thing - the format the
+            // shader declared, so the per-draw resolve can tell a typed declaration from a
+            // formatless one. Kept as one pair rather than two so the move operations below
+            // cannot drift out of sync with a field that only one kind populates.
             Vector<VkFormat> storageImageFormatByBinding;
             Vector<Bool> storageImageUsesBindingFormatByBinding;
             Vector<String> storageBlockNameByBinding;
@@ -140,6 +172,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // PROGRAM rather than of the variant: the zeroed variant leaves the variable
             // declared, so both variants answer the same and the draw path can ask either.
             Bool readsBaseVertexBuiltin = false;
+            // Some pre-rasterization stage assigns gl_ViewportIndex. Its pipeline declares
+            // viewportCount = the renderer's rasterizable viewport count instead of 1, and its
+            // draws push the whole viewport/scissor array; every other program keeps the
+            // single-viewport fast path untouched. Part of the program's identity (folded into
+            // the pipeline hash through programHash), so no memo can serve the wrong shape.
+            Bool writesViewportIndexBuiltin = false;
+            // This program has a tessellation EVALUATION stage and no tessellation CONTROL
+            // stage. GL allows that (4.6 core 11.2.2: with no control shader the input patch
+            // is passed through unmodified, the output patch size is PATCH_VERTICES, and the
+            // levels come from the PATCH_DEFAULT_*_LEVEL state); Vulkan does not - either both
+            // tessellation stages are present or neither
+            // (VUID-VkGraphicsPipelineCreateInfo-pStages-00730). So the draw path has to supply
+            // the pass-through stage GL describes; see GetOrCreatePassthroughTessControlStage.
+            Bool needsPassthroughTessControl = false;
+            // ...and the pass-through this renderer can synthesize carries gl_Position and
+            // nothing else, so it is only correct when the evaluation stage's inputs are
+            // built-ins. A user-defined varying would arrive at the evaluation stage
+            // UNWRITTEN once a control stage sits between it and the vertex stage, which is
+            // silently wrong pixels rather than a crash - so those programs are declined
+            // instead (PipelineFactory::CreatePipeline refuses the pipeline and the draw is
+            // skipped). See ReflectPassthroughTessControlNeed.
+            Bool passthroughTessControlEmulatable = false;
             // Frame-boundary counter value of the last GetOrCreateProgram hit; drives
             // cache eviction (see OnFrameBoundary). Mutable: the draw snapshot's memoised
             // entry pointer re-stamps use through a const reference (StampProgramUse).
@@ -163,6 +217,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // a pipeline failure would be reported against the wrong SPIR-V.
                 stageSpirvDigests = std::move(other.stageSpirvDigests);
                 descriptorSetLayout = other.descriptorSetLayout;
+                usesUpdateAfterBind = other.usesUpdateAfterBind;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
                 activeBindings = std::move(other.activeBindings);
@@ -191,9 +246,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
                 readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
+                needsPassthroughTessControl = other.needsPassthroughTessControl;
+                passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
+                other.usesUpdateAfterBind = false;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
                 other.declinedDescriptors = false;
@@ -205,6 +264,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
                 other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
+                other.needsPassthroughTessControl = false;
+                other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
             }
             VkProgramObject& operator=(VkProgramObject&& other) noexcept {
@@ -217,6 +279,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 modules = std::move(other.modules);
                 stageSpirvDigests = std::move(other.stageSpirvDigests); // travels with `modules` - see the move ctor
                 descriptorSetLayout = other.descriptorSetLayout;
+                usesUpdateAfterBind = other.usesUpdateAfterBind;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
                 activeBindings = std::move(other.activeBindings);
@@ -245,9 +308,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
                 readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
+                needsPassthroughTessControl = other.needsPassthroughTessControl;
+                passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
+                other.usesUpdateAfterBind = false;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
                 other.declinedDescriptors = false;
@@ -259,6 +326,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
                 other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
+                other.needsPassthroughTessControl = false;
+                other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
                 return *this;
             }
@@ -302,15 +372,45 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             virtual void OnProgramEvicted(HashType programHash, VkDescriptorSetLayout descriptorSetLayout) = 0;
         };
 
-        explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings = 16,
+        // How this factory's compute modules implement GL_KHR_shader_subgroup. Computed
-                                Bool shaderDrawParametersEnabled = false,
+        // once at renderer initialization (SubgroupSupportPolicy.h + the device's
-                                Bool unformattedFloatStorageImagesEnabled = false)
+        // subgroup properties) so lowering can never disagree with the advertised
+        // capabilities. Native subgroup operations always execute natively; the two
+        // repair passes patch modules AROUND them, and the emulation only replaces them
+        // on opted-in devices with no subgroup support at all.
+        struct SubgroupLoweringPolicy {
+            Bool emulateSubgroups = false;      // MOBILEGL_MAGMA_EMULATE_SUBGROUP, no-native-support devices
+            Bool fixIterationRPSubgroupScratch = false; // patch iterationRP's under-declared scratch
+            Bool fixIterationRPBarrier = false; // repair Program 203's shared-scratch race
+            Bool deriveNumSubgroups = false;    // repair the NumSubgroups builtin
+            Bool requireFullSubgroups = false;  // computeFullSubgroups enabled on the device
+            Uint32 nativeSubgroupSize = 0;
+            // Full-subgroup launches are bounded by this device limit; a dispatch whose
+            // workgroup needs more subgroups than this cannot request the flag.
+            Uint32 maxComputeWorkgroupSubgroups = 0;
+            // VkPhysicalDeviceLimits::maxComputeSharedMemorySize; bounds the scratch the
+            // emulation pass may add (0 falls back to the Vulkan minimum, 16384).
+            Uint32 maxComputeSharedMemoryBytes = 0;
+        };
+
+        explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings,
+                                Bool shaderDrawParametersEnabled,
+                                Bool unformattedFloatStorageImagesEnabled,
+                                Bool enableSpirvValidation,
+                                UpdateAfterBindLimits updateAfterBindLimits,
+                                SubgroupLoweringPolicy subgroupPolicy)
             : m_device(device), m_maxBindings(maxBindings), m_config(config),
               m_shaderDrawParametersEnabled(shaderDrawParametersEnabled),
-              m_unformattedFloatStorageImagesEnabled(unformattedFloatStorageImagesEnabled) {
+              m_unformattedFloatStorageImagesEnabled(unformattedFloatStorageImagesEnabled),
+              m_enableSpirvValidation(enableSpirvValidation),
+              m_updateAfterBindLimits(updateAfterBindLimits),
+              m_subgroupPolicy(subgroupPolicy) {
             VkProgramObject::s_device = device;
         }
-        ~ProgramFactory() = default;
+        // Destroys the pass-through tessellation control modules. Runs while the device is
+        // still alive for the same reason ~VkProgramObject's does: this factory outlives
+        // nothing that owns the device.
+        ~ProgramFactory();
         ProgramFactory(const ProgramFactory&) = delete;
 
         HashType ComputeHash(const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags) const;
@@ -362,6 +462,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Shared by the two above: does any entry point list an input variable decorated with
         // this builtin?
         static Bool ReflectedDeclaresInputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
+        // True when an entry point writes the ViewportIndex builtin (gl_ViewportIndex), i.e. when
+        // the program can route primitives to a viewport other than 0 and its pipeline therefore
+        // has to declare more than one. Asks about OUTPUT variables because that is the direction
+        // a pre-rasterization stage declares it in.
+        static Bool ReflectedWritesViewportIndexBuiltin(const SpvReflectShaderModule& reflectModule);
+        static Bool ReflectedDeclaresOutputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
+
+        // The pass-through tessellation control stage GL 4.6 core 11.2.2 describes for a
+        // program that has an evaluation stage and no control stage, for an input patch of
+        // `patchVertices` control points. Returned BY VALUE (a stage description is a POD, and
+        // the cache below is a rehashing map, so a pointer into it would not survive the next
+        // distinct patch size). `.module == VK_NULL_HANDLE` means the stage could not be built:
+        // the caller then has no control stage to inject, and CreatePipeline refuses the
+        // pipeline rather than handing the driver a half-tessellated one.
+        //
+        // Keyed on the patch size because GL takes the output patch size from PATCH_VERTICES,
+        // which is draw state, not link state - the CTS case that motivated this links at the
+        // default 3 and draws at 4. The pipeline cache already re-keys on patchControlPoints,
+        // so the module a pipeline was built with is part of that pipeline's identity.
+        // Compiling is bounded by the number of distinct patch sizes a program draws with
+        // (MAX_PATCH_VERTICES = 32 in the worst case, one or two in practice) and only ever
+        // happens for the rare program that has no control stage at all.
+        VkPipelineShaderStageCreateInfo GetOrCreatePassthroughTessControlStage(Uint32 patchVertices);
+
+        // Source of the module above. Exposed for tests: the generated GLSL is the whole
+        // contract with the evaluation stage, so it is worth pinning independently of a device.
+        static String BuildPassthroughTessControlSource(Uint32 patchVertices);
 
     private:
         struct ProgramLookupCache {
@@ -375,11 +502,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void ReflectVertexInputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                      const Vector<Vector<Uint>>& spirv,
                      VkProgramObject& entry) const;
+        void ReflectViewportIndexUsage(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                       const Vector<Vector<Uint>>& spirv,
+                                       VkProgramObject& entry) const;
         void ReflectFragmentOutputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                         const Vector<Vector<Uint>>& spirv,
                         VkProgramObject& entry) const;
         void ReflectLayout(const MG_State::GLState::ProgramObject& program, const Vector<Vector<Uint>>& spirv,
                            VkProgramObject& entry) const;
+        // Fills needsPassthroughTessControl / passthroughTessControlEmulatable off the linked
+        // modules. Const and reflection-only: it decides nothing about the pipeline, it only
+        // records what the evaluation stage's input interface is made of.
+        void ReflectPassthroughTessControlNeed(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                               const Vector<Vector<Uint>>& spirv,
+                                               VkProgramObject& entry) const;
 
         VkDevice m_device = VK_NULL_HANDLE;
         Uint32 m_maxBindings = 0;
@@ -391,6 +527,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // True only when the logical device enabled both
         // shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
         Bool m_unformattedFloatStorageImagesEnabled = false;
+        // Startup snapshot used only by internally synthesized shader modules, which do not
+        // originate from a ProgramLinkTask.
+        Bool m_enableSpirvValidation = false;
+        // Device feature and limit gate resolved before vkCreateDevice. Keeping it in
+        // the factory lets each reflected layout choose ordinary descriptors when its
+        // own counts would exceed the update-after-bind budget.
+        UpdateAfterBindLimits m_updateAfterBindLimits{};
+        SubgroupLoweringPolicy m_subgroupPolicy{};
         // See SetDefaultFramebufferHeight. 0 means "not known yet"; the FragCoordYFlip bit is
         // never set before the swapchain exists, so no variant can be compiled against it.
         Uint32 m_defaultFramebufferHeight = 0;
@@ -400,6 +544,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // See GetCacheStructureEpoch(). Starts at 1 so a zero-initialized memo can never match.
         Uint64 m_cacheStructureEpoch = 1;
         IEvictionObserver* m_evictionObserver = nullptr;
+        // Pass-through tessellation control stages by input patch size. Never evicted: at most
+        // MAX_PATCH_VERTICES entries exist for the lifetime of the device, and every pipeline
+        // ever built from one keeps referencing its module. A failed build is cached as
+        // VK_NULL_HANDLE so a broken generator costs one compile, not one per draw.
+        UnorderedMap<Uint32, VkPipelineShaderStageCreateInfo> m_passthroughTessControlStages;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.cpp

@@ -157,7 +157,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         MGLOG_I("Got %d surface formats:", swapchainCapabilities.surfaceFormats.size());
         for (const auto& sf : swapchainCapabilities.surfaceFormats) {
-            MGLOG_I("    [%s, %s]", string_VkFormat(sf.format), string_VkColorSpaceKHR(sf.colorSpace));
+            MGLOG_D("    [%s, %s]", string_VkFormat(sf.format), string_VkColorSpaceKHR(sf.colorSpace));
         }
 
         const auto pickedSurfaceFormat = ChooseSwapchainSurfaceFormat(swapchainCapabilities.surfaceFormats);
@@ -166,7 +166,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         MGLOG_I("Got %d present modes:", swapchainCapabilities.presentModes.size());
         for (const auto& pm : swapchainCapabilities.presentModes) {
-            MGLOG_I("    %s", string_VkPresentModeKHR(pm));
+            MGLOG_D("    %s", string_VkPresentModeKHR(pm));
         }
 
         const auto presentMode = ChooseSwapchainPresentMode(swapchainCapabilities.presentModes);

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -156,13 +156,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             frame.descriptorPools.clear();
 
             VkDescriptorPool initialPool = VK_NULL_HANDLE;
-            if (!CreateDescriptorPool(m_setsPerFrame, initialPool)) {
+            if (!CreateDescriptorPool(m_setsPerFrame, false, initialPool)) {
-                MGLOG_E("UniformDescriptorBinder::Initialize failed: cannot create frame descriptor pool %u",
+                MGLOG_E_ONCE("UniformDescriptorBinder::Initialize failed: cannot create frame descriptor pool %u",
                         frameIndex);
                 Shutdown();
                 return false;
             }
-            frame.descriptorPools.push_back({initialPool, m_setsPerFrame, 0});
+            frame.descriptorPools.push_back({initialPool, m_setsPerFrame, 0, false});
             MGLOG_D("UniformDescriptorBinder: frame %u descriptor pool created (maxSets=%u)", frameIndex,
                     m_setsPerFrame);
         }
@@ -305,7 +305,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // texture/sampler resolution, completeness probe, sync, layout handling, sampler
         // and view lookups - would recompute the identical descriptor.
         if (trustUnchangedHint && descriptorMemoUsable && binding < m_samplerResolveMemo.size() &&
-            m_samplerResolveMemo[binding].infoValid) {
+            m_samplerResolveMemo[binding].infoValid &&
+            m_samplerResolveMemo[binding].infoProgramLifetimeId == program.GetLifetimeId()) {
             outImageInfo = m_samplerResolveMemo[binding].info;
             return true;
         }
@@ -345,13 +346,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             fallbackHolder = GetFallbackTexture(preferredTarget);
             texture = fallbackHolder.get();
             if (texture == nullptr) {
-                MGLOG_E("ResolveSamplerDescriptor: no fallback texture available for binding=%u ('%s') "
+                MGLOG_E_ONCE("ResolveSamplerDescriptor: no fallback texture available for binding=%u ('%s') "
                         "location=%d unit=%d target=%d",
                         binding, programObj.samplerNameByBinding[binding].c_str(), location, unit,
                         static_cast<Int>(preferredTarget));
                 return false;
             }
-            MGLOG_W(
+            MGLOG_W_ONCE(
                 "ResolveSamplerDescriptor: using fallback texture for unbound sampler binding=%u ('%s') location=%d unit=%d target=%d",
                 binding, programObj.samplerNameByBinding[binding].c_str(), location, unit,
                 static_cast<Int>(preferredTarget));
@@ -360,7 +361,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const MG_State::GLState::SamplerObject* samplerToUse =
             samplerOverride ? samplerOverride.get() : texture->GetSamplerObject().get();
         if (samplerToUse == nullptr) {
-            MGLOG_E(
+            MGLOG_E_ONCE(
                 "ResolveSamplerDescriptor: sampler binding %u ('%s') has no sampler object (textureId=%d location=%d unit=%d)",
                 binding, programObj.samplerNameByBinding[binding].c_str(), texture->GetExternalIndex(), location,
                 unit);
@@ -368,7 +369,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         VkTextureManager::TextureResource* resource = m_textureManager->SyncTextureAndGetDescriptor(*texture);
         if (resource == nullptr) {
-            MGLOG_E(
+            MGLOG_E_ONCE(
                 "ResolveSamplerDescriptor: sampler binding %u ('%s') failed to create/sync texture resource (textureId=%d target=%d location=%d unit=%d)",
                 binding, programObj.samplerNameByBinding[binding].c_str(), texture->GetExternalIndex(),
                 static_cast<Int>(texture->GetTarget()), location, unit);
@@ -380,7 +381,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Int attachmentLevel = 0;
             if (drawFbo &&
                 FindFramebufferAttachmentForTexture(*drawFbo, *texture, attachmentType, attachmentLevel)) {
-                MGLOG_W("ResolveSamplerDescriptor: framebuffer feedback loop detected: textureId=%d is bound "
+                MGLOG_W_ONCE("ResolveSamplerDescriptor: framebuffer feedback loop detected: textureId=%d is bound "
                         "for sampling at binding=%u, but is also attached to drawFbo=%u as %s (level=%d, "
                         "trackedLayout=%d)",
                         texture->GetExternalIndex(), binding, drawFbo->GetExternalIndex(),
@@ -390,7 +391,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             const Bool readyForSampling = m_textureManager->TransitionTextureForSampling(commandBuffer, *texture);
             if (!readyForSampling) {
-                MGLOG_E("ResolveSamplerDescriptor: failed to transition textureId=%d for sampler binding=%u",
+                MGLOG_E_ONCE("ResolveSamplerDescriptor: failed to transition textureId=%d for sampler binding=%u",
                         texture->GetExternalIndex(), binding);
                 return false;
             }
@@ -432,7 +433,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         }
         if (sampledViewFormat == VK_FORMAT_UNDEFINED) {
-            MGLOG_E("ResolveSamplerDescriptor: no compatible sampled view for binding=%u ('%s') "
+            MGLOG_E_ONCE("ResolveSamplerDescriptor: no compatible sampled view for binding=%u ('%s') "
                     "textureId=%d imageFormat=%d numericDomain=%d",
                     binding, programObj.samplerNameByBinding[binding].c_str(), texture->GetExternalIndex(),
                     static_cast<Int>(resource->format), static_cast<Int>(numericDomain));
@@ -445,7 +446,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 ? resource->sampledView
                 : m_textureManager->GetOrCreateSampledImageView(*texture, sampledViewFormat);
         if (sampledImageView == VK_NULL_HANDLE) {
-            MGLOG_E("ResolveSamplerDescriptor: failed to resolve sampled view for binding=%u ('%s') "
+            MGLOG_E_ONCE("ResolveSamplerDescriptor: failed to resolve sampled view for binding=%u ('%s') "
                     "textureId=%d imageFormat=%d viewFormat=%d numericDomain=%d",
                     binding, programObj.samplerNameByBinding[binding].c_str(), texture->GetExternalIndex(),
                     static_cast<Int>(resource->format), static_cast<Int>(sampledViewFormat),
@@ -504,6 +505,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (binding < m_samplerResolveMemo.size()) {
             if (descriptorMemoUsable) {
                 m_samplerResolveMemo[binding].info = outImageInfo;
+                m_samplerResolveMemo[binding].infoProgramLifetimeId = program.GetLifetimeId();
                 m_samplerResolveMemo[binding].infoValid = true;
             } else {
                 // An arrayed binding publishes nothing here, and clears what a previous program
@@ -540,11 +542,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         outImageInfo = {
             .sampler = m_samplerManager->GetOrCreateSampler(*samplerBindingOverride.sampler,
-                                                            *samplerBindingOverride.texture),
+                                                            *samplerBindingOverride.texture,
+                                                            samplerBindingOverride.forceNearestFiltering,
+                                                            resource->sampledLevelCount),
             .imageView = samplerBindingOverride.imageView != VK_NULL_HANDLE ?
                 samplerBindingOverride.imageView :
                 (resource->sampledView != VK_NULL_HANDLE ? resource->sampledView : resource->fullView),
-            .imageLayout = resource->layout,
+            .imageLayout = samplerBindingOverride.imageLayout != VK_IMAGE_LAYOUT_UNDEFINED ?
+                samplerBindingOverride.imageLayout : resource->layout,
         };
         return outImageInfo.sampler != VK_NULL_HANDLE;
     }
@@ -671,14 +676,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         SharedPtr<MG_State::GLState::ITextureObject> texture;
         if (!ResolveSamplerTexture(program, programObj, binding, texture) || texture == nullptr) {
-            MGLOG_E("ResolveTexelBufferDescriptor: texture buffer binding %u ('%s') is unbound", binding,
+            MGLOG_E_ONCE("ResolveTexelBufferDescriptor: texture buffer binding %u ('%s') is unbound", binding,
                     programObj.samplerNameByBinding[binding].c_str());
             return false;
         }
 
         if (texture->GetStorageType() != TextureStorageType::Buffer ||
             texture->GetTarget() != TextureTarget::TextureBuffer) {
-            MGLOG_E(
+            MGLOG_E_ONCE(
                 "ResolveTexelBufferDescriptor: binding %u ('%s') expected texture buffer, got textureId=%u target=%d storage=%d",
                 binding, programObj.samplerNameByBinding[binding].c_str(), texture->GetExternalIndex(),
                 static_cast<Int>(texture->GetTarget()), static_cast<Int>(texture->GetStorageType()));
@@ -688,14 +693,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto* textureBuffer = static_cast<MG_State::GLState::TextureObjectBuffer*>(texture.get());
         const auto& bufferObject = textureBuffer->GetBufferBindingSlot().GetBoundObject();
         if (bufferObject == nullptr) {
-            MGLOG_E("ResolveTexelBufferDescriptor: texture buffer binding %u ('%s') has no GL buffer bound",
+            MGLOG_E_ONCE("ResolveTexelBufferDescriptor: texture buffer binding %u ('%s') has no GL buffer bound",
                     binding, programObj.samplerNameByBinding[binding].c_str());
             return false;
         }
 
         BufferSlice slice{};
         if (!m_bufferManager->AcquireResidentSlice(BufferKind::TextureBuffer, bufferObject, slice) || !slice.IsValid()) {
-            MGLOG_E("ResolveTexelBufferDescriptor: failed to sync GL buffer %u for texture buffer %u",
+            MGLOG_E_ONCE("ResolveTexelBufferDescriptor: failed to sync GL buffer %u for texture buffer %u",
                     bufferObject->GetExternalIndex(), texture->GetExternalIndex());
             return false;
         }
@@ -703,7 +708,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const auto internalFormat = textureBuffer->GetFormat();
         const VkFormat vkFormat = MG_Util::ConvertTextureInternalFormatToVkEnum(internalFormat);
         if (vkFormat == VK_FORMAT_UNDEFINED) {
-            MGLOG_E("ResolveTexelBufferDescriptor: unsupported texture buffer internal format %d",
+            MGLOG_E_ONCE("ResolveTexelBufferDescriptor: unsupported texture buffer internal format %d",
                     static_cast<Int>(internalFormat));
             return false;
         }
@@ -719,7 +724,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             viewRange = (viewRange / texelSize) * texelSize;
         }
         if (viewRange == 0) {
-            MGLOG_E("ResolveTexelBufferDescriptor: texture buffer %u has empty view range", texture->GetExternalIndex());
+            MGLOG_E_ONCE("ResolveTexelBufferDescriptor: texture buffer %u has empty view range", texture->GetExternalIndex());
             return false;
         }
 
@@ -733,7 +738,151 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkBufferView bufferView = VK_NULL_HANDLE;
         const VkResult result = vkCreateBufferView(m_device, &viewInfo, nullptr, &bufferView);
         if (result != VK_SUCCESS || bufferView == VK_NULL_HANDLE) {
-            MGLOG_E("ResolveTexelBufferDescriptor: vkCreateBufferView failed result=%d format=%d range=%zu",
+            MGLOG_E_ONCE("ResolveTexelBufferDescriptor: vkCreateBufferView failed result=%d format=%d range=%zu",
+                    result, static_cast<Int>(vkFormat), static_cast<SizeT>(viewRange));
+            return false;
+        }
+
+        m_frames[frameIndex].texelBufferViews.push_back(bufferView);
+        outBufferView = bufferView;
+        return true;
+    }
+
+    // GLSL `imageBuffer`. The one image uniform whose Vulkan descriptor is a VkBufferView rather
+    // than a VkImageView, so it is half ResolveStorageImageDescriptor (the resource comes from an
+    // IMAGE unit, i.e. from glBindImageTexture, not from a texture unit) and half
+    // ResolveTexelBufferDescriptor (the descriptor is a buffer view over the GL buffer the
+    // texture is attached to).
+    //
+    // Before this existed the descriptor kind reflected as SPV_REFLECT_DESCRIPTOR_TYPE_STORAGE_-
+    // TEXEL_BUFFER and fell into ReflectDescriptorTypeToBindingKind's `default:`, whose only
+    // complaint is an assert that compiles out above DEBUG - so a release build declared no
+    // binding at all for a uniform the shader still read, and lavapipe segfaulted inside pipeline
+    // creation on the JIT worker thread. KHR-GL44.multi_bind.dispatch_bind_image_textures is the
+    // case that carries it.
+    Bool UniformManager::ResolveStorageTexelBufferDescriptor(const MG_State::GLState::ProgramObject& program,
+                                                             const ProgramFactory::VkProgramObject& programObj,
+                                                             Uint32 binding, Uint32 frameIndex,
+                                                             VkBufferView& outBufferView) {
+        outBufferView = VK_NULL_HANDLE;
+        MOBILEGL_ASSERT(m_bufferManager != nullptr, "ResolveStorageTexelBufferDescriptor: buffer manager is null");
+        MOBILEGL_ASSERT(MG_State::pGLContext != nullptr, "ResolveStorageTexelBufferDescriptor: GL context is null");
+        MOBILEGL_ASSERT(frameIndex < m_frames.size(),
+                        "ResolveStorageTexelBufferDescriptor: frame index out of range");
+        MOBILEGL_ASSERT(binding < programObj.samplerUniformLocationByBinding.size(),
+                        "ResolveStorageTexelBufferDescriptor: binding %u out of range", binding);
+
+        const Int location = programObj.samplerUniformLocationByBinding[binding];
+        if (location < 0) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: binding %u ('%s') has no uniform location", binding,
+                    programObj.samplerNameByBinding[binding].c_str());
+            return false;
+        }
+        const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
+        if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: image unit %d out of range for binding %u", imageUnit,
+                    binding);
+            return false;
+        }
+
+        auto& imageBinding = MG_State::pGLContext->GetImageTextureBinding(imageUnit);
+        const auto& texture = imageBinding.Texture;
+        if (texture == nullptr) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: image unit %d is unbound for binding %u", imageUnit,
+                    binding);
+            return false;
+        }
+        if (texture->GetStorageType() != TextureStorageType::Buffer ||
+            texture->GetTarget() != TextureTarget::TextureBuffer) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: binding %u ('%s') expected a texture buffer on image "
+                    "unit %d, got textureId=%u target=%d storage=%d",
+                    binding, programObj.samplerNameByBinding[binding].c_str(), imageUnit,
+                    texture->GetExternalIndex(), static_cast<Int>(texture->GetTarget()),
+                    static_cast<Int>(texture->GetStorageType()));
+            return false;
+        }
+
+        auto* textureBuffer = static_cast<MG_State::GLState::TextureObjectBuffer*>(texture.get());
+        const auto& bufferObject = textureBuffer->GetBufferBindingSlot().GetBoundObject();
+        if (bufferObject == nullptr) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: texture buffer on image unit %d has no GL buffer bound",
+                    imageUnit);
+            return false;
+        }
+
+        // Unlike the sampled texel buffer, the shader MAY write this one, and those writes land
+        // in GPU memory behind the frontend's CPU shadow - which is what MapBuffer and
+        // GetBufferSubData read. Same two calls, and for the same reason, as the storage-block
+        // path above - but only the residency is unconditional. Marking a GL_READ_ONLY binding
+        // GPU-written would make the next map or readback wait for a dispatch that could not have
+        // changed a byte of it.
+        bufferObject->EnsureGpuResidentStorage();
+        if (imageBinding.Access != GL_READ_ONLY) {
+            bufferObject->MarkGpuWritten();
+        }
+
+        BufferSlice slice{};
+        if (!m_bufferManager->AcquireResidentSlice(BufferKind::TextureBuffer, bufferObject, slice) ||
+            !slice.IsValid()) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: failed to sync GL buffer %u for texture buffer %u",
+                    bufferObject->GetExternalIndex(), texture->GetExternalIndex());
+            return false;
+        }
+
+        // The format the SHADER declared wins over the one glBindImageTexture named, on the same
+        // policy as a storage image: a typed `layout(r32ui) uniform uimageBuffer` must be read as
+        // r32ui whatever the texture's own attachment format says. Falling back, in order:
+        // reflected format, then the bind format, then the texture's attached format.
+        MOBILEGL_ASSERT(binding < programObj.storageImageFormatByBinding.size(),
+                        "ResolveStorageTexelBufferDescriptor: binding %u has no reflected format slot", binding);
+        const auto internalFormat = textureBuffer->GetFormat();
+        const VkFormat resourceFormat = MG_Util::ConvertTextureInternalFormatToVkEnum(internalFormat);
+        const VkFormat reflectedFormat = programObj.storageImageFormatByBinding[binding];
+        VkFormat vkFormat = reflectedFormat;
+        if (vkFormat == VK_FORMAT_UNDEFINED && imageBinding.Format != 0) {
+            vkFormat = MG_Util::ConvertTextureInternalFormatToVkEnum(
+                MG_Util::ConvertGLEnumToTextureInternalFormat(imageBinding.Format));
+        }
+        if (vkFormat == VK_FORMAT_UNDEFINED) {
+            vkFormat = resourceFormat;
+        }
+        if (vkFormat == VK_FORMAT_UNDEFINED) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: unsupported image buffer format (internal=%d bind=0x%x)",
+                    static_cast<Int>(internalFormat), imageBinding.Format);
+            return false;
+        }
+
+        // Sized from the TEXTURE's attached format even though the view may carry a different
+        // one. That is not a shortcut: GL requires the shader's format qualifier, the format
+        // passed to glBindImageTexture and the texture's own internal format to belong to the
+        // same format CLASS (GL 4.6 core, table 8.27), and every member of a class has the same
+        // texel size. So the three can disagree on interpretation and never on bytes - which is
+        // what the range below has to be a whole multiple of.
+        const VkDeviceSize texelSize =
+            static_cast<VkDeviceSize>(MG_Util::GetSizedInternalFormatSizeInBytes(internalFormat));
+        const VkDeviceSize rangeOffset = static_cast<VkDeviceSize>(textureBuffer->GetBufferRangeOffset());
+        const VkDeviceSize rangeSize = static_cast<VkDeviceSize>(textureBuffer->GetBufferRangeSizeInBytes());
+        VkDeviceSize viewRange = std::min(rangeSize, slice.size > rangeOffset ? slice.size - rangeOffset : 0);
+        if (texelSize > 0) {
+            viewRange = (viewRange / texelSize) * texelSize;
+        }
+        if (viewRange == 0) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: texture buffer %u has empty view range",
+                    texture->GetExternalIndex());
+            return false;
+        }
+
+        VkBufferViewCreateInfo viewInfo{};
+        viewInfo.sType = VK_STRUCTURE_TYPE_BUFFER_VIEW_CREATE_INFO;
+        viewInfo.buffer = slice.buffer;
+        viewInfo.format = vkFormat;
+        viewInfo.offset = slice.offset + rangeOffset;
+        viewInfo.range = viewRange;
+
+        VkBufferView bufferView = VK_NULL_HANDLE;
+        const VkResult result = vkCreateBufferView(m_device, &viewInfo, nullptr, &bufferView);
+        if (result != VK_SUCCESS || bufferView == VK_NULL_HANDLE) {
+            MGLOG_E_ONCE("ResolveStorageTexelBufferDescriptor: vkCreateBufferView failed result=%d format=%d range=%zu",
                     result, static_cast<Int>(vkFormat), static_cast<SizeT>(viewRange));
             return false;
         }
@@ -770,7 +919,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto& bindingPoint = MG_State::pGLContext->GetBufferBindingPoint(BufferTarget::ShaderStorage, frontendBinding);
         const auto& bufferObject = bindingPoint.GetBoundObject();
         if (bufferObject == nullptr) {
-            MGLOG_E("ResolveStorageBufferDescriptor: no SSBO bound at frontend binding %u for block '%s'",
+            MGLOG_E_ONCE("ResolveStorageBufferDescriptor: no SSBO bound at frontend binding %u for block '%s'",
                     frontendBinding, programObj.storageBlockNameByBinding[binding].c_str());
             return false;
         }
@@ -785,7 +934,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         BufferSlice slice{};
         if (!m_bufferManager->AcquireResidentSlice(BufferKind::ShaderStorage, bufferObject, slice) || !slice.IsValid()) {
-            MGLOG_E("ResolveStorageBufferDescriptor: failed to sync GL buffer %u for block '%s'",
+            MGLOG_E_ONCE("ResolveStorageBufferDescriptor: failed to sync GL buffer %u for block '%s'",
                     bufferObject->GetExternalIndex(), programObj.storageBlockNameByBinding[binding].c_str());
             return false;
         }
@@ -799,7 +948,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             rangeEnd = bufferSize;
         }
         if (rangeEnd <= rangeStart) {
-            MGLOG_E("ResolveStorageBufferDescriptor: empty SSBO range for block '%s'",
+            MGLOG_E_ONCE("ResolveStorageBufferDescriptor: empty SSBO range for block '%s'",
                     programObj.storageBlockNameByBinding[binding].c_str());
             return false;
         }
@@ -823,7 +972,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const Int baseLocation = programObj.samplerUniformLocationByBinding[binding];
         if (baseLocation < 0) {
-            MGLOG_E("ResolveStorageImageDescriptor: storage image binding %u has no uniform location", binding);
+            MGLOG_E_ONCE("ResolveStorageImageDescriptor: storage image binding %u has no uniform location", binding);
             return false;
         }
         // Per ELEMENT, and this is where an image array differs from a storage-block array: GL
@@ -835,26 +984,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // uniform.
         const Int location = baseLocation + static_cast<Int>(element);
         if (!program.UniformLocationsAliasSameUniform(baseLocation, location)) {
-            MGLOG_E("ResolveStorageImageDescriptor: binding %u element %u is past the end of its image array",
+            MGLOG_E_ONCE("ResolveStorageImageDescriptor: binding %u element %u is past the end of its image array",
                     binding, element);
             return false;
         }
         const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
         if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
-            MGLOG_E("ResolveStorageImageDescriptor: image unit %d out of range for binding %u",
+            MGLOG_E_ONCE("ResolveStorageImageDescriptor: image unit %d out of range for binding %u",
                     imageUnit, binding);
             return false;
         }
 
         auto& imageBinding = MG_State::pGLContext->GetImageTextureBinding(imageUnit);
         if (imageBinding.Texture == nullptr) {
-            MGLOG_E("ResolveStorageImageDescriptor: image unit %d is unbound for binding %u", imageUnit, binding);
+            MGLOG_E_ONCE("ResolveStorageImageDescriptor: image unit %d is unbound for binding %u", imageUnit, binding);
             return false;
         }
 
         const Bool ready = m_textureManager->TransitionTextureForStorageImage(commandBuffer, *imageBinding.Texture);
         if (!ready) {
-            MGLOG_E("ResolveStorageImageDescriptor: failed to transition textureId=%d for image unit %d",
+            MGLOG_E_ONCE("ResolveStorageImageDescriptor: failed to transition textureId=%d for image unit %d",
                     imageBinding.Texture->GetExternalIndex(), imageUnit);
             return false;
         }
@@ -874,7 +1023,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkFormat viewFormat = ResolveStorageImageViewFormat(
             reflectedFormat, imageBinding.Format, resource->format, useBindingFormat);
         if (viewFormat == VK_FORMAT_UNDEFINED) {
-            MGLOG_E("ResolveStorageImageDescriptor: unsupported glBindImageTexture format=0x%x "
+            MGLOG_E_ONCE("ResolveStorageImageDescriptor: unsupported glBindImageTexture format=0x%x "
                     "for binding=%u imageUnit=%d textureId=%d bindingPolicy=%s",
                     imageBinding.Format, binding, imageUnit, imageBinding.Texture->GetExternalIndex(),
                     useBindingFormat ? "true" : "false");
@@ -883,7 +1032,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkImageView view = m_textureManager->GetOrCreateStorageImageView(
             *imageBinding.Texture, mipLevel, viewFormat, imageBinding.Layered != GL_FALSE, imageBinding.Layer);
         if (view == VK_NULL_HANDLE) {
-            MGLOG_E("ResolveStorageImageDescriptor: failed to resolve storage view textureId=%d mip=%u "
+            MGLOG_E_ONCE("ResolveStorageImageDescriptor: failed to resolve storage view textureId=%d mip=%u "
                     "bindingFormat=0x%x imageFormat=%d reflectedFormat=%d selectedFormat=%d bindingPolicy=%s",
                     imageBinding.Texture->GetExternalIndex(), mipLevel, imageBinding.Format,
                     static_cast<Int>(resource->format), static_cast<Int>(reflectedFormat),
@@ -904,7 +1053,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Report that there is no fallback and let the caller decline the draw - aborting the
         // process over an unbound sampler is never the right answer.
         if (target != TextureTarget::Texture2D && target != TextureTarget::TextureRectangle) {
-            MGLOG_E("UniformManager::GetFallbackTexture: no fallback exists for target=%d",
+            MGLOG_E_ONCE("UniformManager::GetFallbackTexture: no fallback exists for target=%d",
                     static_cast<Int>(target));
             return nullptr;
         }
@@ -1080,13 +1229,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 continue;
             }
             if (binding >= programObj.samplerUniformLocationByBinding.size()) {
-                MGLOG_E("CollectStorageImageTextures: binding %u has no uniform-location mapping", binding);
+                MGLOG_E_ONCE("CollectStorageImageTextures: binding %u has no uniform-location mapping", binding);
                 return false;
             }
 
             const Int baseLocation = programObj.samplerUniformLocationByBinding[binding];
             if (baseLocation < 0) {
-                MGLOG_E("CollectStorageImageTextures: binding %u has no image uniform location", binding);
+                MGLOG_E_ONCE("CollectStorageImageTextures: binding %u has no image uniform location", binding);
                 return false;
             }
             // Per ELEMENT, for the same reason the sampled walk above is: an image ARRAY is one
@@ -1098,20 +1247,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             for (Uint32 element = 0; element < descriptorCount; ++element) {
                 const Int location = ResolveDescriptorElementLocation(program, baseLocation, element);
                 if (location < 0) {
-                    MGLOG_E("CollectStorageImageTextures: binding %u element %u is past the end of its image array",
+                    MGLOG_E_ONCE("CollectStorageImageTextures: binding %u element %u is past the end of its image array",
                             binding, element);
                     return false;
                 }
                 const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
                 if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
-                    MGLOG_E("CollectStorageImageTextures: image unit %d is invalid for binding %u element %u",
+                    MGLOG_E_ONCE("CollectStorageImageTextures: image unit %d is invalid for binding %u element %u",
                             imageUnit, binding, element);
                     return false;
                 }
 
                 auto* texture = MG_State::pGLContext->GetImageTextureBinding(imageUnit).Texture.get();
                 if (texture == nullptr) {
-                    MGLOG_E("CollectStorageImageTextures: image unit %d is unbound for binding %u element %u",
+                    MGLOG_E_ONCE("CollectStorageImageTextures: image unit %d is unbound for binding %u element %u",
                             imageUnit, binding, element);
                     return false;
                 }
@@ -1123,6 +1272,87 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
+    Bool UniformManager::SamplerOverlapsWritableImageSubresource(Int samplerBaseLevel, Int samplerMaxLevel,
+                                                                 GLint imageLevel, GLenum imageAccess) {
+        return imageAccess != GL_READ_ONLY && imageLevel >= samplerBaseLevel && imageLevel <= samplerMaxLevel;
+    }
+
+    Bool UniformManager::CollectSamplerImageFeedback(
+        const MG_State::GLState::ProgramObject& program,
+        const ProgramFactory::VkProgramObject& programObj,
+        Vector<SamplerImageFeedbackBinding>& outBindings) const {
+        outBindings.clear();
+        MOBILEGL_ASSERT(MG_State::pGLContext != nullptr,
+                        "CollectSamplerImageFeedback: GL context is null");
+        if (programObj.declinedDescriptors) return true;
+
+        for (const Uint32 samplerBinding : programObj.activeBindings) {
+            if (samplerBinding >= m_maxBindings ||
+                programObj.bindingKinds[samplerBinding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const Uint32 samplerCount = BindingDescriptorCount(programObj, samplerBinding);
+            for (Uint32 samplerElement = 0; samplerElement < samplerCount; ++samplerElement) {
+                MG_State::GLState::ITextureObject* sampledTexture = nullptr;
+                const MG_State::GLState::SamplerObject* sampledSampler = nullptr;
+                if (!ResolveSampledBinding(program, programObj, samplerBinding, samplerElement,
+                                           sampledTexture, sampledSampler) ||
+                    sampledTexture == nullptr || sampledSampler == nullptr ||
+                    MG_State::GLState::SamplesAsIncompleteTexture(sampledTexture, sampledSampler)) {
+                    // ResolveSamplerDescriptor uses a fallback in these cases, which cannot
+                    // alias the image-unit binding of the original texture.
+                    continue;
+                }
+                // Multisample source images intentionally omit TRANSFER_SRC usage. Keep their existing
+                // direct binding instead of turning otherwise valid sampler2DMS/image2DMS dispatches
+                // into failed dispatches; a correct snapshot for them needs a same-sample-count path.
+                const TextureTarget sampledTarget = sampledTexture->GetTarget();
+                if (sampledTarget == TextureTarget::Texture2DMultisample ||
+                    sampledTarget == TextureTarget::Texture2DMultisampleArray) {
+                    continue;
+                }
+                const auto& levelRange = sampledTexture->GetLevelRange();
+                Bool aliasesWritableImage = false;
+                for (const Uint32 imageBinding : programObj.activeBindings) {
+                    if (imageBinding >= m_maxBindings ||
+                        programObj.bindingKinds[imageBinding] != ProgramFactory::DescriptorBindingKind::StorageImage) {
+                        continue;
+                    }
+                    if (imageBinding >= programObj.samplerUniformLocationByBinding.size()) return false;
+                    const Int baseLocation = programObj.samplerUniformLocationByBinding[imageBinding];
+                    if (baseLocation < 0) return false;
+                    const Uint32 imageCount = BindingDescriptorCount(programObj, imageBinding);
+                    for (Uint32 imageElement = 0; imageElement < imageCount; ++imageElement) {
+                        const Int location = ResolveDescriptorElementLocation(program, baseLocation, imageElement);
+                        if (location < 0) return false;
+                        const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
+                        if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
+                            return false;
+                        }
+                        const auto& image = MG_State::pGLContext->GetImageTextureBinding(imageUnit);
+                        // A sampler view exposes all layers of its target; equal texture plus an
+                        // overlapping mip therefore aliases the writable image subresource.
+                        if (image.Texture.get() == sampledTexture &&
+                            SamplerOverlapsWritableImageSubresource(levelRange.x(), levelRange.y(),
+                                                                     image.Level, image.Access)) {
+                            aliasesWritableImage = true;
+                            break;
+                        }
+                    }
+                    if (aliasesWritableImage) break;
+                }
+                if (aliasesWritableImage) {
+                    outBindings.push_back({.samplerBinding = samplerBinding,
+                                           .samplerElement = samplerElement,
+                                           .texture = sampledTexture,
+                                           .sampler = sampledSampler,
+                                           .numericDomain = programObj.samplerNumericDomainByBinding[samplerBinding]});
+                }
+            }
+        }
+        return true;
+    }
+
     Bool UniformManager::ResolveUniformBufferPayload(const MG_State::GLState::ProgramObject& program,
                                                      const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                                      Uint32 arrayElement, UboBindResult& out) const {
@@ -1244,7 +1474,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
-    Bool UniformManager::CreateDescriptorPool(Uint32 maxSets, VkDescriptorPool& outPool) const {
+    Bool UniformManager::CreateDescriptorPool(Uint32 maxSets, Bool updateAfterBind, VkDescriptorPool& outPool) const {
         outPool = VK_NULL_HANDLE;
         if (m_device == VK_NULL_HANDLE || maxSets == 0 || m_maxBindings == 0) {
             return false;
@@ -1262,12 +1492,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const Uint64 descriptorCount64 =
             static_cast<Uint64>(maxSets) * static_cast<Uint64>(std::min(m_maxBindings, kEstimatedBindingsPerSet));
         if (descriptorCount64 > static_cast<Uint64>(std::numeric_limits<Uint32>::max())) {
-            MGLOG_E("UniformDescriptorBinder::CreateDescriptorPool failed: descriptorCount overflow");
+            MGLOG_E_ONCE("UniformDescriptorBinder::CreateDescriptorPool failed: descriptorCount overflow");
             return false;
         }
 
         const Uint32 descriptorCount = static_cast<Uint32>(descriptorCount64);
-        VkDescriptorPoolSize poolSizes[5]{};
+        VkDescriptorPoolSize poolSizes[6]{};
         poolSizes[0].type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC;
         poolSizes[0].descriptorCount = descriptorCount;
         poolSizes[1].type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
@@ -1278,6 +1508,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         poolSizes[3].descriptorCount = descriptorCount;
         poolSizes[4].type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
         poolSizes[4].descriptorCount = descriptorCount;
+        poolSizes[5].type = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
+        poolSizes[5].descriptorCount = descriptorCount;
 
         VkDescriptorPoolCreateInfo poolInfo{};
         poolInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
@@ -1285,38 +1517,43 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // (OnDescriptorSetLayoutDestroyed) so program churn recycles pool capacity.
         // The cost is on set allocation only, which happens when a layout's per-frame
         // cache grows - never on the per-draw reuse path.
-        poolInfo.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT;
+        poolInfo.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT |
+                         (updateAfterBind ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT : 0);
         poolInfo.maxSets = maxSets;
         poolInfo.poolSizeCount = static_cast<Uint32>(std::size(poolSizes));
         poolInfo.pPoolSizes = poolSizes;
 
         const VkResult result = vkCreateDescriptorPool(m_device, &poolInfo, nullptr, &outPool);
         if (result != VK_SUCCESS) {
-            MGLOG_E("UniformDescriptorBinder::CreateDescriptorPool failed: vkCreateDescriptorPool returned %d",
+            MGLOG_E_ONCE("UniformDescriptorBinder::CreateDescriptorPool failed: vkCreateDescriptorPool returned %d",
                     result);
             return false;
         }
         return true;
     }
 
-    Bool UniformManager::GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex) {
+    Bool UniformManager::GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind) {
         if (frame.descriptorPools.empty()) {
             return false;
         }
 
-        const auto& currentBucket = frame.descriptorPools[frame.activeDescriptorPoolIndex];
+        const auto matchingBucket = std::find_if(
-        const Uint32 currentMaxSets = std::max<Uint32>(1, currentBucket.maxSets);
+            frame.descriptorPools.begin(), frame.descriptorPools.end(),
+            [updateAfterBind](const DescriptorPoolBucket& candidate) { return candidate.updateAfterBind == updateAfterBind; });
+        const Uint32 currentMaxSets = matchingBucket != frame.descriptorPools.end()
+                                          ? std::max<Uint32>(1, matchingBucket->maxSets)
+                                          : m_setsPerFrame;
         const Uint32 grownMaxSets = currentMaxSets <= (std::numeric_limits<Uint32>::max() / 2) ? (currentMaxSets * 2)
                                                                                                  : currentMaxSets;
 
         VkDescriptorPool grownPool = VK_NULL_HANDLE;
-        if (!CreateDescriptorPool(grownMaxSets, grownPool)) {
+        if (!CreateDescriptorPool(grownMaxSets, updateAfterBind, grownPool)) {
-            MGLOG_E("UniformDescriptorBinder::GrowFrameDescriptorPool failed: cannot create grown pool (%u -> %u sets)",
+            MGLOG_E_ONCE("UniformDescriptorBinder::GrowFrameDescriptorPool failed: cannot create grown pool (%u -> %u sets)",
                     currentMaxSets, grownMaxSets);
             return false;
         }
 
-        frame.descriptorPools.push_back({grownPool, grownMaxSets, 0});
+        frame.descriptorPools.push_back({grownPool, grownMaxSets, 0, updateAfterBind});
         frame.activeDescriptorPoolIndex = static_cast<Uint32>(frame.descriptorPools.size() - 1);
         MGLOG_D(
             "UniformDescriptorBinder: frame %u descriptor pool exhausted, grew pool (%u -> %u sets), poolCount=%zu",
@@ -1326,14 +1563,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     VkResult UniformManager::AllocateDescriptorSetsFromActivePool(Uint32 frameIndex, const ProgramFactory::VkProgramObject& programObj, VkDescriptorSet& outDescriptorSet) {
         auto& frame = m_frames[frameIndex];
-        if (frame.activeDescriptorPoolIndex >= frame.descriptorPools.size()) {
+        const Bool updateAfterBind = programObj.usesUpdateAfterBind;
-            frame.activeDescriptorPoolIndex = 0;
+        if (frame.activeDescriptorPoolIndex >= frame.descriptorPools.size() ||
-        }
+            frame.descriptorPools[frame.activeDescriptorPoolIndex].updateAfterBind != updateAfterBind ||
-        if (frame.descriptorPools[frame.activeDescriptorPoolIndex].allocatedSets >=
+            frame.descriptorPools[frame.activeDescriptorPoolIndex].allocatedSets >=
                 frame.descriptorPools[frame.activeDescriptorPoolIndex].maxSets) {
             const auto availableBucket = std::find_if(
                 frame.descriptorPools.begin(), frame.descriptorPools.end(),
-                [](const DescriptorPoolBucket& candidate) { return candidate.allocatedSets < candidate.maxSets; });
+                [updateAfterBind](const DescriptorPoolBucket& candidate) {
+                    return candidate.updateAfterBind == updateAfterBind && candidate.allocatedSets < candidate.maxSets;
+                });
             if (availableBucket == frame.descriptorPools.end()) {
                 outDescriptorSet = VK_NULL_HANDLE;
                 return VK_ERROR_OUT_OF_POOL_MEMORY;
@@ -1369,8 +1608,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         } else {
             VkResult allocResult = AllocateDescriptorSetsFromActivePool(frameIndex, programObj, outDescriptorSet);
             if (allocResult == VK_ERROR_OUT_OF_POOL_MEMORY || allocResult == VK_ERROR_FRAGMENTED_POOL) {
-                if (!GrowFrameDescriptorPool(frame, frameIndex)) {
+                if (!GrowFrameDescriptorPool(frame, frameIndex, programObj.usesUpdateAfterBind)) {
-                    MGLOG_E("UniformDescriptorBinder::AcquireDescriptorSet failed: descriptor pool growth failed");
+                    MGLOG_E_ONCE("UniformDescriptorBinder::AcquireDescriptorSet failed: descriptor pool growth failed");
                     return allocResult;
                 }
                 allocResult = AllocateDescriptorSetsFromActivePool(frameIndex, programObj, outDescriptorSet);
@@ -1487,7 +1726,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                              Uint32 frameIndex,
                                                              VkPipelineBindPoint bindPoint,
                                                              const SamplerBindingOverride* samplerBindingOverride,
-                                                             Bool samplerDescriptorsUnchangedHint) {
+                                                             Bool samplerDescriptorsUnchangedHint,
+                                                             const Vector<SamplerBindingOverride>* samplerBindingOverrides) {
         // This program has a descriptor MobileGL could not resolve (see
         // VkProgramObject::declinedDescriptors). Refusing here is the whole of the decline: the
         // binding is still declared in the layout, so the pipeline is consistent with the shader
@@ -1501,7 +1741,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         auto& frame = m_frames[frameIndex];
         if (frame.descriptorPools.empty()) {
-            MGLOG_E("UniformDescriptorBinder::BindProgramUniformBuffers failed: frame descriptor pools are invalid");
+            MGLOG_E_ONCE("UniformDescriptorBinder::BindProgramUniformBuffers failed: frame descriptor pools are invalid");
             return false;
         }
         if (frame.activeDescriptorPoolIndex >= frame.descriptorPools.size()) {
@@ -1514,7 +1754,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // 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);
+        const Bool cacheable = samplerBindingOverride == nullptr &&
+                               (samplerBindingOverrides == nullptr || samplerBindingOverrides->empty());
         if (cacheable && samplerDescriptorsUnchangedHint && m_fastRebindMemo.valid &&
             m_fastRebindMemo.frameIndex == frameIndex &&
             m_fastRebindMemo.programLifetimeId == program.GetLifetimeId() &&
@@ -1578,6 +1819,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // is reachable wherever m_maxBindings is small (it clamps to ~16 on Adreno and Mali),
         // which is exactly where a 7-element CTS sampler array does not fit the slack.
         imageInfos.reserve(m_maxBindings + arrayDescriptorExtra);
+        // Exact, and safe only because it is: BOTH texel kinds (samplerBuffer and imageBuffer)
+        // refuse descriptor arrays at program creation, so each contributes at most one view and
+        // the total cannot exceed the binding count. The branches below take the address of
+        // back(), so making a texel kind array-capable without also giving this the surplus
+        // imageInfos gets would dangle every pTexelBufferView already recorded in `writes`.
         texelBufferViews.reserve(m_maxBindings);
         dynamicOffsets.reserve(programObj.dynamicBindings.size() + uboArrayExtra);
 
@@ -1633,7 +1879,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 VkBufferView bufferView = VK_NULL_HANDLE;
                 if (!ResolveTexelBufferDescriptor(program, programObj, binding, frameIndex, bufferView) ||
                     bufferView == VK_NULL_HANDLE) {
-                    MGLOG_E(
+                    MGLOG_E_ONCE(
                         "UniformDescriptorBinder::BindProgramUniformBuffers failed: texture buffer binding %u has no valid descriptor",
                         binding);
                     return false;
@@ -1644,6 +1890,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 write.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER;
                 write.pTexelBufferView = &texelBufferViews.back();
                 writes.push_back(write);
+            } else if (kind == ProgramFactory::DescriptorBindingKind::StorageTexelBuffer) {
+                // Shares texelBufferViews with the sampled kind above, and may do so safely for
+                // the same reason: neither kind can be an array, so each contributes exactly one
+                // element and the reserve of m_maxBindings cannot be outrun - which is what keeps
+                // the &back() below from dangling when a later binding pushes.
+                VkBufferView bufferView = VK_NULL_HANDLE;
+                if (!ResolveStorageTexelBufferDescriptor(program, programObj, binding, frameIndex, bufferView) ||
+                    bufferView == VK_NULL_HANDLE) {
+                    MGLOG_E_ONCE("UniformDescriptorBinder::BindProgramUniformBuffers failed: image buffer binding %u "
+                            "has no valid descriptor",
+                            binding);
+                    return false;
+                }
+
+                texelBufferViews.push_back(bufferView);
+                fastRebindKindsEligible = false;
+                write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
+                write.pTexelBufferView = &texelBufferViews.back();
+                writes.push_back(write);
             } else if (kind == ProgramFactory::DescriptorBindingKind::StorageBuffer) {
                 // One write per binding, but `descriptorCount` buffer infos: a GLSL block
                 // instance array occupies a single binding whose elements each come from their
@@ -1653,7 +1918,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 for (Uint32 element = 0; element < descriptorCount; ++element) {
                     VkDescriptorBufferInfo bufferInfo{};
                     if (!ResolveStorageBufferDescriptor(program, programObj, binding, element, bufferInfo)) {
-                        MGLOG_E(
+                        MGLOG_E_ONCE(
                             "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage buffer binding %u "
                             "element %u has no valid descriptor",
                             binding, element);
@@ -1680,7 +1945,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     VkDescriptorImageInfo imageInfo{};
                     if (!ResolveStorageImageDescriptor(commandBuffer, program, programObj, binding, element,
                                                        imageInfo)) {
-                        MGLOG_E(
+                        MGLOG_E_ONCE(
                             "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage image binding %u "
                             "element %u has no valid descriptor",
                             binding, element);
@@ -1714,22 +1979,32 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const SizeT firstImageInfoIndex = imageInfos.size();
                 for (Uint32 element = 0; element < descriptorCount; ++element) {
                     VkDescriptorImageInfo imageInfo{};
-                    Bool hasImage = false;
+                    const SamplerBindingOverride* overrideForElement =
-                    if (overrideThisBinding && element == 0) {
+                        overrideThisBinding && element == 0 ? samplerBindingOverride : nullptr;
-                        hasImage = ResolveSamplerDescriptorOverride(*samplerBindingOverride, imageInfo);
+                    if (overrideForElement == nullptr && samplerBindingOverrides != nullptr) {
-                    } else {
+                        const auto overrideIt = std::find_if(
-                        hasImage = ResolveSamplerDescriptor(commandBuffer, program, programObj, binding, element,
+                            samplerBindingOverrides->begin(), samplerBindingOverrides->end(),
-                                                            imageInfo, samplerDescriptorsUnchangedHint);
+                            [binding, element](const SamplerBindingOverride& candidate) {
+                                return candidate.binding == binding && candidate.element == element;
+                            });
+                        if (overrideIt != samplerBindingOverrides->end()) {
+                            overrideForElement = &*overrideIt;
                         }
+                    }
+                    const Bool hasImage = overrideForElement != nullptr
+                                              ? ResolveSamplerDescriptorOverride(*overrideForElement, imageInfo)
+                                              : ResolveSamplerDescriptor(commandBuffer, program, programObj, binding,
+                                                                         element, imageInfo,
+                                                                         samplerDescriptorsUnchangedHint);
                     if (!hasImage) {
-                        MGLOG_E(
+                        MGLOG_E_ONCE(
                             "UniformDescriptorBinder::BindProgramUniformBuffers failed: sampler binding %u element %u "
                             "has no valid texture descriptor",
                             binding, element);
                         return false;
                     }
                     if (imageInfo.sampler == VK_NULL_HANDLE || imageInfo.imageView == VK_NULL_HANDLE) {
-                        MGLOG_E(
+                        MGLOG_E_ONCE(
                             "UniformDescriptorBinder::BindProgramUniformBuffers failed: sampler binding %u element %u "
                             "has null sampler or imageView",
                             binding, element);
@@ -1803,7 +2078,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         } else {
             VkResult allocResult = AcquireDescriptorSet(frameIndex, programObj, descriptorSet);
             if (allocResult != VK_SUCCESS || descriptorSet == VK_NULL_HANDLE) {
-                MGLOG_E("UniformDescriptorBinder::BindProgramUniformBuffers failed: descriptor set acquire returned %d",
+                MGLOG_E_ONCE("UniformDescriptorBinder::BindProgramUniformBuffers failed: descriptor set acquire returned %d",
                         allocResult);
                 return false;
             }

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -26,9 +26,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     public:
         struct SamplerBindingOverride {
             Uint32 binding = 0;
+            Uint32 element = 0;
             MG_State::GLState::ITextureObject* texture = nullptr;
             const MG_State::GLState::SamplerObject* sampler = nullptr;
             VkImageView imageView = VK_NULL_HANDLE;
+            VkImageLayout imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+            Bool forceNearestFiltering = false;
+        };
+
+        struct SamplerImageFeedbackBinding {
+            Uint32 samplerBinding = 0;
+            Uint32 samplerElement = 0;
+            MG_State::GLState::ITextureObject* texture = nullptr;
+            const MG_State::GLState::SamplerObject* sampler = nullptr;
+            SamplerNumericDomain numericDomain = SamplerNumericDomain::Unknown;
         };
 
         Bool Initialize(VkDevice device, VkBufferManager* bufferManager,
@@ -79,6 +90,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool CollectStorageImageTextures(const MG_State::GLState::ProgramObject& program,
                                          const ProgramFactory::VkProgramObject& programObj,
                                          Vector<MG_State::GLState::ITextureObject*>& outTextures) const;
+        Bool CollectSamplerImageFeedback(
+            const MG_State::GLState::ProgramObject& program,
+            const ProgramFactory::VkProgramObject& programObj,
+            Vector<SamplerImageFeedbackBinding>& outBindings) const;
+        static Bool SamplerOverlapsWritableImageSubresource(Int samplerBaseLevel, Int samplerMaxLevel,
+                                                             GLint imageLevel, GLenum imageAccess);
         // 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
@@ -91,7 +108,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                        Uint32 frameIndex,
                                        VkPipelineBindPoint bindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
                                        const SamplerBindingOverride* samplerBindingOverride = nullptr,
-                                       Bool samplerDescriptorsUnchangedHint = false);
+                                       Bool samplerDescriptorsUnchangedHint = false,
+                                       const Vector<SamplerBindingOverride>* samplerBindingOverrides = nullptr);
 
         // 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 +132,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkDescriptorPool handle = VK_NULL_HANDLE;
             Uint32 maxSets = 0;
             Uint32 allocatedSets = 0;
+            Bool updateAfterBind = false;
         };
 
         // A cached descriptor set together with the pool it was allocated from, so a
@@ -175,6 +194,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool ResolveTexelBufferDescriptor(const MG_State::GLState::ProgramObject& program,
                                           const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                           Uint32 frameIndex, VkBufferView& outBufferView);
+        // GLSL `imageBuffer`: the same VkBufferView descriptor as the sampled texel buffer above,
+        // but resolved from an IMAGE unit (glBindImageTexture) rather than a texture unit, and
+        // made GPU-resident-writable because the shader may store to it. No `element` parameter:
+        // an imageBuffer ARRAY is refused at program creation, so a binding is always one
+        // descriptor (see the array gate in RemapDescriptorBindingsForVulkan).
+        Bool ResolveStorageTexelBufferDescriptor(const MG_State::GLState::ProgramObject& program,
+                                                 const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
+                                                 Uint32 frameIndex, VkBufferView& outBufferView);
         // `element` indexes a block INSTANCE array's descriptors; it is 0 for every ordinary
         // block. Each element resolves through its own GL storage block, and so its own GL
         // binding point, buffer and glBindBufferRange window.
@@ -215,8 +242,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void BindDescriptorSetDeduped(VkCommandBuffer commandBuffer, VkPipelineBindPoint bindPoint,
                                       VkPipelineLayout pipelineLayout, VkDescriptorSet descriptorSet,
                                       const Vector<Uint32>& dynamicOffsets);
-        Bool CreateDescriptorPool(Uint32 maxSets, VkDescriptorPool& outPool) const;
+        Bool CreateDescriptorPool(Uint32 maxSets, Bool updateAfterBind, VkDescriptorPool& outPool) const;
-        Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex);
+        Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind);
         VkResult AllocateDescriptorSetsFromActivePool(
             Uint32 frameIndex, const ProgramFactory::VkProgramObject& programObj, VkDescriptorSet& outDescriptorSet);
         VkResult AcquireDescriptorSet(Uint32 frameIndex,
@@ -333,8 +360,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // lifetime id, so a freed-and-reallocated sampler or texture at the same heap address
         // always gets a fresh id and misses (a raw pointer would false-hit that ABA) - so a
         // stale guess can only miss and fall through to the hash, never resolve wrong. Still
-        // reset each frame alongside the descriptor-set cache. Indexed by binding.
+        // reset each frame alongside the descriptor-set cache. Indexed by binding, but the
+        // whole-descriptor entry is additionally keyed by program lifetime: Vulkan binding
+        // numbers are layout-local and unrelated programs routinely reuse binding 0/1.
         struct SamplerResolveMemo {
+            Uint64 infoProgramLifetimeId = 0;
             Uint64 samplerLifetimeId = 0;
             Uint64 textureLifetimeId = 0;
             VkSampler sampler = VK_NULL_HANDLE;

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp

@@ -110,7 +110,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const VkFormat sourceVkFormat =
                 ToVkVertexFormat(attr.Type, attr.Size, attr.Normalized, attr.IsInteger, attr.IsBgra, attr.IsLong);
             if (sourceVkFormat == VK_FORMAT_UNDEFINED) {
-                MGLOG_E("Unsupported vertex attribute layout (location=%u, type=%s, size=%d): the array is "
+                MGLOG_E_ONCE("Unsupported vertex attribute layout (location=%u, type=%s, size=%d): the array is "
                         "enabled but cannot be mapped to a VkFormat",
                         location, MG_Util::ConvertDataTypeToString(attr.Type).c_str(), attr.Size);
                 unsupportedAttribMask |= (1u << location);
@@ -125,7 +125,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     if (fallbackFormat != VK_FORMAT_UNDEFINED && SupportsVertexBufferFormat(fallbackFormat)) {
                         vkFormat = fallbackFormat;
                         conversion = VertexStreamConversion::ScaledIntegerToFloat32;
-                        MGLOG_W("Vertex attribute location=%u format=%d lacks "
+                        MGLOG_W_ONCE("Vertex attribute location=%u format=%d lacks "
                                 "VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT; using float32 stream format=%d "
                                 "(type=%s size=%d normalized=%s integer=%s)",
                                 location, static_cast<Int>(sourceVkFormat), static_cast<Int>(vkFormat),
@@ -135,7 +135,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
 
                 if (conversion == VertexStreamConversion::None) {
-                    MGLOG_E("Unsupported Vulkan vertex format (location=%u, format=%d, type=%s, size=%d): "
+                    MGLOG_E_ONCE("Unsupported Vulkan vertex format (location=%u, format=%d, type=%s, size=%d): "
                             "VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT is unavailable and no semantic fallback exists",
                             location, static_cast<Int>(sourceVkFormat),
                             MG_Util::ConvertDataTypeToString(attr.Type).c_str(), attr.Size);
@@ -146,7 +146,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             const SizeT attribByteSize = GetAttributeByteSize(attr.Type, attr.Size, attr.IsBgra);
             if (attribByteSize == 0) {
-                MGLOG_E("Vertex attribute with unknown component size (location=%u, type=%s): the array is "
+                MGLOG_E_ONCE("Vertex attribute with unknown component size (location=%u, type=%s): the array is "
                         "enabled but cannot be sized",
                         location, MG_Util::ConvertDataTypeToString(attr.Type).c_str());
                 unsupportedAttribMask |= (1u << location);
@@ -175,7 +175,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // unless VK_EXT_legacy_vertex_attributes is available, so deinterleave this one
                 // attribute into a tightly packed transient stream without changing its format.
                 conversion = VertexStreamConversion::Repack;
-                MGLOG_W("Vertex attribute location=%u uses Vulkan-incompatible alignment "
+                MGLOG_W_ONCE("Vertex attribute location=%u uses Vulkan-incompatible alignment "
                         "(offset=%zu stride=%u required=%zu); using a tightly packed stream",
                         location, attr.Offset, sourceStride, requiredAlignment);
             }
@@ -287,8 +287,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (m_frameBoundaryCounter - it->second->lastUsedFrameBoundary > kRetireAgeBoundaries) {
                 it = m_cache.erase(it);
                 // Invalidate every VAO's state-pointer memo: the erased node's
-                // address may be reused by a future insert.
+                // address may be reused by a future insert. Advance through the
-                ++m_evictionEpoch;
+                // process-wide source so the value stays unique across factory
+                // instances (see the member comment).
+                m_evictionEpoch = ++s_evictionEpochSource;
             } else {
                 ++it;
             }

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -125,7 +125,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // construction); a memo is honored only while its recorded epoch
         // matches, so an evicted entry can never be dereferenced through a
         // stale memo.
-        Uint64 m_evictionEpoch = 1;
+        //
+        // Drawn from a process-wide source, never a per-instance counter: the VAO
+        // memos outlive this factory (they live on pGLContext's VAOs, the renderer
+        // is destroyed and recreated on EGL surface release/re-create), so a fresh
+        // factory restarting at a dead factory's epoch value would honor its
+        // dangling entry pointers. The constructor takes a value strictly greater
+        // than anything a predecessor ever stamped, so a dead factory's memo can
+        // never compare equal here - the same never-reused idiom as the lifetime ids.
+        // Single-threaded like the rest of the factory (renderer-thread only).
+        static inline Uint64 s_evictionEpochSource = 0;
+        Uint64 m_evictionEpoch = ++s_evictionEpochSource;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.cpp

@@ -23,7 +23,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
             VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
             VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT | VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
-            VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
+            // "Every usage" has to mean every usage: a buffer texture reached through an IMAGE
+            // unit takes a VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER descriptor, and the write is
+            // invalid unless the buffer was created with this bit. Nothing asked for it until
+            // imageBuffer support existed, so the omission was invisible.
+            VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
         // Appended to kPersistentBackedUsage when VK_EXT_transform_feedback is enabled
         // (see VkBufferManagerInitInfo::transformFeedbackUsageEnabled).
         constexpr VkBufferUsageFlags kTransformFeedbackUsage =
@@ -298,7 +302,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             .requiredFlags = requiredFlags,
         });
         if (!created || resource.buffer.Map() == nullptr) {
-            MGLOG_E("VkBufferManager::CreateResidentStorage failed (size=%llu)",
+            MGLOG_E_ONCE("VkBufferManager::CreateResidentStorage failed (size=%llu)",
                     static_cast<unsigned long long>(size));
             resource.buffer.Destroy();
             resource.storageSize = 0;
@@ -320,7 +324,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return false;
         }
         if (!resource.buffer.Upload(bufferObject.MappedData(), size, 0)) {
-            MGLOG_E("VkBufferManager::SwapStorageAndUploadAll: upload failed");
+            MGLOG_E_ONCE("VkBufferManager::SwapStorageAndUploadAll: upload failed");
             resource.pendingFullUpload = true;
             return false;
         }
@@ -379,6 +383,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         BumpSliceEpoch(*resource);
         // Any cached streaming slice refers to the previous contents.
         resource->transientFrameSerial = 0;
+        // Redefining the store hands any adopted mapping back to the CPU shadow
+        // (BufferObject::RedefineStorage), so a buffer that reaches here persistent-mapped
+        // is an ordinary resident one again: it needs the busy-tracking and conditional
+        // orphan below, and the next AcquirePersistentMap has to mint storage for the new
+        // store rather than hand back a mapping of the old one.
+        resource->persistentMapped = false;
         if (!resource->buffer.IsValid()) {
             return; // streaming-only resource: shadow + serial are enough
         }
@@ -399,7 +409,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         if (!resource->buffer.Upload(bufferObject.MappedData(), size, 0)) {
-            MGLOG_E("VkBufferManager::OnRespecify: in-place upload failed");
+            MGLOG_E_ONCE("VkBufferManager::OnRespecify: in-place upload failed");
             resource->pendingFullUpload = true;
         }
     }
@@ -424,7 +434,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (!IsResourceBusy(*resource)) {
             if (!resource->buffer.Upload(bufferObject.MappedData() + offset,
                                          static_cast<VkDeviceSize>(size), static_cast<VkDeviceSize>(offset))) {
-                MGLOG_E("VkBufferManager::OnSubData: host upload failed");
+                MGLOG_E_ONCE("VkBufferManager::OnSubData: host upload failed");
                 resource->pendingFullUpload = true;
             }
             return;
@@ -461,7 +471,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if ((appAccess & BufferMappingAccessBit::Unsynchronized) || !IsResourceBusy(*resource)) {
             if (!resource->buffer.Upload(bufferObject.MappedData() + offset,
                                          static_cast<VkDeviceSize>(size), static_cast<VkDeviceSize>(offset))) {
-                MGLOG_E("VkBufferManager::OnFlushMappedRange: host upload failed");
+                MGLOG_E_ONCE("VkBufferManager::OnFlushMappedRange: host upload failed");
                 resource->pendingFullUpload = true;
             }
             return;
@@ -553,7 +563,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkDeviceSize size = static_cast<VkDeviceSize>(bufferObject->GetSize());
         if (size == 0) {
-            MGLOG_E("VkBufferManager::AcquireResidentSlice failed: buffer size is zero");
+            MGLOG_E_ONCE("VkBufferManager::AcquireResidentSlice failed: buffer size is zero");
             return false;
         }
 
@@ -575,7 +585,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 return false;
             }
             if (!resource->buffer.Upload(bufferObject->MappedData(), size, 0)) {
-                MGLOG_E("VkBufferManager::AcquireResidentSlice failed: initial upload failed");
+                MGLOG_E_ONCE("VkBufferManager::AcquireResidentSlice failed: initial upload failed");
                 resource->buffer.Destroy();
                 resource->storageSize = 0;
                 resource->usageFlags = 0;
@@ -610,7 +620,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkDeviceSize size = static_cast<VkDeviceSize>(bufferObject->GetSize());
         if (size == 0) {
-            MGLOG_E("VkBufferManager::AcquireStreamedSlice failed: buffer size is zero");
+            MGLOG_E_ONCE("VkBufferManager::AcquireStreamedSlice failed: buffer size is zero");
             return false;
         }
 
@@ -708,7 +718,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         case BufferKind::Uniform:
             return VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
         case BufferKind::TextureBuffer:
-            return VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT;
+            // Both texel roles, for the same reason vertex/index carry both bits: one GL buffer
+            // texture can be read as a samplerBuffer and written as an imageBuffer, and which of
+            // the two it is only becomes known when a shader that uses it is bound - long after
+            // the resident buffer was created. A VkBufferView for a storage-texel descriptor is
+            // invalid unless the buffer was created with the storage bit, so a buffer that
+            // acquired only the uniform bit could never be given one.
+            return VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT;
         case BufferKind::ShaderStorage:
             return VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT;
         case BufferKind::Indirect:

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferObject.cpp

@@ -76,7 +76,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkResult result =
             vmaCreateBuffer(m_allocator, &bufferInfo, &allocationInfo, &m_buffer, &m_allocation, nullptr);
         if (result != VK_SUCCESS) {
-            MGLOG_E("VkBufferObject::Create failed: vmaCreateBuffer returned %d", result);
+            MGLOG_E_ONCE("VkBufferObject::Create failed: vmaCreateBuffer returned %d", result);
             m_allocator = nullptr;
             m_buffer = VK_NULL_HANDLE;
             m_allocation = nullptr;
@@ -108,7 +108,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkResult mapResult = vmaMapMemory(m_allocator, m_allocation, &m_mappedData);
         if (mapResult != VK_SUCCESS || m_mappedData == nullptr) {
-            MGLOG_E("VkBufferObject::Map failed: vmaMapMemory returned %d", mapResult);
+            MGLOG_E_ONCE("VkBufferObject::Map failed: vmaMapMemory returned %d", mapResult);
             m_mappedData = nullptr;
             return nullptr;
         }
@@ -138,14 +138,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const Bool wasMapped = IsMapped();
         void* mapped = wasMapped ? m_mappedData : Map();
         if (mapped == nullptr) {
-            MGLOG_E("VkBufferObject::Upload failed: unable to map buffer");
+            MGLOG_E_ONCE("VkBufferObject::Upload failed: unable to map buffer");
             return false;
         }
 
         Memcpy(static_cast<Uint8*>(mapped) + offset, data, static_cast<SizeT>(size));
         const VkResult flushResult = vmaFlushAllocation(m_allocator, m_allocation, offset, size);
         if (flushResult != VK_SUCCESS) {
-            MGLOG_E("VkBufferObject::Upload failed: vmaFlushAllocation returned %d", flushResult);
+            MGLOG_E_ONCE("VkBufferObject::Upload failed: vmaFlushAllocation returned %d", flushResult);
             if (!wasMapped) {
                 Unmap();
             }
@@ -170,7 +170,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkResult result = vmaInvalidateAllocation(m_allocator, m_allocation, offset, resolvedSize);
         if (result != VK_SUCCESS) {
-            MGLOG_E("VkBufferObject::Invalidate failed: vmaInvalidateAllocation returned %d", result);
+            MGLOG_E_ONCE("VkBufferObject::Invalidate failed: vmaInvalidateAllocation returned %d", result);
             return false;
         }
         return true;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkClearManager.cpp

@@ -166,7 +166,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void VkClearManager::MergeClearPayload(ClearAttachmentPayload& dst, const ClearAttachmentPayload& src) {
         dst.mask |= src.mask;
         if ((src.mask & GL_COLOR_BUFFER_BIT) != 0) {
+            // The whole colour story travels together (same rule as
+            // VkRenderPassManager::QueueRenderbufferClear): a glClearBufferiv/uiv
+            // payload carries its value in colorInt/colorUint and its branch selector
+            // in colorEncoding - dropping them here would leave the pending clear
+            // reading as an all-zero float one.
             dst.color = src.color;
+            dst.colorEncoding = src.colorEncoding;
+            dst.colorInt = src.colorInt;
+            dst.colorUint = src.colorUint;
         }
         if ((src.mask & GL_DEPTH_BUFFER_BIT) != 0) {
             dst.depth = src.depth;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.cpp

@@ -123,7 +123,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         if (!attachment.IsComplete()) {
-            MGLOG_W("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u has an incomplete texture attachment; using VK_ATTACHMENT_UNUSED",
+            MGLOG_W_ONCE("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u has an incomplete texture attachment; using VK_ATTACHMENT_UNUSED",
                     drawBufferIndex,
                     MG_Util::ConvertFramebufferAttachmentTypeToString(attachmentType).c_str(),
                     fbo.GetExternalIndex());
@@ -132,7 +132,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         auto* texture = attachment.GetTexture().get();
         if (texture == nullptr) {
-            MGLOG_W("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u resolved to a null texture; using VK_ATTACHMENT_UNUSED",
+            MGLOG_W_ONCE("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u resolved to a null texture; using VK_ATTACHMENT_UNUSED",
                     drawBufferIndex,
                     MG_Util::ConvertFramebufferAttachmentTypeToString(attachmentType).c_str(),
                     fbo.GetExternalIndex());
@@ -311,7 +311,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
         if (!TryResolveSampleCountFlagBits(renderbuffer->GetSamples(), sampleCount)) {
-            MGLOG_E("GetOrCreateRenderbufferResource: unsupported renderbuffer sample count %d for renderbuffer %u",
+            MGLOG_E_ONCE("GetOrCreateRenderbufferResource: unsupported renderbuffer sample count %d for renderbuffer %u",
                     renderbuffer->GetSamples(),
                     renderbuffer->GetExternalIndex());
             return nullptr;
@@ -457,7 +457,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_physicalDevice, format, imageInfo.imageType, imageInfo.tiling, imageInfo.usage, imageInfo.flags,
             &imageFormatProperties);
         if (imageFormatResult != VK_SUCCESS || (imageFormatProperties.sampleCounts & sampleCount) == 0) {
-            MGLOG_E("GetOrCreateRenderbufferResource: unsupported renderbuffer format=%d samples=%d for renderbuffer %u",
+            MGLOG_E_ONCE("GetOrCreateRenderbufferResource: unsupported renderbuffer format=%d samples=%d for renderbuffer %u",
                     static_cast<Int>(format),
                     static_cast<Int>(sampleCount),
                     renderbuffer->GetExternalIndex());
@@ -831,6 +831,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // recreated since (texture + renderbuffer image epochs), and no pending clear (which alters
         // load ops). Any of these differing forces the full recompute below. Portable to VK 1.1.
         if (activeRenderPass != nullptr && m_rpFastValid && m_rpFastFbo == &fbo &&
+            m_rpFastFboLifetimeId == fbo.GetLifetimeId() &&
             m_rpFastFboVersion == fbo.GetObjectVersion() && m_rpFastSwapchainIndex == swapchainImageIndex &&
             m_rpFastTexEpoch == m_textureManager.GetTextureImageEpoch() &&
             m_rpFastRbEpoch == m_renderbufferImageEpoch &&
@@ -855,6 +856,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // epochs AFTER ComputeHash: its attachment SyncTexture can create an image (bump the epoch).
             m_rpFastValid = true;
             m_rpFastFbo = &fbo;
+            m_rpFastFboLifetimeId = fbo.GetLifetimeId();
             m_rpFastFboVersion = fbo.GetObjectVersion();
             m_rpFastSwapchainIndex = swapchainImageIndex;
             m_rpFastTexEpoch = m_textureManager.GetTextureImageEpoch();
@@ -929,7 +931,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     const auto& renderbuffer = rbAtt.GetRenderbuffer();
                     auto* rbResource = GetOrCreateRenderbufferResource(renderbuffer);
                     if (rbResource == nullptr || (rbResource->aspect & VK_IMAGE_ASPECT_COLOR_BIT) == 0) {
-                        MGLOG_E("GetOrCreateRenderPass: draw buffer slot %u on FBO %u has an unsupported color "
+                        MGLOG_E_ONCE("GetOrCreateRenderPass: draw buffer slot %u on FBO %u has an unsupported color "
                                 "renderbuffer %u; using VK_ATTACHMENT_UNUSED",
                                 i, fbo.GetExternalIndex(), renderbuffer->GetExternalIndex());
                         continue;
@@ -1105,7 +1107,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     adoptRenderPassSampleCount(attachmentSampleCount, "color", texture->GetExternalIndex());
 
                     if (!hasClear && trackedColorLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
-                        MGLOG_W("GetOrCreateRenderPass: color attachment textureId=%d starts with undefined layout and no clear; "
+                        MGLOG_W_ONCE("GetOrCreateRenderPass: color attachment textureId=%d starts with undefined layout and no clear; "
                                 "using LOAD_OP_DONT_CARE",
                                 texture->GetExternalIndex());
                         desc.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
@@ -1161,7 +1163,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             isUsableDepthStencilAttachment(depthAtt) && isUsableDepthStencilAttachment(stencilAtt) &&
             !sameDepthStencilAttachmentObject(depthAtt, stencilAtt);
         if (hasDistinctDepthAndStencilAttachments) {
-            MGLOG_E("GetOrCreateRenderPass: separate depth/stencil attachments are not supported yet; using the depth attachment and ignoring the standalone stencil attachment for framebuffer %u",
+            MGLOG_E_ONCE("GetOrCreateRenderPass: separate depth/stencil attachments are not supported yet; using the depth attachment and ignoring the standalone stencil attachment for framebuffer %u",
                     fbo.GetExternalIndex());
         }
         if (selectedDepthStencilAttachment != nullptr) {
@@ -1223,7 +1225,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             depthAttachmentDescription.finalLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
             depthAttachmentDescription.initialLayout = loadInfo.initialLayout;
             if (trackedDepthLayout == VK_IMAGE_LAYOUT_UNDEFINED && (!clearDepth || !clearStencil)) {
-                MGLOG_W("GetOrCreateRenderPass: depth/stencil attachment id=%d starts with undefined layout "
+                MGLOG_W_ONCE("GetOrCreateRenderPass: depth/stencil attachment id=%d starts with undefined layout "
                         "and partial/no clear; using DONT_CARE for uncleared aspects",
                         depthAttachmentId);
             }
@@ -1507,7 +1509,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             ClearAttachmentPayload clearPayload{};
             SharedPtr<MG_State::GLState::ITextureObject> liveTexture;
             if (pending.hasInlinePayload) {
+                // The inline payload was snapshotted when the entry was CREATED, but the
+                // clear VALUE is not part of the entry's hash - a cache hit with a newer
+                // glClear would replay the creation-time value and drop the new one (the
+                // texture path below is immune because it re-reads the live payload).
+                // Same defense as ClearAttachmentsOnActiveRenderPass: prefer the live
+                // pending clear, fall back to the snapshot only when none is queued.
+                if (s_renderPassManager != nullptr &&
+                    s_renderPassManager->GetPendingRenderbufferClear(pending.renderbuffer, clearPayload)) {
+                    if ((clearPayload.mask & GL_COLOR_BUFFER_BIT) != 0 && pending.renderbuffer != nullptr &&
+                        MG_Util::GetBaseInternalFormatComponentCount(pending.renderbuffer->GetInternalFormat()) ==
+                            3) {
+                        // RGB renderbuffers are backed by an RGBA image; the missing alpha reads as 1.
+                        ForceOpaqueClearAlpha(clearPayload);
+                    }
+                } else {
                     clearPayload = pending.inlinePayload;
+                }
             } else {
                 if (pending.key.texture == nullptr ||
                     !s_clearManager->GetPendingClear(pending.key, clearPayload, liveTexture)) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.h

@@ -289,6 +289,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // or a pending clear. Portable to Vulkan 1.1 (no dynamic_rendering / imageless FB needed).
         Bool m_rpFastValid = false;
         const MG_State::GLState::FramebufferObject* m_rpFastFbo = nullptr;
+        // The FBO's never-reused lifetime id joins the raw pointer + Uint16 version:
+        // a deleted FBO reallocated at the same address whose fresh setup performed
+        // the same number of version bumps would otherwise compare equal (both count
+        // from 0), serving the dead framebuffer's pass to the new object.
+        Uint64 m_rpFastFboLifetimeId = 0;
         Uint16 m_rpFastFboVersion = 0;
         Uint32 m_rpFastSwapchainIndex = 0;
         Uint64 m_rpFastTexEpoch = 0;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -300,8 +300,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels,
+            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels);
-            newResource.arrayLayers);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare destination image");
 
         VkImageLayout srcTrackedLayout = oldResource.layout;
@@ -311,8 +310,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, oldResource.image, srcTrackedLayout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels,
+            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels);
-            oldResource.arrayLayers);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare source image");
 
         Vector<VkImageCopy> copyRegions;
@@ -344,8 +342,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, oldResource.layout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask,
-            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels,
+            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels);
-            newResource.arrayLayers);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to restore destination layout");
 
         VK_VERIFY(vkEndCommandBuffer(commandBuffer), "vkEndCommandBuffer(texture preserve)");
@@ -950,7 +947,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
         const VkComponentMapping sampledComponents = ResolveSampledViewComponents(texture, formatInfo);
-        const VkImageAspectFlags sampledAspect = ResolveSampledImageViewAspectMask(resource->aspect);
+        const VkImageAspectFlags sampledAspect =
+            ResolveSampledImageViewAspectMask(resource->aspect, texture.GetDepthStencilTextureMode());
         perMipSampledView = CreateImageView(resource->image, resource->format, sampledAspect, resource->viewType,
                                             mipLevel, 1, 0, resource->arrayLayers, &sampledComponents);
         if (perMipSampledView == VK_NULL_HANDLE) {
@@ -974,13 +972,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return resource->sampledView;
         }
         if (!AreSampledImageViewFormatsCompatible(resource->format, format)) {
-            MGLOG_E("%s: incompatible sampled image view format=%d for textureId=%d imageFormat=%d",
+            MGLOG_E_ONCE("%s: incompatible sampled image view format=%d for textureId=%d imageFormat=%d",
                     __func__, static_cast<Int>(format), texture.GetExternalIndex(),
                     static_cast<Int>(resource->format));
             return VK_NULL_HANDLE;
         }
         if ((resource->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
-            MGLOG_E("%s: textureId=%d needs mutable image format=%d for sampled view format=%d",
+            MGLOG_E_ONCE("%s: textureId=%d needs mutable image format=%d for sampled view format=%d",
                     __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1000,7 +998,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkFormatProperties formatProperties{};
         vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
         if ((formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT) == 0) {
-            MGLOG_E("%s: sampled image view format=%d lacks VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT "
+            MGLOG_E_ONCE("%s: sampled image view format=%d lacks VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT "
                     "for textureId=%d (available=0x%x)",
                     __func__, static_cast<Int>(format), texture.GetExternalIndex(),
                     static_cast<Uint32>(formatProperties.optimalTilingFeatures));
@@ -1014,7 +1012,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource->sampledBaseMipLevel, resource->sampledLevelCount, 0, resource->arrayLayers,
             &sampledComponents, VK_IMAGE_USAGE_SAMPLED_BIT);
         if (view == VK_NULL_HANDLE) {
-            MGLOG_E("%s: failed to create sampled image view textureId=%d imageFormat=%d viewFormat=%d",
+            MGLOG_E_ONCE("%s: failed to create sampled image view textureId=%d imageFormat=%d viewFormat=%d",
                     __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1042,14 +1040,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             format = resource->format;
         }
         if (!AreStorageImageViewFormatsCompatible(resource->format, format)) {
-            MGLOG_E("%s: incompatible storage image view format=%d for textureId=%d imageFormat=%d",
+            MGLOG_E_ONCE("%s: incompatible storage image view format=%d for textureId=%d imageFormat=%d",
                     __func__, static_cast<Int>(format), texture.GetExternalIndex(),
                     static_cast<Int>(resource->format));
             return VK_NULL_HANDLE;
         }
         if (format != resource->format &&
             (resource->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
-            MGLOG_E("%s: textureId=%d needs mutable image format=%d for storage view format=%d",
+            MGLOG_E_ONCE("%s: textureId=%d needs mutable image format=%d for storage view format=%d",
                     __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1069,7 +1067,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 viewType = VK_IMAGE_VIEW_TYPE_2D;
                 break;
             case VK_IMAGE_VIEW_TYPE_3D:
-                MGLOG_E("%s: non-layered 3D storage views are unsupported for textureId=%d",
+                MGLOG_E_ONCE("%s: non-layered 3D storage views are unsupported for textureId=%d",
                         __func__, texture.GetExternalIndex());
                 return VK_NULL_HANDLE;
             default:
@@ -1078,7 +1076,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             if (viewType != resource->viewType) {
                 if (layer < 0 || static_cast<Uint32>(layer) >= resource->arrayLayers) {
-                    MGLOG_E("%s: storage image layer=%d is out of range for textureId=%d arrayLayers=%u",
+                    MGLOG_E_ONCE("%s: storage image layer=%d is out of range for textureId=%d arrayLayers=%u",
                             __func__, layer, texture.GetExternalIndex(), resource->arrayLayers);
                     return VK_NULL_HANDLE;
                 }
@@ -1113,7 +1111,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkFormatProperties formatProperties{};
         vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
         if ((formatProperties.optimalTilingFeatures & requiredFormatFeatures) != requiredFormatFeatures) {
-            MGLOG_E("%s: storage image view format=%d lacks required features=0x%x for textureId=%d "
+            MGLOG_E_ONCE("%s: storage image view format=%d lacks required features=0x%x for textureId=%d "
                     "(available=0x%x)",
                     __func__, static_cast<Int>(format), static_cast<Uint32>(requiredFormatFeatures),
                     texture.GetExternalIndex(), static_cast<Uint32>(formatProperties.optimalTilingFeatures));
@@ -1124,7 +1122,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  mipLevel, 1, baseArrayLayer, layerCount, nullptr,
                                                  VK_IMAGE_USAGE_STORAGE_BIT);
         if (view == VK_NULL_HANDLE) {
-            MGLOG_E("%s: failed to create storage image view for textureId=%d mip=%u imageFormat=%d viewFormat=%d",
+            MGLOG_E_ONCE("%s: failed to create storage image view for textureId=%d mip=%u imageFormat=%d viewFormat=%d",
                     __func__, texture.GetExternalIndex(), mipLevel, static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1190,7 +1188,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool lowerTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, lowerMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, 0, writtenMipLevel, resource.arrayLayers);
+                    resource.aspect, 0, writtenMipLevel);
                 MOBILEGL_ASSERT(lowerTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition lower mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1202,8 +1200,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool upperTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, upperMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel,
+                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel);
-                    resource.arrayLayers);
                 MOBILEGL_ASSERT(upperTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition upper mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1222,7 +1219,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return true;
         }
         if (resource->layout == VK_IMAGE_LAYOUT_UNDEFINED) {
-            MGLOG_W("TransitionTextureForSampling: textureId=%d is still in VK_IMAGE_LAYOUT_UNDEFINED before sampling",
+            MGLOG_W_ONCE("TransitionTextureForSampling: textureId=%d is still in VK_IMAGE_LAYOUT_UNDEFINED before sampling",
                     texture.GetExternalIndex());
         }
 
@@ -1256,8 +1253,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const Bool ok = TransitionImageLayout(commandBuffer, resource->image, resource->layout, targetLayout, srcStageMask,
                                               s_sampledReadStages, srcAccessMask,
-                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels,
+                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
-                                              resource->arrayLayers);
         MOBILEGL_ASSERT(ok, "TransitionTextureForSampling: transition failed for textureId=%d", texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
         StampResourceRecordingUse(*resource);
@@ -1287,7 +1283,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               VK_IMAGE_LAYOUT_GENERAL, srcStageMask,
                                               VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, srcAccessMask,
                                               VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
-                                              resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+                                              resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionTextureForStorageImage: transition failed for textureId=%d",
                         texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
@@ -1295,6 +1291,158 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ok;
     }
 
+    Bool VkTextureManager::SnapshotTextureForSampling(VkCommandBuffer commandBuffer,
+                                                      MG_State::GLState::ITextureObject& texture,
+                                                      SamplerNumericDomain numericDomain,
+                                                      VkPipelineStageFlags consumerShaderStageMask,
+                                                      SampledTextureSnapshot& outSnapshot) {
+        outSnapshot = {};
+        TextureResource* source = SyncTextureAndGetDescriptor(texture);
+        if (source == nullptr || source->image == VK_NULL_HANDLE || source->sampleCount != VK_SAMPLE_COUNT_1_BIT ||
+            source->sampledLevelCount == 0) {
+            return false;
+        }
+
+        const VkFormat sampledFormat = ResolveSampledImageViewFormat(source->format, numericDomain);
+        if (sampledFormat == VK_FORMAT_UNDEFINED ||
+            !AreSampledImageViewFormatsCompatible(source->format, sampledFormat)) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: textureId=%d cannot create sampled view format=%d from image format=%d",
+                         texture.GetExternalIndex(), static_cast<Int>(sampledFormat), static_cast<Int>(source->format));
+            return false;
+        }
+        if (sampledFormat != source->format &&
+            (source->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: textureId=%d needs unavailable mutable image format=%d for sampled view=%d",
+                         texture.GetExternalIndex(), static_cast<Int>(source->format), static_cast<Int>(sampledFormat));
+            return false;
+        }
+
+        VkImageType imageType = VK_IMAGE_TYPE_2D;
+        switch (source->viewType) {
+        case VK_IMAGE_VIEW_TYPE_1D:
+        case VK_IMAGE_VIEW_TYPE_1D_ARRAY:
+            imageType = VK_IMAGE_TYPE_1D;
+            break;
+        case VK_IMAGE_VIEW_TYPE_3D:
+            imageType = VK_IMAGE_TYPE_3D;
+            break;
+        default:
+            break;
+        }
+
+        TextureResource snapshot{};
+        VkImageCreateInfo imageInfo{};
+        imageInfo.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
+        imageInfo.flags = source->imageCreateFlags;
+        imageInfo.imageType = imageType;
+        imageInfo.extent = {source->extent.width, source->extent.height, source->depth};
+        imageInfo.mipLevels = source->mipLevels;
+        imageInfo.arrayLayers = source->arrayLayers;
+        imageInfo.format = source->format;
+        imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
+        imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+        imageInfo.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
+        imageInfo.samples = VK_SAMPLE_COUNT_1_BIT;
+        imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+
+        // Keep the temporary's view-format list just as narrow as the source's sampler use. This
+        // has no storage-image usage, so unlike an app image binding the exact list is knowable.
+        Vector<VkFormat> viewFormats;
+        VkImageFormatListCreateInfo formatListInfo{};
+        if (m_imageFormatListSupported && (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+            viewFormats.push_back(source->format);
+            if (sampledFormat != source->format) {
+                viewFormats.push_back(sampledFormat);
+            }
+            formatListInfo.sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO;
+            formatListInfo.viewFormatCount = static_cast<Uint32>(viewFormats.size());
+            formatListInfo.pViewFormats = viewFormats.data();
+            imageInfo.pNext = &formatListInfo;
+        }
+
+        VmaAllocationCreateInfo allocationInfo{};
+        allocationInfo.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
+        allocationInfo.requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+        const VkResult createResult =
+            vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &snapshot.image, &snapshot.allocation, nullptr);
+        if (createResult != VK_SUCCESS) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: vmaCreateImage failed result=%d textureId=%d", createResult,
+                         texture.GetExternalIndex());
+            return false;
+        }
+
+        snapshot.extent = source->extent;
+        snapshot.depth = source->depth;
+        snapshot.arrayLayers = source->arrayLayers;
+        snapshot.mipLevels = source->mipLevels;
+        snapshot.sampledBaseMipLevel = source->sampledBaseMipLevel;
+        snapshot.sampledLevelCount = source->sampledLevelCount;
+        snapshot.format = source->format;
+        snapshot.aspect = source->aspect;
+        snapshot.viewType = source->viewType;
+        snapshot.sampleCount = VK_SAMPLE_COUNT_1_BIT;
+        snapshot.imageCreateFlags = imageInfo.flags;
+        snapshot.usageFlags = imageInfo.usage;
+
+        const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
+        const VkComponentMapping sampledComponents = ResolveSampledViewComponents(texture, formatInfo);
+        const VkImageAspectFlags sampledAspect =
+            ResolveSampledImageViewAspectMask(snapshot.aspect, texture.GetDepthStencilTextureMode());
+        snapshot.sampledView = CreateImageView(snapshot.image, sampledFormat, sampledAspect, snapshot.viewType,
+                                               snapshot.sampledBaseMipLevel, snapshot.sampledLevelCount, 0,
+                                               snapshot.arrayLayers, &sampledComponents);
+        if (snapshot.sampledView == VK_NULL_HANDLE) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: failed to create sampled view textureId=%d", texture.GetExternalIndex());
+            return false;
+        }
+
+        VkPipelineStageFlags sourceStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+        VkAccessFlags sourceAccessMask = 0;
+        const VkImageLayout sourceLayout = source->layout;
+        GetImageTransitionSourceState(sourceLayout, sourceStageMask, sourceAccessMask);
+        if (!TransitionImageLayout(commandBuffer, source->image, source->layout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                                   sourceStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT, sourceAccessMask,
+                                   VK_ACCESS_TRANSFER_READ_BIT, source->aspect, 0, source->mipLevels) ||
+            !TransitionImageLayout(commandBuffer, snapshot.image, snapshot.layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                                   VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
+                                   VK_ACCESS_TRANSFER_WRITE_BIT, snapshot.aspect, snapshot.sampledBaseMipLevel,
+                                   snapshot.sampledLevelCount)) {
+            return false;
+        }
+
+        Vector<VkImageCopy> copyRegions;
+        copyRegions.reserve(snapshot.sampledLevelCount);
+        for (Uint32 level = snapshot.sampledBaseMipLevel;
+             level < snapshot.sampledBaseMipLevel + snapshot.sampledLevelCount; ++level) {
+            VkImageCopy copy{};
+            copy.srcSubresource = {source->aspect, level, 0, source->arrayLayers};
+            copy.dstSubresource = {snapshot.aspect, level, 0, snapshot.arrayLayers};
+            copy.extent = {std::max(source->extent.width >> level, 1u),
+                           std::max(source->extent.height >> level, 1u),
+                           std::max(source->depth >> level, 1u)};
+            copyRegions.push_back(copy);
+        }
+        vkCmdCopyImage(commandBuffer, source->image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, snapshot.image,
+                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, static_cast<Uint32>(copyRegions.size()), copyRegions.data());
+
+        if (!TransitionImageLayout(commandBuffer, snapshot.image, snapshot.layout,
+                                   ResolveSampledReadOnlyLayout(snapshot.aspect), VK_PIPELINE_STAGE_TRANSFER_BIT,
+                                   consumerShaderStageMask, VK_ACCESS_TRANSFER_WRITE_BIT,
+                                   VK_ACCESS_SHADER_READ_BIT, snapshot.aspect, snapshot.sampledBaseMipLevel,
+                                   snapshot.sampledLevelCount) ||
+            !TransitionImageLayout(commandBuffer, source->image, source->layout, sourceLayout,
+                                   VK_PIPELINE_STAGE_TRANSFER_BIT, consumerShaderStageMask,
+                                   VK_ACCESS_TRANSFER_READ_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+                                   source->aspect, 0, source->mipLevels)) {
+            return false;
+        }
+
+        StampResourceRecordingUse(*source);
+        outSnapshot = {.imageView = snapshot.sampledView, .layout = snapshot.layout};
+        DeferResourceRelease(Move(snapshot));
+        return true;
+    }
+
     void VkTextureManager::MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture) {
         m_storageImageTextures.insert(MakeTextureIdentity(&texture));
     }
@@ -1354,8 +1502,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  VkImageLayout& trackedLayout, VkImageLayout newLayout,
                                                  VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask,
                                                  VkAccessFlags srcAccessMask, VkAccessFlags dstAccessMask,
-                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel, Uint32 levelCount,
+                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel,
-                                                 Uint32 layerCount) {
+                                                 Uint32 levelCount) {
         MOBILEGL_ASSERT(image != VK_NULL_HANDLE, "TransitionImageLayout: m_image == VK_NULL_HANDLE");
         MOBILEGL_ASSERT(!((dstAccessMask & VK_ACCESS_TRANSFER_READ_BIT) != 0 &&
                           (dstStageMask & VK_PIPELINE_STAGE_TRANSFER_BIT) == 0),
@@ -1380,7 +1528,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         barrier.subresourceRange.baseMipLevel = baseMipLevel;
         barrier.subresourceRange.levelCount = levelCount;
         barrier.subresourceRange.baseArrayLayer = 0;
-        barrier.subresourceRange.layerCount = layerCount;
+        // Every layer, always - see the declaration for why layout tracking leaves no other
+        // correct answer. VK_REMAINING_ARRAY_LAYERS rather than the image's own `arrayLayers`
+        // because those are not the same number for a 3D image: MobileGL creates 3D images
+        // 2D_ARRAY_COMPATIBLE and their arrayLayers is 1, which today Vulkan reads as "all depth
+        // slices" but will read as "depth slice 0" once VK_KHR_maintenance9 is enabled. The
+        // validation layer warns about that literal 1 by name.
+        barrier.subresourceRange.layerCount = VK_REMAINING_ARRAY_LAYERS;
         vkCmdPipelineBarrier(commandBuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr, 1, &barrier);
 
         trackedLayout = newLayout;
@@ -1573,7 +1727,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // targets this manager has no Vulkan image shape for yet (cube map arrays above all).
             // Declining the sync leaves the texture unbacked - wrong, but recoverable - where an
             // assertion would take the whole process down instead.
-            MGLOG_W("SyncTextureResource: unsupported uploadTarget=%s textureTarget=%s textureId=%d size=(%d,%d,%d) "
+            MGLOG_W_ONCE("SyncTextureResource: unsupported uploadTarget=%s textureTarget=%s textureId=%d size=(%d,%d,%d) "
                     "mipLevels=%u vkViewType=%d",
                     MG_Util::ConvertTextureUploadTargetToString(uploadTarget).c_str(),
                     MG_Util::ConvertTextureTargetToString(texture.GetTarget()).c_str(), texture.GetExternalIndex(),
@@ -1802,7 +1956,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // Losing reinterpreted views only degrades the formatless-image feature for
                 // this texture; failing creation would lose the texture entirely, so retry
                 // as a plain immutable-format image.
-                MGLOG_W("%s: mutable image format=%d is unsupported for textureId=%d; creating "
+                MGLOG_W_ONCE("%s: mutable image format=%d is unsupported for textureId=%d; creating "
                         "without VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT (format reinterpretation "
                         "will be unavailable for it)",
                         __func__, static_cast<Int>(format), texture.GetExternalIndex());
@@ -1820,7 +1974,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // Losing 2D-array compatibility only costs per-slice framebuffer attachment for this
                 // format; failing creation would lose the texture entirely. Remembered so later syncs
                 // neither reprobe nor flag-mismatch against this image and recreate it.
-                MGLOG_W("%s: VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT is unsupported for format=%d "
+                MGLOG_W_ONCE("%s: VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT is unsupported for format=%d "
                         "textureId=%d; creating without it (per-slice framebuffer attachment will be "
                         "unavailable for it)",
                         __func__, static_cast<Int>(format), texture.GetExternalIndex());
@@ -1839,6 +1993,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         texture.GetExternalIndex(),
                         MG_Util::ConvertTextureUploadTargetToString(uploadTarget).c_str(),
                         static_cast<Int>(format), static_cast<Uint32>(imageInfo.usage));
+                // The preserved image was written by GPU work that may still be in flight
+                // (preserve requires layout != UNDEFINED); park it on the deferred ring
+                // like every other destruction path instead of letting the unique_ptr
+                // destroy it synchronously under the GPU.
+                if (preservedResource) {
+                    DeferResourceRelease(Move(*preservedResource));
+                }
                 return false;
             }
         }
@@ -1852,13 +2013,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkResult createImageResult =
             vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &resource.image, &resource.allocation, nullptr);
         if (createImageResult != VK_SUCCESS) {
-            MGLOG_F("SyncTextureResource: vmaCreateImage failed (%d) textureId=%d extent=%ux%u depth=%u layers=%u "
+            // E_ONCE, not F: the comment above says it - this is a soft failure the caller
+            // recovers from, and it re-fires on every sync of every texture the driver refuses.
+            MGLOG_E_ONCE("SyncTextureResource: vmaCreateImage failed (%d) textureId=%d extent=%ux%u depth=%u layers=%u "
                     "mips=%u samples=%d format=%d",
                     createImageResult, texture.GetExternalIndex(), imageInfo.extent.width, imageInfo.extent.height,
                     imageInfo.extent.depth, imageInfo.arrayLayers, imageInfo.mipLevels,
                     static_cast<Int>(imageInfo.samples), static_cast<Int>(imageInfo.format));
             resource.image = VK_NULL_HANDLE;
             resource.allocation = nullptr;
+            // Same as the probe failure above: the preserved live image must go through
+            // the deferred ring, never a synchronous destructor while frames that
+            // reference it are still in flight.
+            if (preservedResource) {
+                DeferResourceRelease(Move(*preservedResource));
+            }
             return false;
         }
         ++m_textureImageEpoch; // a new attachment image invalidates cached render passes
@@ -2238,7 +2407,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (resource.fullView == VK_NULL_HANDLE) {
             return false;
         }
-        const VkImageAspectFlags sampledAspect = ResolveSampledImageViewAspectMask(resource.aspect);
+        const VkImageAspectFlags sampledAspect =
+            ResolveSampledImageViewAspectMask(resource.aspect, texture.GetDepthStencilTextureMode());
         resource.sampledView = CreateImageView(resource.image, resource.format, sampledAspect, resource.viewType,
                                                baseMipLevel, levelCount, 0, resource.arrayLayers, &sampledComponents);
         if (resource.sampledView == VK_NULL_HANDLE) {
@@ -2424,7 +2594,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const Bool srcIsD24S8 = outResource.format == VK_FORMAT_D24_UNORM_S8_UINT;
             const Bool srcIsD32FS8 = outResource.format == VK_FORMAT_D32_SFLOAT_S8_UINT;
             if (!srcIsD24S8 && !srcIsD32FS8) {
-                MGLOG_E("UploadDirtyMipLevels: unsupported combined depth-stencil format %d for textureId=%d",
+                MGLOG_E_ONCE("UploadDirtyMipLevels: unsupported combined depth-stencil format %d for textureId=%d",
                         static_cast<Int>(outResource.format), mipmapTexture.GetExternalIndex());
                 for (const auto& item : uploadItems) {
                     mipmapTexture.MarkStorageDirty(item.target, item.level, false);
@@ -2601,7 +2771,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                         VK_PIPELINE_STAGE_TRANSFER_BIT,
                                         uploadSrcAccessMask,
                                         VK_ACCESS_TRANSFER_WRITE_BIT,
-                                        aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                        aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL failed");
 
         // Array textures keep their GL "depth" in VkImage array layers, so the
@@ -2705,7 +2875,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                    s_sampledReadStages,
                                    VK_ACCESS_TRANSFER_WRITE_BIT,
                                    VK_ACCESS_SHADER_READ_BIT,
-                                   aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                   aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to sampled read-only layout failed");
         outResource.layout = finalLayout;
 
@@ -2860,10 +3030,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
-    VkImageAspectFlags VkTextureManager::ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect) {
+    VkImageAspectFlags VkTextureManager::ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect,
+                                                                           GLenum depthStencilTextureMode) {
         if ((imageAspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0) {
             return VK_IMAGE_ASPECT_COLOR_BIT;
         }
+        // A sampled view of a combined depth/stencil image may name exactly one aspect
+        // (VUID-VkDescriptorImageInfo-imageView-01976), and GL_DEPTH_STENCIL_TEXTURE_MODE is
+        // what picks it - the whole content of GL_ARB_stencil_texturing. Depth stays the
+        // default, so nothing that never sets the mode changes shape. The texture's params
+        // version moves with the mode, which is what makes the cached views be rebuilt.
+        if (depthStencilTextureMode == GL_STENCIL_INDEX && (imageAspect & VK_IMAGE_ASPECT_STENCIL_BIT) != 0) {
+            return VK_IMAGE_ASPECT_STENCIL_BIT;
+        }
         if ((imageAspect & VK_IMAGE_ASPECT_DEPTH_BIT) != 0) {
             return VK_IMAGE_ASPECT_DEPTH_BIT;
         }

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -310,6 +310,11 @@ public:
         static inline VmaAllocator s_allocator = VK_NULL_HANDLE;
     };
 
+    struct SampledTextureSnapshot {
+        VkImageView imageView = VK_NULL_HANDLE;
+        VkImageLayout layout = VK_IMAGE_LAYOUT_UNDEFINED;
+    };
+
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
     void BeginFrame(Uint32 frameIndex);
@@ -343,6 +348,13 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+    // Copies the complete sampler-visible mip range into a transient sampled image. The source is
+    // restored to its prior layout, so image-store descriptors continue to name the original image.
+    // The transient ownership is tied to the current frame slot and is safe through its submission.
+    Bool SnapshotTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture,
+                                    SamplerNumericDomain numericDomain,
+                                    VkPipelineStageFlags consumerShaderStageMask,
+                                    SampledTextureSnapshot& outSnapshot);
 
     // Recording-generation bookkeeping for the pre-pass command stream. The
     // generation advances every time the frame command buffer (re)begins
@@ -379,17 +391,33 @@ public:
     // true - a false positive merely ends the render pass, a false negative would skip a barrier.
     Bool NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const;
 
-    static VkImageAspectFlags ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect);
+    // `depthStencilTextureMode` is the texture's GL_DEPTH_STENCIL_TEXTURE_MODE; it only decides
+    // anything for an image that carries both aspects. Defaulted so the call sites that have no
+    // texture in hand keep the depth-aspect answer they have always given.
+    static VkImageAspectFlags ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect,
+                                                                GLenum depthStencilTextureMode = GL_DEPTH_COMPONENT);
     static VkFormat ResolveSampledImageViewFormat(VkFormat imageFormat, SamplerNumericDomain numericDomain);
     static Bool AreSampledImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
     static Bool AreStorageImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
 
+    // Moves `image` to `newLayout` and writes the new layout back through `trackedLayout`.
+    //
+    // The barrier covers EVERY array layer of the image, and there is deliberately no layer
+    // parameter to say otherwise: layout here is tracked per IMAGE (one `TextureResource::layout`,
+    // or one caller-owned variable), so a barrier narrower than the image would leave the layers it
+    // skipped in the old layout while the tracker claims they moved. Every transfer against a
+    // framebuffer attachment above layer 0 - glReadPixels, glBlitFramebuffer, glCopyTexSubImage,
+    // glCopyImageSubData - then ran its copy on a layer no barrier had transitioned.
+    //
+    // The mip range IS a parameter, because mip levels really are transitioned piecewise (see
+    // UpdateTrackedImageLayoutAfterAttachmentWrite and the mipmap generation loops): those callers
+    // move the complement of the level they wrote so the whole image converges on one layout again.
+    // Nothing does, or can, do that per layer.
     static Bool TransitionImageLayout(VkCommandBuffer commandBuffer, VkImage image, VkImageLayout& trackedLayout,
                                VkImageLayout newLayout, VkPipelineStageFlags srcStageMask,
                                VkPipelineStageFlags dstStageMask, VkAccessFlags srcAccessMask,
                                VkAccessFlags dstAccessMask, VkImageAspectFlags aspectMask,
-                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1,
+                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1);
-                               Uint32 layerCount = 1);
 
     SizeT CollectGarbage();
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTimerQueryManager.cpp

@@ -15,7 +15,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         MOBILEGL_ASSERT(initInfo.device != VK_NULL_HANDLE, "VkTimerQueryManager::Initialize requires valid VkDevice");
         MOBILEGL_ASSERT(initInfo.frameCount > 0, "VkTimerQueryManager::Initialize requires non-zero frame count");
         if (initInfo.timestampValidBits == 0 || initInfo.timestampPeriodNs <= 0.0f || initInfo.slotsPerPool == 0) {
-            MGLOG_W("VkTimerQueryManager: timestamps unsupported (validBits=%u, period=%f, slots=%u)",
+            MGLOG_W_ONCE("VkTimerQueryManager: timestamps unsupported (validBits=%u, period=%f, slots=%u)",
                     initInfo.timestampValidBits, initInfo.timestampPeriodNs, initInfo.slotsPerPool);
             return false;
         }
@@ -35,7 +35,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         for (auto& poolState : m_pools) {
             const VkResult result = vkCreateQueryPool(m_device, &poolInfo, nullptr, &poolState.pool);
             if (result != VK_SUCCESS) {
-                MGLOG_E("VkTimerQueryManager: vkCreateQueryPool failed with %s", VkResultToString(result));
+                MGLOG_E_ONCE("VkTimerQueryManager: vkCreateQueryPool failed with %s", VkResultToString(result));
                 Shutdown();
                 return false;
             }
@@ -90,7 +90,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto& poolState = m_pools[frameIndex];
         if (poolState.cursor >= m_slotsPerPool) {
             if (!poolState.exhaustionWarned) {
-                MGLOG_W("VkTimerQueryManager: frame %u timestamp pool exhausted (%u slots); further timer queries "
+                MGLOG_W_ONCE("VkTimerQueryManager: frame %u timestamp pool exhausted (%u slots); further timer queries "
                         "this frame fall back to the frontend path",
                         frameIndex, m_slotsPerPool);
                 poolState.exhaustionWarned = true;
@@ -120,7 +120,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_device, m_pools[record.poolIndex].pool, record.slot, 1, sizeof(resultWithAvailability),
             resultWithAvailability, sizeof(Uint64), VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WITH_AVAILABILITY_BIT);
         if (result != VK_SUCCESS && result != VK_NOT_READY) {
-            MGLOG_E("VkTimerQueryManager: vkGetQueryPoolResults failed with %s", VkResultToString(result));
+            MGLOG_E_ONCE("VkTimerQueryManager: vkGetQueryPoolResults failed with %s", VkResultToString(result));
             return false;
         }
         if (resultWithAvailability[1] == 0) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -229,6 +229,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                     GLint dstY, GLint width, GLint height, VkImageLayout srcRestoreLayout,
                                     VkImageLayout dstRestoreLayout, Bool stencilAspect);
         static SizeT GetReadbackTexelSize(VkFormat sourceFormat);
+        // Map a GL bottom-left-origin rectangle into the display-oriented swapchain image.
+        // Quarter-turn surface transforms swap the copy extent's axes.
+        static Bool MapDefaultFramebufferReadbackRect(GLint x, GLint y, GLsizei width, GLsizei height,
+                                                      VkExtent2D imageExtent,
+                                                      VkSurfaceTransformFlagBitsKHR preTransform,
+                                                      VkOffset2D* imageOffset, VkExtent2D* imageCopyExtent);
+        // Reorder a tightly packed block copied with MapDefaultFramebufferReadbackRect back into
+        // GL row order. The input block has swapped dimensions for 90/270 degree transforms.
+        static Bool RemapDefaultFramebufferReadback(const Uint8* rawPixels, Uint32 logicalWidth,
+                                                    Uint32 logicalHeight,
+                                                    VkSurfaceTransformFlagBitsKHR preTransform,
+                                                    SizeT texelSize, Uint8* outPixels);
         static Bool ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
                                           GLsizei width, GLsizei height, GLenum destinationFormat,
                                           GLenum destinationType, SizeT destinationRowStride,
@@ -298,6 +310,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The samplerAnisotropy device feature was granted, so GL_TEXTURE_MAX_ANISOTROPY_EXT is
         // honored rather than accepted-and-ignored.
         Bool IsSamplerAnisotropySupported() const { return m_samplerAnisotropyFeatureEnabled; }
+        // ARB_base_instance extends indirect command records with a non-zero firstInstance and
+        // requires gl_InstanceID to remain zero-based. Vulkan needs both features to honor that
+        // complete contract: one legalizes the command word, the other enables the shader rebase.
+        Bool IsNonZeroIndirectBaseInstanceSupported() const {
+            return m_drawIndirectFirstInstanceFeatureEnabled && m_shaderDrawParametersFeatureEnabled;
+        }
         // Ensures the frame command buffer is recording (same lazy pattern as
         // SetupDraw) and writes a bottom-of-pipe timestamp into the current
         // frame's pool. Null when unsupported or the pool is exhausted.
@@ -536,7 +554,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool m_samplerAnisotropyFeatureEnabled = false;
         Bool m_shaderDrawParametersExtensionEnabled = false;
         Bool m_shaderDrawParametersFeatureEnabled = false;
+        // Native subgroup topology, queried at device creation for the compute-module
+        // subgroup repairs (SubgroupSupportPolicy.h) and the REQUIRE_FULL_SUBGROUPS
+        // stage flag; 0 / false when the device has no usable compute subgroups or
+        // MOBILEGL_DISABLE_SUBGROUP forced them off.
+        Uint32 m_nativeSubgroupSize = 0;
+        Bool m_nativeSubgroupSupported = false;
+        Bool m_computeFullSubgroupsFeatureEnabled = false;
+        // VkPhysicalDeviceSubgroupSizeControlProperties::maxComputeWorkgroupSubgroups;
+        // 0 when the extension (and therefore the full-subgroups flag) is unavailable.
+        Uint32 m_maxComputeWorkgroupSubgroups = 0;
         Bool m_unformattedFloatStorageImagesEnabled = false;
+        // Set only after descriptor-indexing feature AND property queries prove that
+        // update-after-bind is legal for every descriptor category this renderer emits.
+        ProgramFactory::UpdateAfterBindLimits m_updateAfterBindLimits{};
         // fillModeNonSolid gates VK_POLYGON_MODE_LINE/_POINT (glPolygonMode); independentBlend gates
         // per-draw-buffer color write masks (glColorMaski). Both are cached at device creation and
         // drive a runtime fallback when the device lacks them.
@@ -547,6 +578,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // needs no feature). Both cached at device creation and drive a hard-fail-at-draw when absent.
         Bool m_dualSrcBlendFeatureEnabled = false;
         Bool m_primitiveTopologyListRestartFeatureEnabled = false;
+        // multiViewport gates rasterizing into more than one of ARB_viewport_array's 16 viewports
+        // (gl_ViewportIndex). m_maxRasterizableViewports is min(MAX_VIEWPORTS, device limit), or 1
+        // when the feature is off, and is the viewportCount a gl_ViewportIndex-writing pipeline
+        // declares - it is NOT what GL_MAX_VIEWPORTS reports, which is the frontend state width.
+        Bool m_multiViewportFeatureEnabled = false;
+        Uint32 m_maxRasterizableViewports = 1;
         // Union of shader stages sampled-read barriers may name; built at device creation
         // because geometry/tessellation stage bits are invalid in a barrier when their
         // feature is off (VUID-vkCmdPipelineBarrier-srcStageMask-04090/-04091), and
@@ -770,6 +807,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint32 m_lastLodProgramVersion = 0;
         Uint64 m_lastLodBindGeneration = 0;
         Uint64 m_lastLodParamsSum = 0;
+        // Sampling-resolution generation at probe time. The probe reads the effective
+        // sampler's filters/aniso/LOD range, whose setters bump only this counter -
+        // the params-version sum above never moves for them.
+        Uint64 m_lastLodSamplingGeneration = 0;
         ProgramFactory::CompileOptionFlags m_lastLodBaseFlags = {};
         ProgramFactory::CompileOptionFlags m_lastLodResultFlags = {};
 
@@ -807,6 +848,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 vaoLifetimeId = 0;
             Uint32 vaoConfigVersion = 0;
             const void* drawFbo = nullptr;
+            // Never-reused lifetime id beside the raw pointer + Uint16 version: a
+            // deleted FBO recycled at the same address with the same fresh version
+            // count would otherwise compare equal (same ABA as the render-pass
+            // manager's fast-path memo).
+            Uint64 drawFboLifetimeId = 0;
             Uint16 fboVersion = 0;
             Bool drawFboIsDefault = false;
             Uint renderStateVersion = 0;
@@ -830,6 +876,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // 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;
+            // The snapshotting draw's pipeline viewportCount. A pure function of the PROGRAM
+            // (writesViewportIndexBuiltin) and of a device feature fixed at renderer init, both
+            // of which the programLifetimeId/programVersion guards above already pin - carried
+            // here so the fast path does not re-fetch the program object to re-derive it.
+            Uint32 viewportCount = 1;
             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
@@ -897,6 +948,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // already sampleable.
         Vector<VkTextureManager::TextureResource*> m_sampledResourcesScratch;
         Vector<MG_State::GLState::ITextureObject*> m_storageImageTexturesScratch;
+        Vector<UniformManager::SamplerImageFeedbackBinding> m_samplerImageFeedbackScratch;
+        Vector<UniformManager::SamplerBindingOverride> m_samplerImageBindingOverridesScratch;
         Vector<VkBuffer> m_vertexBuffersScratch;
         Vector<VkDeviceSize> m_vertexOffsetsScratch;
         Vector<VkVertexInputAttributeDescription> m_patchedAttributesScratch;
@@ -1023,6 +1076,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkBuffer indexVkBuffer = VK_NULL_HANDLE;
             VkDeviceSize indexSliceOffset = 0;
             Uint64 indexFrameSerial = 0;
+            // The EBO carried a host map when the slice was recorded - the mirror of
+            // anyBufferMapped on the vertex half. A shadow-backed (non-adopted)
+            // persistent map mutates its shadow with no API call and no epoch bump, so
+            // the one-compare rescue must decline and re-run the acquire, whose
+            // SyncPersistentMappedRange is the push-down. A map taken AFTER the record
+            // is already covered: AcquirePersistentMap bumps the slice epoch for the
+            // request itself, adopted or declined.
+            Bool indexBufferMapped = false;
 
             // Bound per draw (first bindingCount elements).
             VkBuffer vkBuffers[kMaxBindings] = {};
@@ -1116,11 +1177,34 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             FrameContext::FrameData& frame,
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj);
+        // Vulkan forbids a sampled descriptor and writable storage descriptor from naming the
+        // same image subresource in one shader operation. Snapshot only the sampler side; the
+        // storage descriptor continues to name the application texture.
+        Bool PrepareSamplerImageFeedbackSnapshots(
+            FrameContext::FrameData& frame,
+            const MG_State::GLState::ProgramObject& program,
+            const ProgramFactory::VkProgramObject& programObj,
+            VkPipelineStageFlags consumerShaderStageMask);
 
         // 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);
+        // viewportCount is the bound pipeline's declared viewport count: 1 for every program that
+        // does not write gl_ViewportIndex (the memoized fast path), otherwise the renderer's
+        // rasterizable viewport count, which takes the unmemoized array path.
+        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo,
+                                       Uint32 viewportCount = 1);
+        void ApplyMultiViewportDynamicState(VkCommandBuffer commandBuffer, Uint32 viewportCount, const IntVec2& extent,
+                                            VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo);
+        VkRect2D ComputeGLScissorRect(Uint32 index, const IntVec2& extent,
+                                      VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo) const;
+        // How many viewports a draw with this program rasterizes into: 1 unless the program
+        // assigns gl_ViewportIndex AND the device enabled multiViewport. Both the pipeline's
+        // baked viewportCount and the dynamic arrays come from this one answer, so they cannot
+        // disagree.
+        Uint32 ResolveDrawViewportCount(Bool programWritesViewportIndex) const {
+            return programWritesViewportIndex && m_multiViewportFeatureEnabled ? m_maxRasterizableViewports : 1u;
+        }
 
         Bool UploadAndBindVertexBuffers(VkCommandBuffer commandBuffer, const MG_State::GLState::VertexArrayObject& vao,
                                         const ProgramFactory::VkProgramObject& programObj,

MobileGL/MG_Backend/DirectVulkan/SubgroupSupportPolicy.h

@@ -0,0 +1,63 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/SubgroupSupportPolicy.h
+// Copyright (c) 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 <Config.h>
+#include <Includes.h>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+    // The single decision point for how DirectVulkan implements GL_KHR_shader_subgroup,
+    // shared by capability advertisement (BackendObject) and module lowering
+    // (VulkanRenderer / ProgramFactory) so the two can never disagree.
+    //
+    // Native subgroups are the implementation whenever the device has them, whatever
+    // their width - subgroup operations execute on the hardware paths they were made
+    // for. Module-level repairs keep the GL contract intact around them:
+    //   - FixIterationRPSubgroupScratchPass patches the one known pack bug: iterationRP's
+    //     prefixSumCache[32], under-declared for sub-16-lane devices (8-lane lavapipe);
+    //   - FixIterationRPBarrierPass repairs Program 203's race between two reductions
+    //     reusing that scratch, when explicitly enabled;
+    //   - DeriveNumSubgroupsPass replaces the one builtin drivers get wrong
+    //     (gl_NumSubgroups) with the value the rest of the topology implies.
+    // The 32-lane shared-memory emulation (EmulateSubgroupsPass) is a LAST RESORT for
+    // devices with no subgroup support at all, and only when the user opts in with
+    // MOBILEGL_MAGMA_EMULATE_SUBGROUP=1; it never replaces available native operations.
+
+    inline constexpr Uint32 kEmulatedSubgroupSize = 32u;
+    inline constexpr Uint32 kEmulatedSubgroupStages = GL_COMPUTE_SHADER_BIT;
+    inline constexpr Uint32 kEmulatedSubgroupFeatures =
+        GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_VOTE_BIT_KHR |
+        GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR | GL_SUBGROUP_FEATURE_BALLOT_BIT_KHR |
+        GL_SUBGROUP_FEATURE_SHUFFLE_BIT_KHR | GL_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT_KHR |
+        GL_SUBGROUP_FEATURE_CLUSTERED_BIT_KHR | GL_SUBGROUP_FEATURE_QUAD_BIT_KHR;
+
+    inline Bool ShouldEmulateSubgroups(const Bool nativeSubgroupSupported) {
+        return MG_Config::Features.MagmaEmulateSubgroup && !nativeSubgroupSupported &&
+               !MG_Config::Features.DisableSubgroup;
+    }
+
+    inline Bool ShouldFixIterationRPSubgroupScratch() {
+        // Auto is ON: the patch is fingerprint-gated to iterationRP's reduction and
+        // grows one under-declared array; every other module passes through untouched.
+        return MG_Config::Features.FixIterationRPSubgroupScratch !=
+               MG_Config::QuirkOverride::ForceOff;
+    }
+
+    inline Bool ShouldFixIterationRPBarrier() {
+        return MG_Config::Features.IterationRPFixBarrier;
+    }
+
+    inline Bool ShouldDeriveNumSubgroups() {
+        // Auto is ON: gl_NumSubgroups must agree with the gl_SubgroupID range for the GL
+        // contract to hold, and the derived ceil() value is the one the renderer can pin
+        // with REQUIRE_FULL_SUBGROUPS - the driver builtin is the value with no
+        // cross-driver guarantee (Adreno returns 1 for an 8-subgroup dispatch).
+        return MG_Config::Features.DeriveNumSubgroups != MG_Config::QuirkOverride::ForceOff;
+    }
+} // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/VkIncludes.h

@@ -74,6 +74,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 // The context line (__VA_ARGS__ = its own format string + args) must be a SEPARATE log
 // call: appending its format to the base format while its arguments precede the base
 // arguments makes every conversion read the wrong slot (a %s pulling an int crashes).
+//
+// MGLOG_F and deliberately NOT latched. VK_VERIFY is the invariant-check macro: a Vulkan call
+// MobileGL believes it has already made legal came back non-success, which is a
+// should-never-happen state, not an expected failure mode a user hits. Those fast-fail loudly
+// and keep saying so - the log-quietness rules that latch W/E cover expected failures (driver
+// capability gaps, app misuse), not broken internal invariants. MOBILEGL_ASSERT below traps in
+// a DEBUG build; MGLOG_F is what makes the same condition visible in an INFO test run, where
+// the assert is compiled out by contract.
+//
+// A soft, recoverable failure must therefore NOT be routed through VK_VERIFY. Check the
+// VkResult directly and report it with MGLOG_E_ONCE - see VkTextureManager::SyncTextureResource,
+// where a driver legitimately refuses an image the format pre-check accepted.
 #define VK_VERIFY(expr, ...)                                                                                           \
     do {                                                                                                               \
         VkResult _vk_verify_result = (expr);                                                                           \

MobileGL/MG_Impl/EGLImpl/EGLImpl.cpp

@@ -21,7 +21,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         EGLStateContext* GetState() {
             if (!MG_State::pEGLContext) {
-                MGLOG_E("pEGLContext is null. MG_State may not be initialized.");
+                MGLOG_E_ONCE("pEGLContext is null. MG_State may not be initialized.");
             }
             return MG_State::pEGLContext.get();
         }
@@ -146,7 +146,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             state->DestroySurface(dpy, surface);
             return EGL_NO_SURFACE;
         }
@@ -172,11 +172,11 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             return EGL_FALSE;
         }
         if (!backendObject->SwapEGLBuffers(dpy, draw)) {
-            MGLOG_E("eglSwapBuffers failed on thread=%s dpy=%p draw=%p", CurrentThreadIdString().c_str(), dpy, draw);
+            MGLOG_E_ONCE("eglSwapBuffers failed on thread=%s dpy=%p draw=%p", CurrentThreadIdString().c_str(), dpy, draw);
             state->SetError(EGL_BAD_SURFACE);
             return EGL_FALSE;
         }
@@ -211,7 +211,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             return EGL_FALSE;
         }
         if (!backendObject->InitializeEGLDisplay(dpy, major, minor)) {
@@ -265,7 +265,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         if (releaseCurrentRequest) {
             if (auto* backendObject = MG_Backend::pActiveBackendObject.get()) {
                 if (!backendObject->MakeEGLCurrent(dpy, draw, read, ctx)) {
-                    MGLOG_E("eglMakeCurrent release failed in backend thread=%s", threadId.c_str());
+                    MGLOG_E_ONCE("eglMakeCurrent release failed in backend thread=%s", threadId.c_str());
                     state->MakeCurrent(oldDisplay, oldDraw, oldRead, oldContext);
                     state->SetError(EGL_BAD_ACCESS);
                     return EGL_FALSE;
@@ -277,12 +277,12 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             state->MakeCurrent(oldDisplay, oldDraw, oldRead, oldContext);
             return EGL_FALSE;
         }
         if (!backendObject->MakeEGLCurrent(dpy, draw, read, ctx)) {
-            MGLOG_E("eglMakeCurrent backend attach failed thread=%s dpy=%p draw=%p read=%p ctx=%p", threadId.c_str(),
+            MGLOG_E_ONCE("eglMakeCurrent backend attach failed thread=%s dpy=%p draw=%p read=%p ctx=%p", threadId.c_str(),
                     dpy, draw, read, ctx);
             state->SetError(EGL_BAD_ACCESS);
             state->MakeCurrent(oldDisplay, oldDraw, oldRead, oldContext);
@@ -703,7 +703,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             state->DestroySurface(dpy, surface);
             return EGL_NO_SURFACE;
         }
@@ -726,7 +726,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         }
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             return EGL_FALSE;
         }
         width = std::max<EGLint>(width, 1);
@@ -764,7 +764,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         MGLOG_D("eglGetProcAddress(%s)", name);
         void* proc = MG_Impl::GetProcAddress(name);
         if (!proc) {
-            MGLOG_W("Failed to get function: %s", name);
+            MGLOG_D("Failed to get function: %s", name);
             return nullptr;
         }
         return (__eglMustCastToProperFunctionPointerType)proc;

MobileGL/MG_Impl/GLImpl/Buffer/GL_Buffer.cpp

@@ -18,6 +18,7 @@
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/BufferEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/BufferEnumConverter.h>
+#include <MG_Util/Texture/PixelStoreProcessor.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
     namespace {
@@ -31,6 +32,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             NamedBufferData,
             NamedBufferSubData,
             CopyNamedBufferSubData,
+            ClearBufferData,
+            ClearBufferSubData,
             ClearNamedBufferData,
             ClearNamedBufferSubData,
             MapBufferRange,
@@ -65,6 +68,10 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return "NamedBufferSubData";
             case BufferOp::CopyNamedBufferSubData:
                 return "CopyNamedBufferSubData";
+            case BufferOp::ClearBufferData:
+                return "ClearBufferData";
+            case BufferOp::ClearBufferSubData:
+                return "ClearBufferSubData";
             case BufferOp::ClearNamedBufferData:
                 return "ClearNamedBufferData";
             case BufferOp::ClearNamedBufferSubData:
@@ -143,16 +150,6 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return 0;
             }
 
-            // The pattern is replicated verbatim, which is only the whole story while the client
-            // layout already matches the internal format - the case every entry point in practice
-            // uses, and the only one the conversion machinery here can express. Say so rather than
-            // quietly writing a differently-sized pattern.
-            const SizeT sourceSize = MG_Util::GetInputBytesPerPixel(inputFormat, pixelType);
-            if (sourceSize != elementSize) {
-                MGLOG_W("%s: clear pattern is %zu bytes but internalformat 0x%X stores %zu; "
-                        "converting between them is not implemented",
-                        GetBufferOpName(op), sourceSize, internalformat, elementSize);
-            }
             return elementSize;
         }
 
@@ -194,27 +191,59 @@ namespace MobileGL::MG_Impl::GLImpl {
             return true;
         }
 
-        void ClearNamedBufferRange_State(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size,
+        Bool BuildClearPattern(GLenum internalformat, GLenum format, GLenum type, const void* data,
+                               SizeT patternSize, BufferOp op, Vector<Uint8>& pattern) {
+            const TextureInternalFormat internal = MG_Util::ConvertGLEnumToTextureInternalFormat(internalformat);
+            const TextureInputFormat inputFormat = MG_Util::ConvertGLEnumToTextureInputFormat(format);
+            const TexturePixelDataType inputType = MG_Util::ConvertGLEnumToTexturePixelDataType(type);
+
+            Vector<Uint8> zeroInput;
+            const void* inputPixel = data;
+            if (inputPixel == nullptr) {
+                const SizeT inputSize = MG_Util::GetInputBytesPerPixel(inputFormat, inputType);
+                if (inputSize == 0) {
+                    MG_State::pGLContext->RecordError(
+                        ErrorCode::InvalidValue,
+                        MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", GetBufferOpName(op),
+                                                     "format and type do not describe a source pixel."));
+                    return false;
+                }
+                zeroInput.resize(inputSize);
+                inputPixel = zeroInput.data();
+            }
+
+            if (!MG_Util::PixelStoreProcessor::ConvertOnePixelToInternal(
+                    internal, inputFormat, inputType, inputPixel, pattern)) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", GetBufferOpName(op),
+                        std::format("Cannot convert one ({}, {}) pixel into internalformat 0x{:X}.",
+                                    MG_Util::ConvertGLEnumToString(format), MG_Util::ConvertGLEnumToString(type),
+                                    internalformat)));
+                return false;
+            }
+
+            if (data == nullptr) {
+                // GL defines a null clear value as all zero bits in the destination store, while
+                // retaining the format/type validation above.
+                pattern.assign(patternSize, 0);
+            }
+            return true;
+        }
+
+        void ClearBufferRange_State(const SharedPtr<MG_State::GLState::BufferObject>& bufferObject,
+                                    GLenum internalformat, GLintptr offset, GLsizeiptr size,
                                     GLenum format, GLenum type, const void* data, BufferOp op) {
             const SizeT patternSize = GetClearPatternSize(internalformat, format, type, op);
             if (patternSize == 0) return;
-
-            auto bufferObject = GetNamedBufferObject(buffer, op);
-            if (!bufferObject) return;
             if (!ValidateBufferClearRange(bufferObject, offset, size, patternSize, op)) return;
             if (size == 0) return;
 
-            Vector<Uint8> clearData(static_cast<SizeT>(size));
+            Vector<Uint8> pattern;
-            if (data) {
+            if (!BuildClearPattern(internalformat, format, type, data, patternSize, op, pattern)) return;
-                const auto* pattern = static_cast<const Uint8*>(data);
+            bufferObject->FillSubData({pattern.data(), pattern.size()}, static_cast<SizeT>(offset),
-                for (SizeT at = 0; at < clearData.size(); at += patternSize) {
+                                      static_cast<SizeT>(size));
-                    Memcpy(clearData.data() + at, pattern, patternSize);
-                }
-            } else {
-                Memset(clearData.data(), 0, clearData.size());
-            }
-
-            bufferObject->UploadSubData({clearData.data(), clearData.size()}, static_cast<SizeT>(offset));
         }
 
         auto& GetBufferBindingSlot(BufferTarget target) {
@@ -1197,16 +1226,33 @@ namespace MobileGL::MG_Impl::GLImpl {
                                         static_cast<SizeT>(writeOffset), static_cast<SizeT>(size));
     }
 
+    void ClearBufferData_State(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) {
+        auto bufferObject = GetBoundBufferObject(target, BufferOp::ClearBufferData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
+                               type, data, BufferOp::ClearBufferData);
+    }
+
+    void ClearBufferSubData_State(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size,
+                                  GLenum format, GLenum type, const void* data) {
+        auto bufferObject = GetBoundBufferObject(target, BufferOp::ClearBufferSubData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, offset, size, format, type, data,
+                               BufferOp::ClearBufferSubData);
+    }
+
     void ClearNamedBufferData_State(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data) {
         auto bufferObject = GetNamedBufferObject(buffer, BufferOp::ClearNamedBufferData);
         if (!bufferObject) return;
-        ClearNamedBufferRange_State(buffer, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
+        ClearBufferRange_State(bufferObject, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
                                type, data, BufferOp::ClearNamedBufferData);
     }
 
     void ClearNamedBufferSubData_State(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size,
                                        GLenum format, GLenum type, const void* data) {
-        ClearNamedBufferRange_State(buffer, internalformat, offset, size, format, type, data,
+        auto bufferObject = GetNamedBufferObject(buffer, BufferOp::ClearNamedBufferSubData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, offset, size, format, type, data,
                                BufferOp::ClearNamedBufferSubData);
     }
 
@@ -1662,6 +1708,15 @@ namespace MobileGL::MG_Impl::GLImpl {
         CopyNamedBufferSubData_State(readBuffer, writeBuffer, readOffset, writeOffset, size);
     }
 
+    void ClearBufferData(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) {
+        ClearBufferData_State(target, internalformat, format, type, data);
+    }
+
+    void ClearBufferSubData(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
+                            GLenum type, const void* data) {
+        ClearBufferSubData_State(target, internalformat, offset, size, format, type, data);
+    }
+
     void ClearNamedBufferData(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data) {
         ClearNamedBufferData_State(buffer, internalformat, format, type, data);
     }

MobileGL/MG_Impl/GLImpl/Buffer/GL_Buffer.h

@@ -27,6 +27,9 @@ namespace MobileGL::MG_Impl::GLImpl {
     void NamedBufferSubData(GLuint buffer, GLintptr offset, GLsizeiptr size, const void* data);
     void CopyNamedBufferSubData(GLuint readBuffer, GLuint writeBuffer, GLintptr readOffset, GLintptr writeOffset,
                                 GLsizeiptr size);
+    void ClearBufferData(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data);
+    void ClearBufferSubData(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
+                            GLenum type, const void* data);
     void ClearNamedBufferData(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data);
     void ClearNamedBufferSubData(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
                                  GLenum type, const void* data);

MobileGL/MG_Impl/GLImpl/Drawing/GL_Drawing.cpp

@@ -1259,7 +1259,15 @@ namespace MobileGL::MG_Impl::GLImpl {
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "instancecount must be non-negative."));
             return;
         }
-        if (!MG_State::pGLContext->ValidateTransformFeedbackName(id)) {
+        // "id is not the name of a transform feedback object" has to mean the same thing here
+        // as it does to glIsTransformFeedback, and the two predicates are not interchangeable:
+        // a name glGenTransformFeedbacks handed out is only reserved until it is first bound,
+        // and only the bind turns it into an object (GL 4.6 core 13.2.1). ValidateTransformFeedbackName
+        // answers the reservation question - the right one for glBindTransformFeedback, which is
+        // what turns a reserved name into an object - so using it here let a generated-but-unbound
+        // name through to the completed-span check below and raised INVALID_OPERATION where the
+        // spec asks for INVALID_VALUE. Name 0 is the default object and always drawable.
+        if (id != 0 && !MG_State::pGLContext->IsTransformFeedbackObject(id)) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,

MobileGL/MG_Impl/GLImpl/Exporting/Definitions.cpp

@@ -25,12 +25,12 @@
 #define DECLARE_GL_FUNCTION_STUB_HEAD(type, name, ...) MOBILEGL_GL_API type gl##name(__VA_ARGS__) {
 
 #define DECLARE_GL_FUNCTION_STUB_END(type, name, ...)                                                                  \
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);                                                                   \
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);                                                                   \
     return (type)1;                                                                                                    \
     }
 
 #define DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(type, name, ...)                                                        \
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);                                                                   \
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);                                                                   \
     }
 
 #define DECLARE_GL_FUNCTION_HEAD(type, name, ...) MOBILEGL_GL_API type gl##name(__VA_ARGS__) {
@@ -969,14 +969,14 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL3dv, GLuint index, const GLdoub
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL4dv, GLuint index, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribL4dv, index, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribLPointer, GLuint index, GLint size, GLenum type, GLsizei stride, const void* pointer) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribLPointer, index, size, type, stride, pointer)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetVertexAttribLdv, GLuint index, GLenum pname, GLdouble* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetVertexAttribLdv, index, pname, params)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedfv, index, v)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedfv, index, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
-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_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_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_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_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_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
 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_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)
@@ -985,8 +985,8 @@ DECLARE_GL_FUNCTION_HEAD(void, DrawElementsInstancedBaseVertexBaseInstance, GLen
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetActiveAtomicCounterBufferiv, GLuint program, GLuint bufferIndex, GLenum pname, GLint* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetActiveAtomicCounterBufferiv, program, bufferIndex, pname, params)
 DECLARE_GL_FUNCTION_HEAD(void, DrawTransformFeedbackInstanced, GLenum mode, GLuint id, GLsizei instancecount) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawTransformFeedbackInstanced, mode, id, instancecount)
 DECLARE_GL_FUNCTION_HEAD(void, DrawTransformFeedbackStreamInstanced, GLenum mode, GLuint id, GLuint stream, GLsizei instancecount) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawTransformFeedbackStreamInstanced, mode, id, stream, instancecount)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearBufferData, GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearBufferData, target, internalformat, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearBufferData, GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearBufferData, target, internalformat, format, type, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearBufferSubData, GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearBufferSubData, target, internalformat, offset, size, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearBufferSubData, GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearBufferSubData, target, internalformat, offset, size, format, type, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetInternalformati64v, GLenum target, GLenum internalformat, GLenum pname, GLsizei count, GLint64* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetInternalformati64v, target, internalformat, pname, count, params)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, InvalidateTexSubImage, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, InvalidateTexSubImage, texture, level, xoffset, yoffset, zoffset, width, height, depth)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, InvalidateTexImage, GLuint texture, GLint level) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, InvalidateTexImage, texture, level)
@@ -1061,7 +1061,7 @@ DECLARE_GL_FUNCTION_HEAD(void, TextureSubImage1D, GLuint texture, GLint level, G
 DECLARE_GL_FUNCTION_HEAD(void, TextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLenum type, const void* pixels) DECLARE_GL_FUNCTION_END_NO_RETURN(void, TextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, type, pixels)
 DECLARE_GL_FUNCTION_HEAD(void, TextureSubImage3D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels) DECLARE_GL_FUNCTION_END_NO_RETURN(void, TextureSubImage3D, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, type, pixels)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage1D, GLuint texture, GLint level, GLint xoffset, GLsizei width, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage1D, texture, level, xoffset, width, format, imageSize, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, imageSize, data)
+DECLARE_GL_FUNCTION_HEAD(void, CompressedTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CompressedTextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, imageSize, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage3D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage3D, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, imageSize, data)
 DECLARE_GL_FUNCTION_HEAD(void, CopyTextureSubImage1D, GLuint texture, GLint level, GLint xoffset, GLint x, GLint y, GLsizei width) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CopyTextureSubImage1D, texture, level, xoffset, x, y, width)
 DECLARE_GL_FUNCTION_HEAD(void, CopyTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CopyTextureSubImage2D, texture, level, xoffset, yoffset, x, y, width, height)
@@ -1849,7 +1849,7 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureImage3DEXT, GLuint texture,
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureImage2DEXT, GLuint texture, GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height, GLint border, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureImage2DEXT, texture, target, level, internalformat, width, height, border, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureImage1DEXT, GLuint texture, GLenum target, GLint level, GLenum internalformat, GLsizei width, GLint border, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureImage1DEXT, texture, target, level, internalformat, width, border, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage3DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage3DEXT, texture, target, level, xoffset, yoffset, zoffset, width, height, depth, format, imageSize, bits)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage2DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage2DEXT, texture, target, level, xoffset, yoffset, width, height, format, imageSize, bits)
+DECLARE_GL_FUNCTION_HEAD(void, CompressedTextureSubImage2DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CompressedTextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage1DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLsizei width, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage1DEXT, texture, target, level, xoffset, width, format, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetCompressedTextureImageEXT, GLuint texture, GLenum target, GLint lod, void* img) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetCompressedTextureImageEXT, texture, target, lod, img)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedMultiTexImage3DEXT, GLenum texunit, GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height, GLsizei depth, GLint border, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedMultiTexImage3DEXT, texunit, target, level, internalformat, width, height, depth, border, imageSize, bits)
@@ -2585,7 +2585,7 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, BindTransformFeedbackNV, GLenum target, GLui
 DECLARE_GL_FUNCTION_STUB_HEAD(void, DeleteTransformFeedbacksNV, GLsizei n, const GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DeleteTransformFeedbacksNV, n, ids)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GenTransformFeedbacksNV, GLsizei n, GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GenTransformFeedbacksNV, n, ids)
 MOBILEGL_GL_API GLboolean glIsTransformFeedbackNV(GLuint id) {
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);
     return GL_FALSE;
 }
 DECLARE_GL_FUNCTION_STUB_HEAD(void, PauseTransformFeedbackNV, void) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, PauseTransformFeedbackNV, )
@@ -3181,5 +3181,5 @@ MOBILEGL_GL_API void glVertexAttribDivisorARB(GLuint index, GLuint divisor) {
 }
 
 MOBILEGL_GL_API void glWindowRectanglesEXT(GLenum mode, GLsizei count, const GLint* box) {
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);
 }

MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp

@@ -547,7 +547,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                       GLint dstX1, GLint dstY1, GLbitfield mask, GLenum filter) {
         auto blitNamedFramebuffer = MG_Backend::gBackendFunctionsTable.GL.BlitNamedFramebuffer;
         if (!blitNamedFramebuffer) {
-            MGLOG_E("glBlitNamedFramebuffer skipped: backend does not implement explicit framebuffer blit.");
+            MGLOG_E_ONCE("glBlitNamedFramebuffer skipped: backend does not implement explicit framebuffer blit.");
             return;
         }
         blitNamedFramebuffer(readFramebuffer, drawFramebuffer, srcX0, srcY0, srcX1, srcY1, dstX0, dstY0, dstX1,
@@ -558,7 +558,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                          GLenum buffer, GLint drawbuffer, const GLfloat* value) {
         auto clearNamedFramebufferfv = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferfv;
         if (!clearNamedFramebufferfv) {
-            MGLOG_E("glClearNamedFramebufferfv skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferfv skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferfv(framebuffer, buffer, drawbuffer, value);
@@ -568,7 +568,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                          GLenum buffer, GLint drawbuffer, GLfloat depth, GLint stencil) {
         auto clearNamedFramebufferfi = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferfi;
         if (!clearNamedFramebufferfi) {
-            MGLOG_E("glClearNamedFramebufferfi skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferfi skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferfi(framebuffer, buffer, drawbuffer, depth, stencil);
@@ -578,7 +578,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                          GLenum buffer, GLint drawbuffer, const GLint* value) {
         auto clearNamedFramebufferiv = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferiv;
         if (!clearNamedFramebufferiv) {
-            MGLOG_E("glClearNamedFramebufferiv skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferiv skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferiv(framebuffer, buffer, drawbuffer, value);
@@ -588,7 +588,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                           GLenum buffer, GLint drawbuffer, const GLuint* value) {
         auto clearNamedFramebufferuiv = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferuiv;
         if (!clearNamedFramebufferuiv) {
-            MGLOG_E("glClearNamedFramebufferuiv skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferuiv skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferuiv(framebuffer, buffer, drawbuffer, value);

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -16,6 +16,7 @@
 #include <MG_State/GLState/ErrorState/ErrorInfo.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/BufferEnumConverter.h>
+#include <MG_Util/Converters/GLToMG/RenderStateEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/FramebufferEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/ErrorCodeConverter.h>
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
@@ -27,6 +28,11 @@
 #include <MG_Backend/BackendObjects.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
+    // Declared rather than #included from GL_RenderState.h on purpose: that header also declares
+    // a free function named BlendEquation, which would hide the ::MobileGL::BlendEquation enum
+    // this file's blend-state queries name unqualified.
+    GLboolean IsEnabledi(GLenum target, GLuint index);
+
     namespace {
         enum class IndexedBufferQueryKind {
             Binding,
@@ -339,26 +345,70 @@ namespace MobileGL::MG_Impl::GLImpl {
             return sampler ? static_cast<GLint>(sampler->GetExternalIndex()) : 0;
         }
 
-        // The ARB_viewport_array indexed rectangles. MobileGL keeps exactly one viewport, one
+        // The ARB_viewport_array indexed rectangles. Each of these is genuinely per-viewport
-        // scissor box and one depth range, so every in-range index answers with that single
+        // frontend state (RenderStateParameters::Viewports / ScissorBoxes / DepthRanges), so the
-        // value - but it has to come from the frontend state the non-indexed getters read.
+        // indexed getters must read the indexed storage - the generic path at the bottom of
-        // The generic path at the bottom of GetIntegeri_v is a raw backend passthrough that
+        // GetIntegeri_v is a raw backend passthrough that has no case for them and returned
-        // has no case for these, so routing them through it returned zeros.
+        // zeros, and routing them to the NON-indexed getter (what this used to do) answered every
+        // index with viewport 0's value, which is what
+        // KHR-GL43.viewport_array.{viewport,scissor,depth_range}_api caught.
         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.
+        // Component count of an indexed viewport-array query, so every width of getter writes the
+        // caller's whole buffer instead of just element 0 (GL 4.6 core 22.1).
+        GLsizei IndexedViewportQueryComponents(GLenum target) {
+            return target == GL_DEPTH_RANGE ? 2 : 4;
+        }
+
+        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it. The bound is
+        // the frontend's own state width, which is also exactly what GL_MAX_VIEWPORTS reports -
+        // taking it from the backend caps instead would let a device limit of 1 (a Vulkan device
+        // without the multiViewport feature) make index 1 illegal even though the state exists.
         Bool ValidateViewportQueryIndex(GLuint index, const char* caller) {
-            GLint maxViewports = 0;
+            if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
-            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;
         }
 
+        // The indexed viewport/scissor/depth-range state as floats, which is the widest lossless
+        // shape MobileGL stores (the viewport really is float state; the scissor box is integral
+        // and well inside float's exact range, and every depth range is in [0, 1]). Every indexed
+        // getter width funnels through this so they can never disagree with each other.
+        void ReadIndexedViewportStateFloat(GLenum target, GLuint index, GLfloat* out) {
+            switch (target) {
+            case GL_VIEWPORT: {
+                const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(index);
+                out[0] = viewport.x();
+                out[1] = viewport.y();
+                out[2] = viewport.z();
+                out[3] = viewport.w();
+                return;
+            }
+            case GL_SCISSOR_BOX: {
+                const IntVec4& box = MG_State::pGLContext->GetScissorBoxIndexed(index);
+                out[0] = static_cast<GLfloat>(box.x());
+                out[1] = static_cast<GLfloat>(box.y());
+                out[2] = static_cast<GLfloat>(box.z());
+                out[3] = static_cast<GLfloat>(box.w());
+                return;
+            }
+            case GL_DEPTH_RANGE: {
+                const FloatVec2& range = MG_State::pGLContext->GetDepthRangeIndexed(index);
+                out[0] = range.x();
+                out[1] = range.y();
+                return;
+            }
+            default:
+                MOBILEGL_ASSERT(false, "ReadIndexedViewportStateFloat: unexpected target 0x%x",
+                                static_cast<Uint32>(target));
+                return;
+            }
+        }
+
         void CopyIntsToBooleans(const GLint* src, SizeT count, GLboolean* dst) {
             for (SizeT i = 0; i < count; ++i) {
                 dst[i] = src[i] ? GL_TRUE : GL_FALSE;
@@ -383,7 +433,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         MGLOG_D("glGetString, name: %s", MG_Util::ConvertGLEnumToString(name).c_str());
         if (!activeBackendObject) {
-            MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return (GLubyte*)"Unknown";
         }
 
@@ -442,7 +492,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
         if (!activeBackendObject) {
-            MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return (GLubyte*)"Unknown";
         }
         const auto& rendererInfo = activeBackendObject->GetRendererInfo();
@@ -629,6 +679,17 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.ViewportBoundsRangeMax;
             return;
         }
+        // Viewport 0's rectangle, verbatim. Falling through to the integer width below would
+        // round the fractional rectangle a glViewportIndexedf(0, ...) is allowed to set, and
+        // glGetFloatv(GL_VIEWPORT) is a lossless query of float state.
+        case GL_VIEWPORT: {
+            const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(0);
+            params[0] = viewport.x();
+            params[1] = viewport.y();
+            params[2] = viewport.z();
+            params[3] = viewport.w();
+            return;
+        }
         case GL_MIN_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_MAX_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_FRAGMENT_INTERPOLATION_OFFSET_BITS: {
@@ -792,15 +853,32 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // GL 4.6 core 22.1: an indexed query answers EVERY indexed state, and GL_SCISSOR_TEST is
+        // indexed by viewport just like GL_BLEND is by draw buffer. Without this the integer
+        // width fell through to the backend passthrough and answered GL_INVALID_ENUM, which is
+        // the sticky error KHR-GL43.viewport_array.queries trips over at its next error check.
+        if (MG_Util::ConvertGLEnumToCapabilityInput(target) != CapabilityInput::Unknown) {
+            *data = IsEnabledi(target, index);
+            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
+        // (gl4cMultiBindTests and the viewport_array group both do).
-        // viewport and one scissor box, so every in-range index reports that one.
         case GL_VIEWPORT:
         case GL_SCISSOR_BOX:
+        case GL_DEPTH_RANGE: {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetIntegerv(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                // Round, not truncate: glGetIntegerv on floating-point state rounds to nearest
+                // (GL 4.6 core 22.2), so a 255.875-wide viewport reads back as 256 and not 255.
+                data[i] = static_cast<GLint>(std::lround(values[i]));
+            }
             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:
@@ -927,7 +1005,10 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetFloatv(target, data);
+            // Verbatim, NOT via the integer width: the viewport is float state and
+            // KHR-GL43.viewport_array.viewport_api compares the read-back with ==, so a
+            // glViewportIndexedf(i, 0.125f, ...) has to come back as 0.125f exactly.
+            ReadIndexedViewportStateFloat(target, index, data);
             return;
         }
         GLint ints[4] = {};
@@ -944,7 +1025,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetDoublev(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                data[i] = static_cast<GLdouble>(values[i]);
+            }
             return;
         }
         GLint ints[4] = {};
@@ -1020,7 +1106,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         // frontend-only value simply is not in the driver's table.
         GLint values[4] = {};
         GetIntegeri_v(target, index, values);
-        *data = static_cast<GLint64>(values[0]);
+        // The viewport-array rectangles are the only multi-component indexed state here; every
+        // other pname is scalar, so widening element 0 alone would silently truncate them.
+        const GLsizei components = IsIndexedViewportQuery(target) ? IndexedViewportQueryComponents(target) : 1;
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = static_cast<GLint64>(values[i]);
+        }
     }
 
     void GetInteger64v(GLenum pname, GLint64* params) {
@@ -1954,7 +2045,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
         if (!activeBackendObject) {
-            MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return;
         }
         const auto& rendererInfo = activeBackendObject->GetRendererInfo();
@@ -2192,7 +2283,15 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.MaxViewportHeight;
             break;
         case GL_MAX_VIEWPORTS:
-            *params = dynamicParameters.MaxViewports;
+            // The frontend's own state width, not the backend's device limit. GL 4.3 core
+            // requires MAX_VIEWPORTS >= 16 and every indexed viewport entry point validates
+            // against RenderStateParameters::MAX_VIEWPORTS, so reporting anything else would
+            // either advertise viewports the state cannot hold or reject indices it can. A
+            // Vulkan device without the multiViewport feature reports maxViewports == 1, which
+            // limits what can be RASTERIZED to more than one rectangle (see the multiViewport
+            // gate in VulkanRenderer), not what the GL state can hold; caps.MaxViewports keeps
+            // carrying that device number for exactly that decision.
+            *params = static_cast<GLint>(RenderStateParameters::MAX_VIEWPORTS);
             break;
         case GL_MINOR_VERSION:
             *params = rendererInfo.RendererGLInfo.TargetGLVersion.Minor;
@@ -2248,7 +2347,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = static_cast<GLint>(std::lround(dynamicParameters.MaxTextureMaxAnisotropy));
             break;
         default:
-            MGLOG_E("glGetIntegerv: Invalid enum %s (0x%X)", MG_Util::ConvertGLEnumToString(pname).c_str(), pname);
+            MGLOG_D("glGetIntegerv: Invalid enum %s (0x%X)", MG_Util::ConvertGLEnumToString(pname).c_str(), pname);
             MG_State::pGLContext->RecordError(ErrorCode::InvalidEnum,
                                               MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "GetIntegerv",
                                                                            std::format("Invalid enum: 0x{:X}", pname)));

MobileGL/MG_Impl/GLImpl/Program/GL_Program.cpp

@@ -908,7 +908,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             const SizeT span = UniformStorageSpanInBytes(ttype, size);
             if (pUBO == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
                 offset + span > programObject->GetUBOSize()) {
-                MGLOG_E("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
+                MGLOG_E_ONCE("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
                         program, location);
                 return;
             }
@@ -962,7 +962,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         const SizeT span = UniformStorageSpanInBytes(ttype, size);
         if (pUBO == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
             offset + span > programObject->GetUBOSize()) {
-            MGLOG_E("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
+            MGLOG_E_ONCE("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
                     program, location);
             return;
         }
@@ -1057,21 +1057,20 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        static Bool allowVSOnlyPrograms;
+        // Read fresh every link, never latched in a static: the capability is
-        static Bool initialized = false;
+        // per-backend, and a latch would freeze it across a backend teardown +
-        if (!initialized) {
+        // re-initialization (the previous function-static memo here never even set
+        // its own initialized flag, so it re-read every call anyway - this makes
+        // the always-fresh behavior the stated one). A struct-field read per
+        // glLinkProgram costs nothing.
         const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
         if (!activeBackendObject) {
-                MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return;
         }
-            const auto& rendererInfo = activeBackendObject->GetRendererInfo();
+        const Bool allowVSOnlyPrograms =
-            allowVSOnlyPrograms = (Int)rendererInfo.StaticBackendCapability.AllowVSOnlyPrograms;
+            activeBackendObject->GetRendererInfo().StaticBackendCapability.AllowVSOnlyPrograms;
-        }
-        const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
-        if (activeBackendObject) {
         programObject->SetMaxFragmentOutputColorNumber(activeBackendObject->GetDynamicParameters().MaxDrawBuffers);
-        }
         programObject->Link(!allowVSOnlyPrograms);
     }
 
@@ -1152,7 +1151,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             SizeT writeSize = ItemCount * sizeof(T);
             if (size < writeSize) {
                 // Metadata bug: degrade to a clamped copy instead of killing the process.
-                MGLOG_E("%s: uniform size mismatch at program %u location %u: expected at least %zu bytes, got %zu "
+                MGLOG_E_ONCE("%s: uniform size mismatch at program %u location %u: expected at least %zu bytes, got %zu "
                         "bytes; clamping",
                         __func__, programObject.GetExternalIndex(), location, ItemCount * sizeof(T), size);
                 writeSize = size;
@@ -1173,7 +1172,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                 offset + byteOffsetInsideUniform + writeSize > uboSize) {
                 // Should not happen: linking gives every settable uniform backing
                 // storage. Log and drop the write instead of faulting.
-                MGLOG_E("%s: uniform at program %u location %u has no backing storage (ubo=%p offset=%u size=%zu "
+                MGLOG_E_ONCE("%s: uniform at program %u location %u has no backing storage (ubo=%p offset=%u size=%zu "
                         "uboSize=%zu); dropping write",
                         __func__, programObject.GetExternalIndex(), location, static_cast<void*>(pUBO), offset,
                         writeSize, uboSize);
@@ -1807,7 +1806,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             break;
         }
         default:
-            MGLOG_E("%s: unknown pname = %p %s", __func__, pname, MG_Util::ConvertGLEnumToString(pname).c_str());
+            MGLOG_D("%s: unknown pname = %p %s", __func__, pname, MG_Util::ConvertGLEnumToString(pname).c_str());
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,

MobileGL/MG_Impl/GLImpl/Query/GL_Query.cpp

@@ -648,4 +648,39 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (!ValidateQueryStreamIndex(__FUNCTION__, target, index)) return;
         GetQueryiv(target, pname, params);
     }
+
+    void DestroyAllQueryObjects() {
+        // Detach the registry under the lock, release outside it - same discipline
+        // (and the same accepted teardown race) as DestroyAllSyncObjects. Without
+        // this drain, every query the app left undeleted survived full library
+        // teardown in the process-global registry: the objects and their backend
+        // wrappers leaked across Destroy/Initialize cycles, stale ids kept
+        // answering IsQuery == GL_TRUE in the re-initialized library, and a later
+        // glDeleteQueries could hand the OLD backend's handle to a DIFFERENT
+        // backend's DeleteBackendQuery, which casts it to the wrong wrapper type.
+        UnorderedMap<GLuint, QueryObject*> orphans;
+        {
+            const std::lock_guard<std::mutex> lock(g_queryObjectsMutex);
+            orphans.swap(g_liveQueryObjects);
+            g_activeTimeElapsedQueryId = 0;
+            g_activePrimitivesWrittenQueryId = 0;
+            g_activePrimitivesGeneratedQueryId = 0;
+            g_activeSamplesPassedQueryId = 0;
+        }
+        if (orphans.empty()) {
+            return;
+        }
+        // Backend handles must be released by the backend that created them, so
+        // this runs while the function table is still populated. Both backends'
+        // DeleteBackendQuery are generation-guarded, so a handle whose renderer
+        // or ES context is already gone frees only the wrapper.
+        const auto deleteBackendQuery = MG_Backend::gBackendFunctionsTable.GL.DeleteBackendQuery;
+        for (const auto& [_, queryObject] : orphans) {
+            if (deleteBackendQuery && queryObject->backendHandle) {
+                deleteBackendQuery(queryObject->backendHandle);
+            }
+            delete queryObject;
+        }
+        MGLOG_D("DestroyAllQueryObjects: reclaimed %zu query object(s) the app left undeleted", orphans.size());
+    }
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Query/GL_Query.h

@@ -29,4 +29,13 @@ namespace MobileGL::MG_Impl::GLImpl {
     void GetQueryBufferObjecti64v(GLuint id, GLuint buffer, GLenum pname, GLintptr offset);
     void GetQueryBufferObjectui64v(GLuint id, GLuint buffer, GLenum pname, GLintptr offset);
     void QueryCounter(GLuint id, GLenum target);
+    // Destroys every still-registered query object exactly as DeleteQueries would.
+    // GL requires queries to die with their context; called only from full library
+    // teardown (DestroyImpl), where no context survives on any thread, so the
+    // process-global registry can be drained wholesale. Must run while the backend
+    // function table is still populated: each backend handle has to be released by
+    // the backend that created it, never by a later re-initialized one (whose
+    // DeleteBackendQuery would cast the wrapper to the wrong backend's type).
+    // Same contract as DestroyAllSyncObjects.
+    void DestroyAllQueryObjects();
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.cpp

@@ -20,28 +20,118 @@ namespace MobileGL::MG_Impl::GLImpl {
         return std::clamp(static_cast<Float>(value), 0.0f, 1.0f);
     }
 
-    static Bool ValidateIndexedBlendCapability(GLenum target, GLuint index, const char* functionName) {
+    // GL 4.6 core 17.3.2 and 22.1 give exactly two indexed capabilities: GL_BLEND, indexed by
-        if (target != GL_BLEND) {
+    // draw buffer, and GL_SCISSOR_TEST, indexed by viewport. They have DIFFERENT bounds
+    // (MAX_DRAW_BUFFERS vs MAX_VIEWPORTS), so the limit is picked per target rather than shared.
+    static Bool ValidateIndexedCapability(GLenum target, GLuint index, const char* functionName) {
+        GLuint limit = 0;
+        const char* indexName = nullptr;
+        switch (target) {
+        case GL_BLEND:
+            limit = MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS;
+            indexName = "Buffer";
+            break;
+        case GL_SCISSOR_TEST:
+            limit = RenderStateParameters::MAX_VIEWPORTS;
+            indexName = "Viewport";
+            break;
+        default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
-                                             "Only GL_BLEND is supported for indexed capability state."));
+                                             "Only GL_BLEND and GL_SCISSOR_TEST are supported for indexed "
+                                             "capability state."));
             return false;
         }
 
-        if (index >= MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS) {
+        if (index >= limit) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
-                MakeUnique<GenericErrorInfo>(
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
-                    "MG_Impl/GLImpl", functionName,
+                                             String(indexName) + " index " + std::to_string(index) +
-                    "Buffer index " + std::to_string(index) + " is out of range. Max supported is " +
+                                                 " is out of range. Max supported is " + std::to_string(limit - 1) +
-                        std::to_string(MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS - 1) + "."));
+                                                 "."));
             return false;
         }
 
         return true;
     }
 
+    // ------------------ ARB_viewport_array parameter validation ------------------
+    // All three families share the same two shapes, so they share the two checkers. GL 4.6 core
+    // 13.6.1/17.3.2: an out-of-range index is GL_INVALID_VALUE, and so is a negative width or
+    // height. `first + count == MAX_VIEWPORTS` is LEGAL - only strictly greater is an error,
+    // which KHR-GL43.viewport_array.api_errors checks explicitly in both directions.
+    static Bool ValidateViewportIndex(GLuint index, const char* functionName) {
+        if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                         "Viewport index " + std::to_string(index) +
+                                             " is out of range. Max supported is " +
+                                             std::to_string(RenderStateParameters::MAX_VIEWPORTS - 1) + "."));
+        return false;
+    }
+
+    static Bool ValidateViewportRange(GLuint first, GLsizei count, const char* functionName) {
+        if (count < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "count must not be negative."));
+            return false;
+        }
+        // Widened before adding: first is a GLuint and count a GLsizei, so `first + count` in
+        // 32 bits can wrap past MAX_VIEWPORTS and let an out-of-range range through.
+        const Uint64 last = static_cast<Uint64>(first) + static_cast<Uint64>(count);
+        if (last > RenderStateParameters::MAX_VIEWPORTS) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "first (" + std::to_string(first) + ") + count (" +
+                                                 std::to_string(count) + ") exceeds GL_MAX_VIEWPORTS (" +
+                                                 std::to_string(RenderStateParameters::MAX_VIEWPORTS) + ")."));
+            return false;
+        }
+        return true;
+    }
+
+    template <typename T>
+    static Bool ValidateNonNegativeExtent(T width, T height, const char* functionName) {
+        if (width >= T(0) && height >= T(0)) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "Width and height must be non-negative."));
+        return false;
+    }
+
+    // The array forms are all-or-nothing: one bad element rejects the whole call with a SINGLE
+    // GL_INVALID_VALUE and leaves every rectangle untouched. api_errors relies on both halves -
+    // it passes a full 16-element array with exactly one negative extent and then asserts the
+    // error queue holds exactly one entry.
+    template <typename T>
+    static Bool ValidateArrayExtents(GLsizei count, const T* v, const char* functionName) {
+        for (GLsizei i = 0; i < count; ++i) {
+            if (v[i * 4 + 2] >= T(0) && v[i * 4 + 3] >= T(0)) continue;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "Width and height must be non-negative (element " + std::to_string(i) +
+                                                 ")."));
+            return false;
+        }
+        return true;
+    }
+
+    static Bool ValidateNonNullArray(const void* v, const char* functionName) {
+        if (v != nullptr) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "value pointer cannot be null."));
+        return false;
+    }
+
     static Bool TryConvertBlendEquation(GLenum mode, const char* functionName,
                                         ::MobileGL::BlendEquation& outEquation) {
         outEquation = MG_Util::ConvertGLEnumToBlendEquation(mode);
@@ -93,16 +183,70 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Viewport_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
+        if (!ValidateNonNegativeExtent(width, height, "Viewport_State")) return;
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Viewport_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
 
         MG_State::pGLContext->SetViewport(IntVec4(x, y, width, height));
     }
 
+    // ------------------ ARB_viewport_array setters ------------------
+    void ViewportArrayv_State(GLuint first, GLsizei count, const GLfloat* v) {
+        if (!ValidateViewportRange(first, count, "ViewportArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ViewportArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ViewportArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetViewportIndexed(first + static_cast<GLuint>(i),
+                                                     FloatVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ViewportIndexedf_State(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        if (!ValidateViewportIndex(index, "ViewportIndexedf_State")) return;
+        if (!ValidateNonNegativeExtent(w, h, "ViewportIndexedf_State")) return;
+
+        MG_State::pGLContext->SetViewportIndexed(index, FloatVec4(x, y, w, h));
+    }
+
+    void ScissorArrayv_State(GLuint first, GLsizei count, const GLint* v) {
+        if (!ValidateViewportRange(first, count, "ScissorArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ScissorArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ScissorArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetScissorBoxIndexed(first + static_cast<GLuint>(i),
+                                                       IntVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ScissorIndexed_State(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        if (!ValidateViewportIndex(index, "ScissorIndexed_State")) return;
+        if (!ValidateNonNegativeExtent(width, height, "ScissorIndexed_State")) return;
+
+        MG_State::pGLContext->SetScissorBoxIndexed(index, IntVec4(left, bottom, width, height));
+    }
+
+    void DepthRangeArrayv_State(GLuint first, GLsizei count, const GLdouble* v) {
+        if (!ValidateViewportRange(first, count, "DepthRangeArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "DepthRangeArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetDepthRangeIndexed(
+                first + static_cast<GLuint>(i),
+                FloatVec2(ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 0])),
+                          ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 1]))));
+        }
+    }
+
+    void DepthRangeIndexed_State(GLuint index, GLdouble n, GLdouble f) {
+        if (!ValidateViewportIndex(index, "DepthRangeIndexed_State")) return;
+
+        MG_State::pGLContext->SetDepthRangeIndexed(
+            index, FloatVec2(ClampUnitFloat(static_cast<GLfloat>(n)), ClampUnitFloat(static_cast<GLfloat>(f))));
+    }
+
     void StencilOpSeparate_State(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         Bool applyFront = false;
         Bool applyBack = false;
@@ -175,12 +319,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Scissor_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
+        if (!ValidateNonNegativeExtent(width, height, "Scissor_State")) return;
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Scissor_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
 
         MG_State::pGLContext->SetScissorBox(IntVec4(x, y, width, height));
     }
@@ -336,7 +475,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     GLboolean IsEnabledi_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "IsEnabledi_State")) {
+        if (!ValidateIndexedCapability(target, index, "IsEnabledi_State")) {
             return GL_FALSE;
         }
 
@@ -392,7 +531,14 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         GLint values[4] = {};
         GetIntegeri_v(target, index, values);
-        *data = values[0] != 0 ? GL_TRUE : GL_FALSE;
+        // The ARB_viewport_array rectangles are the only multi-component indexed state that
+        // reaches here; writing element 0 alone would leave the caller's other three untouched.
+        const GLsizei components = target == GL_VIEWPORT || target == GL_SCISSOR_BOX
+            ? 4
+            : (target == GL_DEPTH_RANGE ? 2 : 1);
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = values[i] != 0 ? GL_TRUE : GL_FALSE;
+        }
     }
 
     GLboolean IsEnabled_State(GLenum cap) {
@@ -725,7 +871,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Disablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Disablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Disablei_State")) {
             return;
         }
 
@@ -743,7 +889,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Enablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Enablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Enablei_State")) {
             return;
         }
 
@@ -797,6 +943,44 @@ namespace MobileGL::MG_Impl::GLImpl {
         Viewport_State(x, y, width, height);
     }
 
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v) {
+        ViewportArrayv_State(first, count, v);
+    }
+
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        ViewportIndexedf_State(index, x, y, w, h);
+    }
+
+    void ViewportIndexedfv(GLuint index, const GLfloat* v) {
+        // The index is validated before the pointer is touched: glViewportIndexedfv(MAX, nullptr)
+        // must be one GL_INVALID_VALUE, not a null dereference.
+        if (!ValidateViewportIndex(index, "ViewportIndexedfv")) return;
+        if (!ValidateNonNullArray(v, "ViewportIndexedfv")) return;
+        ViewportIndexedf_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v) {
+        ScissorArrayv_State(first, count, v);
+    }
+
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        ScissorIndexed_State(index, left, bottom, width, height);
+    }
+
+    void ScissorIndexedv(GLuint index, const GLint* v) {
+        if (!ValidateViewportIndex(index, "ScissorIndexedv")) return;
+        if (!ValidateNonNullArray(v, "ScissorIndexedv")) return;
+        ScissorIndexed_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v) {
+        DepthRangeArrayv_State(first, count, v);
+    }
+
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f) {
+        DepthRangeIndexed_State(index, n, f);
+    }
+
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         StencilOpSeparate_State(face, sfail, dpfail, dppass);
     }

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.h

@@ -20,6 +20,16 @@ namespace MobileGL::MG_Impl::GLImpl {
     void Enablei(GLenum target, GLuint index);
     void BlendFunc(GLenum sfactor, GLenum dfactor);
     void Viewport(GLint x, GLint y, GLsizei width, GLsizei height);
+    // ARB_viewport_array (core since GL 4.1). Every one of these addresses the same 16-element
+    // indexed state the classic glViewport/glScissor/glDepthRange trio broadcasts to.
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v);
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h);
+    void ViewportIndexedfv(GLuint index, const GLfloat* v);
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v);
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height);
+    void ScissorIndexedv(GLuint index, const GLint* v);
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v);
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f);
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass);
     void StencilOp(GLenum fail, GLenum zfail, GLenum zpass);
     void StencilMaskSeparate(GLenum face, GLuint mask);

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.cpp

@@ -621,7 +621,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         // the process down, which is never an acceptable answer to a query - see the same reasoning
         // above for the compressed-format path.
         void RecordUnsupportedLevelQueryStorage(const char* caller, GLenum pname) {
-            MGLOG_I("%s: glGetTexLevelParameter(pname=%s) is not implemented for texture-buffer "
+            MGLOG_W_ONCE("%s: glGetTexLevelParameter(pname=%s) is not implemented for texture-buffer "
                     "storage; recording GL_INVALID_OPERATION instead of terminating",
                     caller, MG_Util::ConvertGLEnumToString(pname).c_str());
             MG_State::pGLContext->RecordError(
@@ -870,7 +870,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                 MG_Util::GetInputBytesPerPixel(MG_Util::ConvertGLEnumToTextureInputFormat(format),
                                                MG_Util::ConvertGLEnumToTexturePixelDataType(type));
             if (readBytesPerTexel != bytesPerTexel) {
-                MGLOG_I("%s: cannot copy into a %zu-byte texel from a %zu-byte readback layout", caller,
+                MGLOG_W_ONCE("%s: cannot copy into a %zu-byte texel from a %zu-byte readback layout", caller,
                         bytesPerTexel, readBytesPerTexel);
                 return false;
             }
@@ -1376,6 +1376,47 @@ namespace MobileGL::MG_Impl::GLImpl {
         return true;
     }
 
+    // The same rules for the COMPRESSED entry points, whose payload size is the imageSize the
+    // caller passed rather than something derived from a (format, type) pair - and which have no
+    // datum size, so the alignment rule above does not apply to them. Shared by
+    // glCompressedTexImage2D and glCompressedTexSubImage2D so the two cannot drift; the point
+    // that is easy to get wrong and that KHR-GL44.buffer_storage.map_persistent_texture exists to
+    // check is the first one: a PERSISTENT mapping stays a legal transfer source.
+    Bool ValidateCompressedUnpackBufferSource(const void* data, SizeT imageSize, const char* caller) {
+        const auto& unpackBuffer =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelUnpack).GetBoundObject();
+        if (!unpackBuffer) return true;
+
+        if (unpackBuffer->IsMapped() && !(unpackBuffer->GetMappingAccess() & BufferMappingAccessBit::Persistent)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller, "Pixel unpack buffer is currently mapped."));
+            return false;
+        }
+
+        const SizeT offset = reinterpret_cast<SizeT>(data);
+        const SizeT bufferSize = unpackBuffer->GetSize();
+        if (offset > bufferSize || imageSize > bufferSize - offset) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Unpacking would read past the end of the pixel unpack buffer."));
+            return false;
+        }
+        return true;
+    }
+
+    // Where a compressed upload reads its blocks from: `data` is an offset into the bound unpack
+    // buffer when there is one, and a client pointer otherwise. Only meaningful once
+    // ValidateCompressedUnpackBufferSource has passed. Null means there is nothing to read, which
+    // GL leaves undefined and which callers must not dereference.
+    const void* CompressedUnpackSource(const void* data) {
+        const auto& unpackBuffer =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelUnpack).GetBoundObject();
+        if (!unpackBuffer) return data;
+        return reinterpret_cast<const char*>(unpackBuffer->MappedData()) + reinterpret_cast<SizeT>(data);
+    }
+
     void TexSubImage3D_State(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width,
                              GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels) {
         TextureUploadTarget textureUploadTarget = MG_Util::ConvertGLEnumToTextureUploadTarget(target);
@@ -1449,7 +1490,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (xoffset + width > static_cast<GLsizei>(texelSize.x()) ||
             yoffset + height > static_cast<GLsizei>(texelSize.y()) ||
             zoffset + depth > static_cast<GLsizei>(texelSize.z())) {
-            MGLOG_E("TexSubImage3D_State: Specified region exceeds texture level dimensions");
+            MGLOG_E_ONCE("TexSubImage3D_State: Specified region exceeds texture level dimensions");
             free(processedPixels);
             return;
         }
@@ -1558,7 +1599,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             {width, height, 1}, false, inputSize);
 
         if (!processedPixels || inputSize == 0) {
-            MGLOG_E("TexSubImage2D_State: Failed to process pixel data for TexSubImage2D, width: %d, height: %d", width,
+            MGLOG_E_ONCE("TexSubImage2D_State: Failed to process pixel data for TexSubImage2D, width: %d, height: %d", width,
                     height);
             if (processedPixels) free(processedPixels);
             return;
@@ -1572,7 +1613,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (xoffset + width > static_cast<GLsizei>(texelSize.x()) ||
             yoffset + height > static_cast<GLsizei>(texelSize.y())) {
-            MGLOG_E("TexSubImage2D_State: Specified region exceeds texture dimensions");
+            MGLOG_E_ONCE("TexSubImage2D_State: Specified region exceeds texture dimensions");
             free(processedPixels);
             return;
         }
@@ -2123,7 +2164,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (processedPixels && imageSize > 0) {
             if (imageSize != internalBytes) {
-                MGLOG_W("%s: Processed pixel data size (%zu) does not match expected size (%zu). "
+                MGLOG_W_ONCE("%s: Processed pixel data size (%zu) does not match expected size (%zu). "
                         "This may indicate an alignment or processing issue.",
                         __func__, imageSize, internalBytes);
             }
@@ -2238,6 +2279,23 @@ namespace MobileGL::MG_Impl::GLImpl {
             DiscardMipmapChainOnBaseRespecification(textureMipmapObject, textureUploadTarget, level);
             textureMipmapObject->AllocateStorage(textureUploadTarget, level,
                                                  {{width, height, 1}, internalBytes});
+            // GL 4.6 core 8.5: a SPECIFIC compressed internalformat (unlike a generic
+            // GL_COMPRESSED_* one, where the implementation is free to choose) commits the
+            // level to that format - GL_TEXTURE_COMPRESSED must then answer true for it and
+            // GL_TEXTURE_INTERNAL_FORMAT must report it, which is how an application asks for
+            // the size to hand glCompressedTexSubImage2D afterwards. Only the tag and the size
+            // are recorded: there is no BC/ETC codec here, so the texel shadow keeps the
+            // uncompressed storage this format resolved to (which is also what lets the level
+            // sample as the application's texels), and the compressed image the tag describes
+            // is zero-filled - the one reproducible answer glGetCompressedTexImage can give for
+            // an image nothing ever compressed. AllocateStorage above clears the tag, so this
+            // has to follow it.
+            const auto compressedInfo = MG_Util::GetCompressedFormatInfo(static_cast<GLenum>(internalformat));
+            if (compressedInfo.blockWidth != 0) {
+                textureMipmapObject->SetMipmapCompressedImage(
+                    textureUploadTarget, level, static_cast<GLenum>(internalformat), nullptr,
+                    MG_Util::CalculateCompressedTextureImageSize(compressedInfo, {width, height, 1}));
+            }
         }
 
         if (!originalPixels) {
@@ -2252,7 +2310,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (processedPixels && imageSize > 0) {
             if (imageSize != internalBytes) {
-                MGLOG_W("TexImage2D_State: Processed pixel data size (%zu) does not match expected size (%zu). "
+                MGLOG_W_ONCE("TexImage2D_State: Processed pixel data size (%zu) does not match expected size (%zu). "
                         "This may indicate an alignment or processing issue.",
                         imageSize, internalBytes);
             }
@@ -2965,9 +3023,9 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_TEXTURE_INTERNAL_FORMAT:
             if (params) {
                 // A level stored compressed must report the token it was given, not the
-                // uncompressed format backing it (GL 4.6 core 8.11). Only glCompressedTexImage* sets
+                // uncompressed format backing it (GL 4.6 core 8.11). glCompressedTexImage2D sets
-                // that tag, so every level created by glTexImage*D - including one given a compressed
+                // that tag, and so does a glTexImage2D given a SPECIFIC compressed internalformat;
-                // internalformat - still answers with its resolved storage format.
+                // every other level answers with its resolved storage format.
                 const GLenum compressedFormat = GetCompressedLevelFormat(textureObject, textureUploadTarget, level);
                 *params = (compressedFormat != GL_NONE)
                               ? (GLint)compressedFormat
@@ -3103,9 +3161,9 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_TEXTURE_INTERNAL_FORMAT:
             if (params) {
                 // A level stored compressed must report the token it was given, not the
-                // uncompressed format backing it (GL 4.6 core 8.11). Only glCompressedTexImage* sets
+                // uncompressed format backing it (GL 4.6 core 8.11). glCompressedTexImage2D sets
-                // that tag, so every level created by glTexImage*D - including one given a compressed
+                // that tag, and so does a glTexImage2D given a SPECIFIC compressed internalformat;
-                // internalformat - still answers with its resolved storage format.
+                // every other level answers with its resolved storage format.
                 const GLenum compressedFormat = GetCompressedLevelFormat(textureObject, textureUploadTarget, level);
                 *params = (GLfloat)((compressedFormat != GL_NONE)
                                         ? compressedFormat
@@ -3324,12 +3382,84 @@ namespace MobileGL::MG_Impl::GLImpl {
                          dstY, dstZ, srcWidth, srcHeight, srcDepth);
     }
 
+    namespace {
+        // The eleven targets GL 4.6 core 18.3.2 accepts. GL_TEXTURE_BUFFER, the six cube FACE
+        // enums and every PROXY enum all convert to a TextureTarget this frontend recognises,
+        // so ValidateTextureTarget lets them through; here they are INVALID_ENUM.
+        Bool ValidateCopyImageTarget(GLenum target, const char* endpointName) {
+            switch (target) {
+            case GL_RENDERBUFFER:
+            case GL_TEXTURE_1D:
+            case GL_TEXTURE_1D_ARRAY:
+            case GL_TEXTURE_2D:
+            case GL_TEXTURE_2D_ARRAY:
+            case GL_TEXTURE_2D_MULTISAMPLE:
+            case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+            case GL_TEXTURE_3D:
+            case GL_TEXTURE_CUBE_MAP:
+            case GL_TEXTURE_CUBE_MAP_ARRAY:
+            case GL_TEXTURE_RECTANGLE:
+                return true;
+            default:
+                break;
+            }
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("{} is not a target glCopyImageSubData accepts as the {}.",
+                                MG_Util::ConvertGLEnumToString(target), endpointName)));
+            return false;
+        }
+
+        IntVec3 GetCopyImageLevelSize(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                      TextureUploadTarget uploadTarget, GLint level) {
+            const auto* mipmapTexture = MG_State::GLState::AsMipmapTexture(textureObject.get());
+            if (!mipmapTexture) return textureObject->GetBaseSize();
+            return mipmapTexture->GetMipmapTexelSize(uploadTarget, static_cast<Uint>(level));
+        }
+
+        // glCopyImageSubData names an object that must already exist, and GL 4.6 core 18.3.2
+        // spells the failure INVALID_VALUE - "if either name does not correspond to a valid
+        // object". The shared ValidateTextureObject says INVALID_OPERATION, which is right for
+        // the ~30 entry points that reach it through a BOUND object (where the name was never
+        // in question and the fault is the binding), so this is a local rule rather than a
+        // change to the helper.
+        Bool ValidateCopyImageObjectExists(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                           const char* endpointName) {
+            if (textureObject) return true;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("The {} name does not correspond to an existing image object.", endpointName)));
+            return false;
+        }
+
+        // Same split for the target/object disagreement: GL 4.6 core 18.3.2 makes a target that
+        // does not match the object INVALID_ENUM, where the shared uniformity helper records
+        // INVALID_OPERATION for the upload paths that share it.
+        Bool ValidateCopyImageTargetMatchesObject(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                                  TextureTarget target, const char* endpointName) {
+            if (!textureObject || textureObject->GetTarget() == target) return true;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("The {} target {} does not match the target the object was created with ({}).",
+                                endpointName, MG_Util::ConvertTextureTargetToString(target),
+                                MG_Util::ConvertTextureTargetToString(textureObject->GetTarget()))));
+            return false;
+        }
+    } // namespace
+
     Bool ValidateCopyImageSubData_State(const SharedPtr<MG_State::GLState::ITextureObject>& srcTexture,
-                                        GLenum srcTarget, GLint srcLevel,
+                                        GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY,
                                         const SharedPtr<MG_State::GLState::ITextureObject>& dstTexture,
-                                        GLenum dstTarget, GLint dstLevel,
+                                        GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY,
                                         GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
-        if (!TextureImpl::ValidateTextureObject(srcTexture) || !TextureImpl::ValidateTextureObject(dstTexture)) {
+        if (!ValidateCopyImageObjectExists(srcTexture, "source") ||
+            !ValidateCopyImageObjectExists(dstTexture, "destination")) {
             return false;
         }
         const auto srcTextureTarget = MG_Util::ConvertGLEnumToTextureTarget(srcTarget);
@@ -3338,14 +3468,30 @@ namespace MobileGL::MG_Impl::GLImpl {
             !TextureImpl::ValidateTextureTarget(dstTextureTarget)) {
             return false;
         }
-        if (!TextureImpl::ValidateTextureTargetUniformity(srcTexture, srcTextureTarget) ||
+        // GL_TEXTURE_BUFFER and the cube FACE enums convert to a target this frontend knows, but
-            !TextureImpl::ValidateTextureTargetUniformity(dstTexture, dstTextureTarget)) {
+        // 18.3.2 does not accept them here - only the eleven whole-image targets do.
+        if (!ValidateCopyImageTarget(srcTarget, "source") || !ValidateCopyImageTarget(dstTarget, "destination")) {
+            return false;
+        }
+        if (!ValidateCopyImageTargetMatchesObject(srcTexture, srcTextureTarget, "source") ||
+            !ValidateCopyImageTargetMatchesObject(dstTexture, dstTextureTarget, "destination")) {
             return false;
         }
         if (!TextureImpl::ValidateTextureLevelNumber(srcLevel) ||
             !TextureImpl::ValidateTextureLevelNumber(dstLevel)) {
             return false;
         }
+        // ValidateTextureLevelNumber only bounds the index by GL_MAX_TEXTURE_SIZE; it cannot
+        // see that this particular texture stops at level 0. Both backends turn <level> into an
+        // image subresource with no further checking (DirectVulkan builds a VkImageCopy from it,
+        // DirectGLES forwards it to the ES copy), so a level the texture never had reached the
+        // driver as an out-of-range mip index - on Adreno that is a SIGSEGV inside
+        // vkCmdCopyImage, which is what KHR-GL43.copy_image.non_existent_mipmap used to do to
+        // the whole glcts process. The answer the spec asks for is GL_INVALID_VALUE.
+        if (!TextureImpl::ValidateTextureLevelExists(srcTexture, srcLevel, __func__) ||
+            !TextureImpl::ValidateTextureLevelExists(dstTexture, dstLevel, __func__)) {
+            return false;
+        }
         if (srcWidth < 0 || srcHeight < 0 || srcDepth < 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
@@ -3356,7 +3502,44 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (srcWidth == 0 || srcHeight == 0 || srcDepth == 0) {
             return false;
         }
-        if (!TextureImpl::ValidateBaseInternalFormatMatch(srcTexture->GetFormat(), dstTexture->GetFormat())) {
+        // A multisample image can only be copied to one with the same sample count, and a
+        // single-sample image reports zero - so this one comparison is also what rejects
+        // copying between a multisample target and a non-multisample one.
+        if (srcTexture->GetSamples() != dstTexture->GetSamples()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("The two images have different sample counts ({} vs. {}).",
+                                srcTexture->GetSamples(), dstTexture->GetSamples())));
+            return false;
+        }
+        // 18.3.2: both images must be complete. An incomplete one has no defined texels to copy
+        // and no defined storage to copy into.
+        if (!srcTexture->IsComplete() || !dstTexture->IsComplete()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("A copied image is incomplete (source complete: {}, destination complete: {}).",
+                                srcTexture->IsComplete(), dstTexture->IsComplete())));
+            return false;
+        }
+        const auto srcUploadTarget = GetPrimaryUploadTarget(srcTexture);
+        const auto dstUploadTarget = GetPrimaryUploadTarget(dstTexture);
+        const auto srcBlock = TextureImpl::ResolveCopyImageTexelBlock(
+            srcTexture->GetFormat(), GetCompressedLevelFormat(srcTexture, srcUploadTarget, srcLevel));
+        const auto dstBlock = TextureImpl::ResolveCopyImageTexelBlock(
+            dstTexture->GetFormat(), GetCompressedLevelFormat(dstTexture, dstUploadTarget, dstLevel));
+        if (!TextureImpl::ValidateCopyImageFormatCompatibility(srcBlock, dstBlock)) {
+            return false;
+        }
+        const IntVec3 srcLevelSize = GetCopyImageLevelSize(srcTexture, srcUploadTarget, srcLevel);
+        const IntVec3 dstLevelSize = GetCopyImageLevelSize(dstTexture, dstUploadTarget, dstLevel);
+        if (!TextureImpl::ValidateCopyImageBlockAlignment(srcBlock, srcX, srcY, srcWidth, srcHeight,
+                                                          srcLevelSize.x(), srcLevelSize.y(), "source") ||
+            !TextureImpl::ValidateCopyImageBlockAlignment(dstBlock, dstX, dstY, srcWidth, srcHeight,
+                                                          dstLevelSize.x(), dstLevelSize.y(), "destination")) {
             return false;
         }
         return true;
@@ -3468,10 +3651,170 @@ namespace MobileGL::MG_Impl::GLImpl {
         RecordUnsupportedCompressedFormat(__func__);
     }
 
+    // Replaces a block-aligned rectangle of the compressed image glCompressedTexImage2D (or a
+    // compressed glTexImage2D) shadowed for this level. Same deviation as the image call it
+    // patches: the uncompressed texel shadow beside it is NOT touched, because there is no
+    // BC/ETC codec here to decode the incoming blocks with - so what changes is the image
+    // glGetCompressedTexImage hands back, not what the level samples as. Marking the texels
+    // dirty would therefore only re-upload bytes that did not change.
     void CompressedTexSubImage2D_State(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
                                        GLsizei height, GLenum format, GLsizei imageSize, const void* data) {
-        // TODO: implement compressed upload - see CompressedTexImage2D_State.
+        // ======================= Converting ================================
+        const auto textureUploadTarget = MG_Util::ConvertGLEnumToTextureUploadTarget(target);
+        const auto textureTarget = MG_Util::ConvertGLEnumToTextureTarget(target);
+        // Zero block width doubles as "format is not a specific compressed format", the
+        // INVALID_ENUM case - one lookup answers both questions.
+        const auto compressedInfo = MG_Util::GetCompressedFormatInfo(format);
+
+        // ===================== Error Checking ==============================
+        if (!TextureImpl::ValidateTextureUploadTarget(textureUploadTarget)) return;
+        // A proxy holds no image to modify; only the glTexImage*/glCompressedTexImage* pair
+        // accepts one.
+        if (TextureImpl::IsProxyTextureTarget(textureUploadTarget)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "A proxy target has no texture image to modify."));
+            return;
+        }
+        if (!TextureImpl::ValidateTextureLevelNumber(level)) return;
+        if (!TextureImpl::ValidateTextureLevelWithUploadTarget(textureUploadTarget, level)) return;
+        if (width < 0 || height < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "width and height must be non-negative."));
+            return;
+        }
+        if (compressedInfo.blockWidth == 0) {
             RecordUnsupportedCompressedFormat(__func__);
+            return;
+        }
+
+        auto& textureObject = GetTextureObjectByTarget(textureUploadTarget, textureTarget);
+        if (!TextureImpl::ValidateTextureObject(textureObject)) return;
+        auto* textureMipmapObject = MG_State::GLState::AsMipmapTexture(textureObject.get());
+        if (textureMipmapObject == nullptr) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Texture storage is not mipmap-backed."));
+            return;
+        }
+        // GL 4.6 core 8.7: INVALID_OPERATION unless the image being modified is stored in
+        // exactly this compressed format. That is also what makes the block arithmetic below
+        // sound - the level's grid is measured with THIS format's block size.
+        const GLenum levelFormat =
+            textureMipmapObject->GetMipmapCompressedFormat(textureUploadTarget, static_cast<Uint>(level));
+        if (levelFormat != format) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "format does not match the internal format of the texture image."));
+            return;
+        }
+
+        const IntVec3 levelSize = textureMipmapObject->GetMipmapTexelSize(textureUploadTarget, static_cast<Uint>(level));
+        // Written as a subtraction rather than `xoffset + width > levelSize.x()`: both operands
+        // are application-supplied GLints, so the sum is free to overflow, and a signed overflow
+        // is undefined behaviour that a compiler may resolve by assuming the check passes.
+        // levelSize is our own and non-negative, and the offsets are known non-negative by the
+        // time the subtraction runs, so this form cannot wrap.
+        if (xoffset < 0 || yoffset < 0 || width > levelSize.x() - xoffset || height > levelSize.y() - yoffset) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "The replaced region does not lie within the texture image."));
+            return;
+        }
+        // GL 4.6 core 8.7 for block-based formats: the region must start on a block boundary
+        // and must either be a whole number of blocks wide/high or run to the image's edge.
+        const Int blockWidth = static_cast<Int>(compressedInfo.blockWidth);
+        const Int blockHeight = static_cast<Int>(compressedInfo.blockHeight);
+        const Bool alignedX = (xoffset % blockWidth == 0) &&
+                              (width % blockWidth == 0 || xoffset + width == levelSize.x());
+        const Bool alignedY = (yoffset % blockHeight == 0) &&
+                              (height % blockHeight == 0 || yoffset + height == levelSize.y());
+        if (!alignedX || !alignedY) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "The replaced region is not aligned to the format's compressed blocks."));
+            return;
+        }
+        // Exactly the size the format and dimensions imply, which is also what keeps the copy
+        // below in bounds.
+        const SizeT expectedImageSize =
+            MG_Util::CalculateCompressedTextureImageSize(compressedInfo, {width, height, 1});
+        if (imageSize < 0 || static_cast<SizeT>(imageSize) != expectedImageSize) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "imageSize does not match the compressed image size."));
+            return;
+        }
+
+        // ======================= Processing ================================
+        if (!ValidateCompressedUnpackBufferSource(data, expectedImageSize, __func__)) return;
+        const void* compressedBytes = CompressedUnpackSource(data);
+        if (expectedImageSize == 0) return; // a zero-sized region is a legal no-op
+        if (compressedBytes == nullptr) {
+            // No unpack buffer and a null client pointer: there is nothing to read. GL leaves
+            // this undefined rather than erroring, and dereferencing it is the one answer that
+            // is never acceptable.
+            MGLOG_D("%s: null data with no pixel unpack buffer bound, nothing to replace", __func__);
+            return;
+        }
+
+        // Once per process: the call is about to succeed, and what it does is narrower than what
+        // an application has every right to expect from it. Before this existed the call answered
+        // GL_INVALID_ENUM, which was wrong but at least visible; a silent success that leaves the
+        // sampled texels untouched is the kind of thing that costs a day to find from the other
+        // end. MGLOG_W is the right level and now survives at INFO; it sat at MGLOG_I only
+        // while the Log.h ordering compiled warnings out of the builds that ship.
+        static std::atomic<Bool> announcedNoCodec{false};
+        if (!announcedNoCodec.exchange(true)) {
+            MGLOG_W("%s: the compressed blocks are stored verbatim and returned by "
+                    "glGetCompressedTexImage, but there is no BC/ETC decoder here, so they do not "
+                    "reach the texels this level SAMPLES as. Upload through glTexSubImage2D for "
+                    "that.",
+                    __func__);
+        }
+
+        // The level's compressed image is stored as one blob, so the rectangle is patched into
+        // a copy of it and the whole thing handed back. Compressed sub-image uploads are not a
+        // hot path, and this keeps the storage layer's compressed API to the two calls it has.
+        const SizeT blobSize =
+            textureMipmapObject->GetMipmapCompressedByteSize(textureUploadTarget, static_cast<Uint>(level));
+        const void* existing =
+            textureMipmapObject->MapMipmapCompressedImage(textureUploadTarget, static_cast<Uint>(level));
+        if (blobSize == 0 || existing == nullptr) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "The texture level holds no compressed image to modify."));
+            return;
+        }
+        Vector<Uint8> blob(blobSize);
+        Memcpy(blob.data(), existing, blobSize);
+
+        const SizeT blockByteSize = compressedInfo.blockByteSize;
+        const SizeT levelBlocksX = (static_cast<SizeT>(levelSize.x()) + compressedInfo.blockWidth - 1) /
+                                   compressedInfo.blockWidth;
+        const SizeT levelRowBytes = levelBlocksX * blockByteSize;
+        const SizeT regionBlocksX = (static_cast<SizeT>(width) + compressedInfo.blockWidth - 1) /
+                                    compressedInfo.blockWidth;
+        const SizeT regionBlocksY = (static_cast<SizeT>(height) + compressedInfo.blockHeight - 1) /
+                                    compressedInfo.blockHeight;
+        const SizeT firstBlockX = static_cast<SizeT>(xoffset) / compressedInfo.blockWidth;
+        const SizeT firstBlockY = static_cast<SizeT>(yoffset) / compressedInfo.blockHeight;
+        const SizeT regionRowBytes = regionBlocksX * blockByteSize;
+        const auto* source = static_cast<const Uint8*>(compressedBytes);
+        for (SizeT row = 0; row < regionBlocksY; ++row) {
+            const SizeT destOffset = (firstBlockY + row) * levelRowBytes + firstBlockX * blockByteSize;
+            if (destOffset + regionRowBytes > blobSize) break; // a level whose blob predates its size
+            Memcpy(blob.data() + destOffset, source + row * regionRowBytes, regionRowBytes);
+        }
+        textureMipmapObject->SetMipmapCompressedImage(textureUploadTarget, static_cast<Uint>(level), format,
+                                                      blob.data(), blobSize);
     }
 
     void CompressedTexSubImage1D_State(GLenum target, GLint level, GLint xoffset, GLsizei width, GLenum format,
@@ -3564,28 +3907,8 @@ namespace MobileGL::MG_Impl::GLImpl {
         // SetMipmapCompressedImage re-arms it.
         textureMipmapObject->AllocateStorage(textureUploadTarget, level, {{width, height, 1}, internalBytes});
 
-        const void* compressedBytes = data;
+        if (!ValidateCompressedUnpackBufferSource(data, expectedImageSize, __func__)) return;
-        const auto& pixelUnpackBufferObject =
+        const void* compressedBytes = CompressedUnpackSource(data);
-            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelUnpack).GetBoundObject();
-        if (pixelUnpackBufferObject) {
-            if (pixelUnpackBufferObject->IsMapped()) {
-                MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidOperation,
-                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                                 "Pixel unpack buffer is currently mapped."));
-                return;
-            }
-            const SizeT offset = reinterpret_cast<SizeT>(data);
-            const SizeT bufferSize = pixelUnpackBufferObject->GetSize();
-            if (offset > bufferSize || expectedImageSize > bufferSize - offset) {
-                MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidOperation,
-                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                                 "Unpacking would read past the end of the pixel unpack buffer."));
-                return;
-            }
-            compressedBytes = reinterpret_cast<const char*>(pixelUnpackBufferObject->MappedData()) + offset;
-        }
         textureMipmapObject->SetMipmapCompressedImage(textureUploadTarget, level, internalformat, compressedBytes,
                                                       expectedImageSize);
         textureMipmapObject->MarkStorageDirty(textureUploadTarget, level, true);
@@ -4095,6 +4418,13 @@ namespace MobileGL::MG_Impl::GLImpl {
         // core 8.19). Allocating only the primary one left the object cube-incomplete, so every
         // framebuffer it was attached to reported GL_FRAMEBUFFER_INCOMPLETE_ATTACHMENT. Every other
         // 2D target has exactly one upload target, so this loop is a no-op change for them.
+        // A specific compressed internalformat commits every level it allocates to that
+        // format, the same way glTexImage2D does - and here it matters twice over, because
+        // immutable storage plus glCompressedTexSubImage2D IS the modern way to upload a
+        // compressed texture: without the tag that sub-image call finds an uncompressed
+        // level and refuses it. Zero width means a generic (implementation's choice)
+        // format, which MobileGL answers with uncompressed storage, so it is not tagged.
+        const auto compressedInfo = MG_Util::GetCompressedFormatInfo(internalformat);
         for (const auto uploadTarget : textureObject->GetUploadTargets()) {
             for (GLsizei level = 0; level < levels; ++level) {
                 const GLsizei levelWidth = std::max<GLsizei>(1, width >> level);
@@ -4103,6 +4433,13 @@ namespace MobileGL::MG_Impl::GLImpl {
                     static_cast<SizeT>(levelWidth) * static_cast<SizeT>(levelHeight) * bytesPerPixel;
                 textureMipmapObject->AllocateStorage(uploadTarget, level, {{levelWidth, levelHeight, 1}, byteSize});
                 textureMipmapObject->MarkStorageDirty(uploadTarget, level, false);
+                if (compressedInfo.blockWidth != 0) {
+                    // After AllocateStorage, which clears the tag.
+                    textureMipmapObject->SetMipmapCompressedImage(
+                        uploadTarget, static_cast<Uint>(level), internalformat, nullptr,
+                        MG_Util::CalculateCompressedTextureImageSize(compressedInfo,
+                                                                     {levelWidth, levelHeight, 1}));
+                }
             }
             // See TextureStorage1D.
             textureMipmapObject->TruncateMipmapLevels(uploadTarget, static_cast<Uint>(levels));
@@ -4472,6 +4809,14 @@ namespace MobileGL::MG_Impl::GLImpl {
         free(processedPixels);
     }
 
+    void CompressedTextureSubImage2D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
+                                     GLsizei height, GLenum format, GLsizei imageSize, const void* data) {
+        auto textureObject = GetTextureObjectByName(texture, __func__);
+        WithTemporarilyBoundNamedTexture(textureObject, [&](GLenum target) {
+            CompressedTexSubImage2D_State(target, level, xoffset, yoffset, width, height, format, imageSize, data);
+        });
+    }
+
     void TextureSubImage3D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width,
                            GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels) {
         auto textureObject = GetTextureObjectByName(texture, __func__);
@@ -5369,10 +5714,15 @@ namespace MobileGL::MG_Impl::GLImpl {
     void CopyImageSubData(GLuint srcName, GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY, GLint srcZ,
                           GLuint dstName, GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY, GLint dstZ,
                           GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
-        auto srcTexture = GetTextureObjectByName(srcName, __func__);
+        // A missing name is INVALID_VALUE here, where GetTextureObjectByName's own diagnostic is
-        auto dstTexture = GetTextureObjectByName(dstName, __func__);
+        // INVALID_OPERATION - so resolve through the plain lookup, which answers a null
-        if (!ValidateCopyImageSubData_State(srcTexture, srcTarget, srcLevel, dstTexture, dstTarget, dstLevel,
+        // SharedPtr, and let the validator record the error this entry point owes.
-                                            srcWidth, srcHeight, srcDepth)) {
+        const SharedPtr<MG_State::GLState::ITextureObject> srcTexture =
+            MG_State::pGLContext->GetTextureObject(srcName);
+        const SharedPtr<MG_State::GLState::ITextureObject> dstTexture =
+            MG_State::pGLContext->GetTextureObject(dstName);
+        if (!ValidateCopyImageSubData_State(srcTexture, srcTarget, srcLevel, srcX, srcY, dstTexture, dstTarget,
+                                            dstLevel, dstX, dstY, srcWidth, srcHeight, srcDepth)) {
             return;
         }
         CopyImageSubData_Backend(srcTexture, srcTarget, srcLevel, srcX, srcY, srcZ, dstTexture, dstTarget, dstLevel,

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.h

@@ -37,6 +37,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                            GLenum format, GLenum type, const void* pixels);
     void TextureSubImage3D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width,
                            GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels);
+    void CompressedTextureSubImage2D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
+                                     GLsizei height, GLenum format, GLsizei imageSize, const void* data);
     void TextureParameterf(GLuint texture, GLenum pname, GLfloat param);
     void TextureParameterfv(GLuint texture, GLenum pname, const GLfloat* params);
     void TextureParameteri(GLuint texture, GLenum pname, GLint param);

MobileGL/MG_Impl/GLImpl/Texture/Validators.cpp

@@ -15,6 +15,7 @@
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToMG/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToStr/TextureEnumConverter.h>
+#include <MG_Util/Metrics/TextureMetrics.h>
 
 namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
     Bool ValidateTextureTarget(TextureTarget target) {
@@ -353,6 +354,63 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         return true;
     }
 
+    Bool ValidateTextureLevelExists(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject, Int level,
+                                    const char* caller) {
+        // A null object is somebody else's error to report - ValidateTextureObject runs
+        // first at every call site and has already recorded it.
+        if (!textureObject) return false;
+
+        const auto* mipmapTexture = MG_State::GLState::AsMipmapTexture(textureObject.get());
+        if (mipmapTexture == nullptr) {
+            // The only non-mipmap storage class is a buffer texture, and GL_TEXTURE_BUFFER is
+            // not a target glCopyImageSubData accepts at all (it is in the CTS's invalid-target
+            // set). Declining here is not the error code the spec asks for - that would be
+            // INVALID_ENUM from a target check this validator is not - but it does keep a
+            // texture with no image levels whatsoever from reaching a backend that would
+            // dereference a backend texture it never created.
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Texture has no mipmap levels to address."));
+            return false;
+        }
+
+        // What this number is, exactly, because two other things are almost it and neither is
+        // safe to assume: it is the number of level SLOTS the shadow has allocated - holes
+        // included, since MipmapStorage::AllocateLevel grows to level+1 and never fills the gap.
+        // For a cube map MipmapUploadTargetArray reports face +X's chain rather than the union.
+        //
+        // The guarantee that matters is one-sided: this count is always >= the level count the
+        // backends derive (VkTextureManager::GetUploadMipLevelCount stops at the first level
+        // with a non-positive extent, so it can only be shorter). That is the safe direction -
+        // no copy to a level the texture genuinely has is ever rejected here. It is NOT an
+        // exact match, so the backends keep their own range guard for the band in between: a
+        // chain with a hole (level 0 and 2 defined, 1 not) is accepted by this predicate and
+        // declined by the backend, which is a silent no-op rather than a copy. That band is a
+        // backend storage limitation, not a validation one - rejecting it here with
+        // INVALID_VALUE would be refusing a copy the spec permits.
+        const Uint levelCount = mipmapTexture->GetMipmapLevelCount();
+
+        if (levelCount == 0) {
+            // No image has ever been defined on this texture, so the fault is the texture,
+            // not the number: GL 4.6 core 18.3.2 asks for INVALID_OPERATION when an object a
+            // copy names is an incomplete texture.
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Texture has no image defined at any level."));
+            return false;
+        }
+        if (level < 0 || static_cast<Uint>(level) >= levelCount) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Texture level does not exist in this texture."));
+            return false;
+        }
+        return true;
+    }
+
     Bool ValidateTextureObject(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject) {
         if (!textureObject) {
             MG_State::pGLContext->RecordError(
@@ -458,26 +516,86 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         }
     } // namespace
 
-    Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2) {
+    CopyImageTexelBlock ResolveCopyImageTexelBlock(TextureInternalFormat format, GLenum compressedFormat) {
-        const auto unsizedFormat1 = MG_Util::ConvertInternalFormatToUnsized(format1);
+        CopyImageTexelBlock block{};
-        const auto unsizedFormat2 = MG_Util::ConvertInternalFormatToUnsized(format2);
+        if (compressedFormat != GL_NONE) {
-        if (unsizedFormat1 != unsizedFormat2) {
+            const auto info = MG_Util::GetCompressedFormatInfo(compressedFormat);
-            // The 3-argument GenericErrorInfo constructor used to be spelled as a single
+            if (info.blockByteSize != 0) {
-            // std::format() call whose format string was the component name, so every
+                block.byteSize = info.blockByteSize;
-            // diagnostic collapsed to the literal "MG_Impl/GLImpl". Format the message, then
+                block.blockWidth = info.blockWidth;
-            // hand over component/function/message separately.
+                block.blockHeight = info.blockHeight;
+                block.compressed = true;
+                return block;
+            }
+        }
+        // The size MobileGL actually stores a texel of this format in, which for every format GL
+        // gives a required size is that required size. The handful of legacy formats GL leaves
+        // implementation-defined (R3_G3_B2, RGB4/5/10/12, RGBA2/12) have no view class in table
+        // 8.22 to be compared against anyway, and this is the size that decides whether a raw
+        // copy between them would in fact preserve the bytes.
+        block.byteSize = MG_Util::GetSizedInternalFormatSizeInBytes(format);
+        return block;
+    }
+
+    Bool ValidateCopyImageFormatCompatibility(const CopyImageTexelBlock& srcBlock,
+                                              const CopyImageTexelBlock& dstBlock) {
+        if (srcBlock.byteSize == 0 || dstBlock.byteSize == 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
+                                             "A copied image has no storage whose texel size is known."));
+            return false;
+        }
+        if (srcBlock.byteSize != dstBlock.byteSize) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
                 MakeUnique<GenericErrorInfo>(
-                    "MG_Impl/GLImpl", "ValidateBaseInternalFormatMatch",
+                    "MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
-                    std::format("The base internal format of the two formats do not match ({} vs. {})",
+                    std::format("The two images' texel blocks are different sizes ({} vs. {} bytes), so the "
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat1),
+                                "formats are not copy-compatible.",
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat2))));
+                                srcBlock.byteSize, dstBlock.byteSize)));
+            return false;
+        }
+        // Two compressed images additionally have to agree on the SHAPE of the block, not only
+        // its size: an 8-byte 4x4 block and a hypothetical 8-byte 8x8 one hold different texel
+        // counts, and GL 4.6 core 18.3.2 requires both dimensions to match.
+        if (srcBlock.compressed && dstBlock.compressed &&
+            (srcBlock.blockWidth != dstBlock.blockWidth || srcBlock.blockHeight != dstBlock.blockHeight)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
+                    std::format("The two compressed images have different block dimensions ({}x{} vs. {}x{}).",
+                                srcBlock.blockWidth, srcBlock.blockHeight, dstBlock.blockWidth,
+                                dstBlock.blockHeight)));
             return false;
         }
         return true;
     }
 
+    Bool ValidateCopyImageBlockAlignment(const CopyImageTexelBlock& block, Int x, Int y, Int width, Int height,
+                                         Int imageWidth, Int imageHeight, const char* endpointName) {
+        if (!block.compressed) return true;
+        const Int blockWidth = static_cast<Int>(block.blockWidth);
+        const Int blockHeight = static_cast<Int>(block.blockHeight);
+        if (blockWidth <= 1 && blockHeight <= 1) return true;
+        // The origin is unconditional; the extent gets the "or it reaches the edge of the image"
+        // exemption GL 4.6 core 18.3.2 grants, which is what lets a 16x16 BPTC image be copied
+        // whole even when the last block is partial.
+        const Bool originAligned = (x % blockWidth == 0) && (y % blockHeight == 0);
+        const Bool widthOk = (width % blockWidth == 0) || (x + width == imageWidth);
+        const Bool heightOk = (height % blockHeight == 0) || (y + height == imageHeight);
+        if (originAligned && widthOk && heightOk) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>(
+                "MG_Impl/GLImpl", "ValidateCopyImageBlockAlignment",
+                std::format("The {} region [{}, {}] + [{} x {}] is not aligned to the {}x{} compressed block "
+                            "grid of a {} x {} image.",
+                            endpointName, x, y, width, height, blockWidth, blockHeight, imageWidth, imageHeight)));
+        return false;
+    }
+
     Bool ValidateCopyTexImageBaseFormatSubset(TextureInternalFormat destFormat, TextureInternalFormat srcFormat) {
         const auto unsizedDest = MG_Util::ConvertInternalFormatToUnsized(destFormat);
         const auto unsizedSrc = MG_Util::ConvertInternalFormatToUnsized(srcFormat);

MobileGL/MG_Impl/GLImpl/Texture/Validators.h

@@ -30,6 +30,16 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
                                                           TextureInternalFormat internalFormat,
                                                           TexturePixelDataType type);
     Bool ValidateTextureLevelWithUploadTarget(TextureUploadTarget target, Int level);
+    // "Is <level> a level this texture actually has?", which ValidateTextureLevelNumber above
+    // does NOT answer - that one only bounds the index by GL_MAX_TEXTURE_SIZE and knows nothing
+    // about the object. Entry points that resolve a level straight into a backend image
+    // subresource need this one: a level the texture never had is GL_INVALID_VALUE (GL 4.6 core
+    // 18.3.2), and passing it through instead reaches the driver as an out-of-range subresource.
+    // Note the error split is per-entry-point, so this is not universally reusable:
+    // glClearTexImage owes INVALID_OPERATION for the same out-of-range level and spells its own
+    // copy of this predicate in GL_Texture.cpp (GetClearTextureObject).
+    Bool ValidateTextureLevelExists(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject, Int level,
+                                    const char* caller);
     Bool ValidateTextureObject(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject);
     // Rejects the per-target default texture objects (name 0) with GL_INVALID_OPERATION for entry
     // points that require a GenTextures-created texture, e.g. TexStorage* ("An INVALID_OPERATION
@@ -40,8 +50,32 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
                                          TextureTarget target);
     Bool ValidateTextureSubImageOffsets(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject, Int xoffset,
                                         Int width, Int yoffset = 0, Int height = 0, Int zoffset = 0, Int depth = 0);
-    // Exact base-format equality - what glCopyImageSubData's format compatibility needs.
+    // The texel block of one glCopyImageSubData endpoint, resolved to the two things the
-    Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2);
+    // compatibility rule actually asks about. `compressed` is not redundant with a block bigger
+    // than 1x1: it is what distinguishes "compressed, and so the region is measured in texels of
+    // a blocked image" from "uncompressed, and so it is measured in texels".
+    struct CopyImageTexelBlock {
+        SizeT byteSize = 0;
+        Uint blockWidth = 1;
+        Uint blockHeight = 1;
+        Bool compressed = false;
+    };
+    // `compressedFormat` is the GLenum a glCompressedTexImage* upload recorded for the level, or
+    // GL_NONE. It has to be asked for separately because MobileGL stores every compressed format
+    // in uncompressed storage (ConvertGLEnumToTextureInternalFormat), so the TextureInternalFormat
+    // alone can no longer tell a BPTC image from the RGBA8 backing it.
+    CopyImageTexelBlock ResolveCopyImageTexelBlock(TextureInternalFormat format, GLenum compressedFormat);
+    // GL 4.6 core 18.3.2: the two images must be COMPATIBLE, and compatible means their texel
+    // blocks are the same SIZE - not that they share a base internal format. RGBA32UI into
+    // RGBA32F is legal (both 128-bit) while RGBA8 into RGBA32F is not, and a compressed image
+    // pairs with an uncompressed one whose texel is as big as the compressed block.
+    Bool ValidateCopyImageFormatCompatibility(const CopyImageTexelBlock& srcBlock,
+                                              const CopyImageTexelBlock& dstBlock);
+    // GL 4.6 core 18.3.2: for a compressed image the region's origin must sit on a block
+    // boundary and its size must be a whole number of blocks - unless the edge it runs to is
+    // the edge of the image.
+    Bool ValidateCopyImageBlockAlignment(const CopyImageTexelBlock& block, Int x, Int y, Int width, Int height,
+                                         Int imageWidth, Int imageHeight, const char* endpointName);
     // GL 4.6 SS 8.6 subset rule for glCopyTexImage*: the read buffer must supply every component
     // the requested internalformat asks for, but may supply more.
     Bool ValidateCopyTexImageBaseFormatSubset(TextureInternalFormat destFormat, TextureInternalFormat srcFormat);

MobileGL/MG_Impl/GLImpl/VertexArray/GL_VertexArray.cpp

@@ -527,7 +527,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (!MG_Backend::pActiveBackendObject ||
             !MG_Backend::pActiveBackendObject->GetDynamicParameters().SupportsFloat64VertexAttributes) {
-            MGLOG_I("VertexAttribLFormat: attribute %u asked for a 64-bit (GL_DOUBLE) format, but this "
+            MGLOG_W_ONCE("VertexAttribLFormat: attribute %u asked for a 64-bit (GL_DOUBLE) format, but this "
                     "backend has no double-precision vertex attribute support - see the "
                     "\"64-bit vertex attributes\" / \"shaderFloat64\" POST row for what that costs",
                     attribindex);

MobileGL/MG_Impl/GLXImpl/Exporting/Definitions.cpp

@@ -166,32 +166,32 @@ MOBILEGL_GLX_API int glXSwapIntervalSGI(int interval) {
 
 // Legacy entry points some loaders probe for; harmless no-op stubs.
 MOBILEGL_GLX_API void glXCopyContext(Display*, void*, void*, unsigned long) {
-    MGLOG_W("glx: glXCopyContext is not supported");
+    MGLOG_W_ONCE("glx: glXCopyContext is not supported");
 }
 
 MOBILEGL_GLX_API unsigned long glXCreateGLXPixmap(Display*, void*, unsigned long) {
-    MGLOG_W("glx: glXCreateGLXPixmap is not supported");
+    MGLOG_W_ONCE("glx: glXCreateGLXPixmap is not supported");
     return 0;
 }
 
 MOBILEGL_GLX_API void glXDestroyGLXPixmap(Display*, unsigned long) {}
 
 MOBILEGL_GLX_API unsigned long glXCreatePixmap(Display*, void*, unsigned long, const int*) {
-    MGLOG_W("glx: glXCreatePixmap is not supported");
+    MGLOG_W_ONCE("glx: glXCreatePixmap is not supported");
     return 0;
 }
 
 MOBILEGL_GLX_API void glXDestroyPixmap(Display*, unsigned long) {}
 
 MOBILEGL_GLX_API unsigned long glXCreatePbuffer(Display*, void*, const int*) {
-    MGLOG_W("glx: glXCreatePbuffer is not supported");
+    MGLOG_W_ONCE("glx: glXCreatePbuffer is not supported");
     return 0;
 }
 
 MOBILEGL_GLX_API void glXDestroyPbuffer(Display*, unsigned long) {}
 
 MOBILEGL_GLX_API void glXUseXFont(unsigned long, int, int, int) {
-    MGLOG_W("glx: glXUseXFont is not supported");
+    MGLOG_W_ONCE("glx: glXUseXFont is not supported");
 }
 
 MOBILEGL_GLX_API void glXSelectEvent(Display*, unsigned long, unsigned long) {}

MobileGL/MG_Impl/GLXImpl/GLXImpl.cpp

@@ -149,7 +149,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
                     fns->Sync = reinterpret_cast<decltype(fns->Sync)>(dlsym(fns->Library, "XSync"));
                 }
                 if (!fns->Valid()) {
-                    MGLOG_E("glx: failed to load libX11 entry points");
+                    MGLOG_E_ONCE("glx: failed to load libX11 entry points");
                 }
                 return fns;
             }();
@@ -314,7 +314,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             Uint32 width = 0;
             Uint32 height = 0;
             if (!QueryDrawableSize(dpy, drawable, width, height)) {
-                MGLOG_E("glx: XGetGeometry failed for drawable 0x%lx", drawable);
+                MGLOG_E_ONCE("glx: XGetGeometry failed for drawable 0x%lx", drawable);
                 return nullptr;
             }
 
@@ -326,7 +326,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             EGLSurface surface = EGLImpl::CreatePlatformWindowSurface(
                 context.Display, context.Config, reinterpret_cast<void*>(drawable), attribs);
             if (surface == EGL_NO_SURFACE) {
-                MGLOG_E("glx: failed to create window surface for drawable 0x%lx (%ux%u)", drawable,
+                MGLOG_E_ONCE("glx: failed to create window surface for drawable 0x%lx (%ux%u)", drawable,
                         width, height);
                 return nullptr;
             }
@@ -347,7 +347,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
             EGLDisplay display = EnsureDisplay();
             if (display == EGL_NO_DISPLAY) {
-                MGLOG_E("glx: no EGL display");
+                MGLOG_E_ONCE("glx: no EGL display");
                 return nullptr;
             }
             EGLImpl::BindAPI(EGL_OPENGL_API);
@@ -376,13 +376,13 @@ namespace MobileGL::MG_Impl::GLXImpl {
             EGLint configCount = 0;
             if (!EGLImpl::ChooseConfig(display, configAttribs, &config, 1, &configCount) ||
                 configCount <= 0) {
-                MGLOG_E("glx: eglChooseConfig failed");
+                MGLOG_E_ONCE("glx: eglChooseConfig failed");
                 return nullptr;
             }
 
             EGLContext eglContext = EGLImpl::CreateContext(display, config, shareContext, contextAttribs);
             if (eglContext == EGL_NO_CONTEXT) {
-                MGLOG_E("glx: eglCreateContext failed");
+                MGLOG_E_ONCE("glx: eglCreateContext failed");
                 return nullptr;
             }
 
@@ -931,7 +931,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
 
         if (!EGLImpl::MakeCurrent(object->Display, surface->Surface, surface->Surface,
                                   object->Context)) {
-            MGLOG_E("glx: eglMakeCurrent failed (drawable=0x%lx, ctx=%p)", drawable, context);
+            MGLOG_E_ONCE("glx: eglMakeCurrent failed (drawable=0x%lx, ctx=%p)", drawable, context);
             return 0;
         }
         t_current = {dpy, drawable, drawable, context};
@@ -943,7 +943,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
         if (context && draw != read) {
             // MobileGL's backends reject split draw/read surfaces; bind the draw
             // drawable for both, which is what every real caller here needs.
-            MGLOG_W("glx: glXMakeContextCurrent draw 0x%lx != read 0x%lx, using draw for both", draw,
+            MGLOG_W_ONCE("glx: glXMakeContextCurrent draw 0x%lx != read 0x%lx, using draw for both", draw,
                     read);
         }
         const int result = MakeCurrent(dpy, draw, context);
@@ -958,7 +958,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
         auto& surfaces = DrawableSurfaces();
         auto it = surfaces.find(drawable);
         if (it == surfaces.end()) {
-            MGLOG_W("glx: glXSwapBuffers with no surface for drawable 0x%lx", drawable);
+            MGLOG_W_ONCE("glx: glXSwapBuffers with no surface for drawable 0x%lx", drawable);
             return;
         }
         SyncSurfaceSize(dpy, drawable, it->second);

MobileGL/MG_Impl/GLXImpl/LookUp/LookUp.cpp

@@ -31,7 +31,7 @@ namespace MG_Impl::GLXImpl {
 #endif
         void* proc = MobileGL::MG_Impl::GetProcAddress(name);
         if (!proc) {
-            MGLOG_W("Failed to get function: %s", (const char*)name);
+            MGLOG_D("Failed to get function: %s", (const char*)name);
             return nullptr;
         }
 

MobileGL/MG_Impl/GetProcAddress.cpp

@@ -1403,7 +1403,7 @@ namespace MobileGL::MG_Impl {
         GETPROC(glFramebufferTextureMultiviewOVR, name);
         // GETPROC(glNamedFramebufferTextureMultiviewOVR, name);
 
-        MGLOG_W("GetProcAddress(%s) = nullptr!", name);
+        MGLOG_D("GetProcAddress(%s) = nullptr!", name);
         return nullptr;
     }
 } // namespace MobileGL::MG_Impl

MobileGL/MG_Impl/NSOpenGLImpl/NSOpenGLImpl.cpp

@@ -269,7 +269,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             }
             id metalLayerClass = reinterpret_cast<id>(objc_getClass("CAMetalLayer"));
             if (!metalLayerClass) {
-                MGLOG_E("NSOpenGLImpl: CAMetalLayer class not found");
+                MGLOG_E_ONCE("NSOpenGLImpl: CAMetalLayer class not found");
                 return nil;
             }
 
@@ -310,7 +310,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
                                                        static_cast<GLint>(geometry.DrawableSize.width),
                                                        static_cast<GLint>(geometry.DrawableSize.height));
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: failed to attach drawable: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: failed to attach drawable: %s", CGLImpl::ErrorString(error));
             }
         }
 
@@ -325,7 +325,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             }
             const auto error = CGLImpl::SetCurrentContext(context);
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: makeCurrentContext failed: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: makeCurrentContext failed: %s", CGLImpl::ErrorString(error));
             }
         }
 
@@ -345,7 +345,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             }
             const auto error = CGLImpl::FlushDrawable(context);
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: flushBuffer failed: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: flushBuffer failed: %s", CGLImpl::ErrorString(error));
             }
         }
 
@@ -377,7 +377,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
                                                        static_cast<GLint>(geometry.DrawableSize.width),
                                                        static_cast<GLint>(geometry.DrawableSize.height));
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: update failed to attach drawable: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: update failed to attach drawable: %s", CGLImpl::ErrorString(error));
                 return;
             }
             CGLImpl::UpdateContext(context);
@@ -421,7 +421,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             SEL selector = sel_registerName(selectorName);
             Method method = class_getInstanceMethod(cls, selector);
             if (!method) {
-                MGLOG_W("NSOpenGLImpl: missing instance method %s", selectorName);
+                MGLOG_W_ONCE("NSOpenGLImpl: missing instance method %s", selectorName);
                 return;
             }
             if (original) {
@@ -434,7 +434,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             SEL selector = sel_registerName(selectorName);
             Method method = class_getClassMethod(cls, selector);
             if (!method) {
-                MGLOG_W("NSOpenGLImpl: missing class method %s", selectorName);
+                MGLOG_W_ONCE("NSOpenGLImpl: missing class method %s", selectorName);
                 return;
             }
             method_setImplementation(method, replacement);
@@ -444,7 +444,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             Class pixelFormatClass = objc_getClass("NSOpenGLPixelFormat");
             Class contextClass = objc_getClass("NSOpenGLContext");
             if (!pixelFormatClass || !contextClass) {
-                MGLOG_W("NSOpenGLImpl: NSOpenGL classes are not loaded; hooks not installed");
+                MGLOG_W_ONCE("NSOpenGLImpl: NSOpenGL classes are not loaded; hooks not installed");
                 return false;
             }
 

MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.cpp

@@ -56,7 +56,7 @@ extern "C" HGLRC WINAPI wglCreateLayerContext(HDC hdc, int iLayerPlane) {
 }
 
 extern "C" BOOL WINAPI wglCopyContext(HGLRC, HGLRC, UINT) {
-    MGLOG_W("wglCopyContext is not supported");
+    MGLOG_W_ONCE("wglCopyContext is not supported");
     SetLastError(ERROR_NOT_SUPPORTED);
     return FALSE;
 }
@@ -132,24 +132,24 @@ extern "C" DWORD WINAPI wglSwapMultipleBuffers(UINT n, CONST WGLSWAP* ps) {
 // ---- Font rendering (legacy immediate-mode feature; not supported) ----
 
 extern "C" BOOL WINAPI wglUseFontBitmapsA(HDC, DWORD, DWORD, DWORD) {
-    MGLOG_W("wglUseFontBitmapsA is not supported");
+    MGLOG_W_ONCE("wglUseFontBitmapsA is not supported");
     return FALSE;
 }
 
 extern "C" BOOL WINAPI wglUseFontBitmapsW(HDC, DWORD, DWORD, DWORD) {
-    MGLOG_W("wglUseFontBitmapsW is not supported");
+    MGLOG_W_ONCE("wglUseFontBitmapsW is not supported");
     return FALSE;
 }
 
 extern "C" BOOL WINAPI wglUseFontOutlinesA(HDC, DWORD, DWORD, DWORD, FLOAT, FLOAT, int,
                                                  LPGLYPHMETRICSFLOAT) {
-    MGLOG_W("wglUseFontOutlinesA is not supported");
+    MGLOG_W_ONCE("wglUseFontOutlinesA is not supported");
     return FALSE;
 }
 
 extern "C" BOOL WINAPI wglUseFontOutlinesW(HDC, DWORD, DWORD, DWORD, FLOAT, FLOAT, int,
                                                  LPGLYPHMETRICSFLOAT) {
-    MGLOG_W("wglUseFontOutlinesW is not supported");
+    MGLOG_W_ONCE("wglUseFontOutlinesW is not supported");
     return FALSE;
 }
 

MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp

@@ -215,7 +215,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
             Uint32 width = 0;
             Uint32 height = 0;
             if (!QueryClientSize(hwnd, width, height)) {
-                MGLOG_E("wgl: GetClientRect failed for HWND %p", hwnd);
+                MGLOG_E_ONCE("wgl: GetClientRect failed for HWND %p", hwnd);
                 return nullptr;
             }
 
@@ -227,7 +227,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
             EGLSurface surface =
                 EGLImpl::CreatePlatformWindowSurface(context.Display, context.Config, hwnd, attribs);
             if (surface == EGL_NO_SURFACE) {
-                MGLOG_E("wgl: failed to create window surface for HWND %p (%ux%u)", hwnd, width, height);
+                MGLOG_E_ONCE("wgl: failed to create window surface for HWND %p (%ux%u)", hwnd, width, height);
                 return nullptr;
             }
 
@@ -244,7 +244,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
             const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
             EGLDisplay display = EnsureDisplay();
             if (display == EGL_NO_DISPLAY) {
-                MGLOG_E("wgl: no EGL display");
+                MGLOG_E_ONCE("wgl: no EGL display");
                 return nullptr;
             }
             EGLImpl::BindAPI(EGL_OPENGL_API);
@@ -275,13 +275,13 @@ namespace MobileGL::MG_Impl::WGLImpl {
             EGLConfig config = nullptr;
             EGLint configCount = 0;
             if (!EGLImpl::ChooseConfig(display, configAttribs, &config, 1, &configCount) || configCount <= 0) {
-                MGLOG_E("wgl: eglChooseConfig failed");
+                MGLOG_E_ONCE("wgl: eglChooseConfig failed");
                 return nullptr;
             }
 
             EGLContext eglContext = EGLImpl::CreateContext(display, config, shareContext, contextAttribs);
             if (eglContext == EGL_NO_CONTEXT) {
-                MGLOG_E("wgl: eglCreateContext failed");
+                MGLOG_E_ONCE("wgl: eglCreateContext failed");
                 return nullptr;
             }
 
@@ -612,7 +612,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
         auto& surfaces = WindowSurfaces();
         auto it = surfaces.find(hwnd);
         if (it == surfaces.end()) {
-            MGLOG_W("wglSwapBuffers: no surface for HWND %p", hwnd);
+            MGLOG_W_ONCE("wglSwapBuffers: no surface for HWND %p", hwnd);
             return FALSE;
         }
         SyncSurfaceSize(hwnd, it->second);
@@ -685,7 +685,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
         }
 
         if (!EGLImpl::MakeCurrent(object->Display, surface->Surface, surface->Surface, object->Context)) {
-            MGLOG_E("wglMakeCurrent: eglMakeCurrent failed (hdc=%p, hglrc=%p)", hdc, hglrc);
+            MGLOG_E_ONCE("wglMakeCurrent: eglMakeCurrent failed (hdc=%p, hglrc=%p)", hdc, hglrc);
             return FALSE;
         }
         t_current = {hdc, hglrc};

MobileGL/MG_IntegrationTest/CMakeLists.txt

@@ -24,9 +24,14 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 set(MGL_ITEST_ROOT ${CMAKE_CURRENT_LIST_DIR}/../..)
 
-# Only meaningful where MobileGL_s exists (i.e. not Android).
+# Desktop links the static implementation directly. Android runs the same
-if (NOT TARGET MobileGL_s)
+# executable from adb shell and links the shipping shared library instead.
-    message(STATUS "MobileGL_s is not available; skipping the integration test module")
+if (ANDROID)
+    set(MGL_ITEST_MOBILEGL_TARGET MobileGL)
+elseif (TARGET MobileGL_s)
+    set(MGL_ITEST_MOBILEGL_TARGET MobileGL_s)
+else()
+    message(STATUS "No MobileGL library target is available; skipping the integration test module")
     return()
 endif()
 
@@ -60,17 +65,31 @@ add_executable(MobileGLIntegrationTest
     Scenarios/FragCoordOriginScenario.cpp
     Scenarios/ClearThenReadPixelsScenario.cpp
     Scenarios/DepthStencilReadbackScenario.cpp
+    Scenarios/DepthStencilReadbackMatrixScenario.cpp
+    Scenarios/DepthStencilReadbackAttachmentShapeScenario.cpp
+    Scenarios/ClipDistanceScenario.cpp
+    Scenarios/ViewportArrayScenario.cpp
     Scenarios/SsboArrayLengthScenario.cpp
     Scenarios/DoublePrecisionScenario.cpp
     Scenarios/UniformInitializerScenario.cpp
     Scenarios/SwizzleAccessRoutineScenario.cpp
+    Scenarios/IterationRPFirstReductionScenario.cpp
+    Scenarios/IterationRPProgram203Scenario.cpp
+    Scenarios/IterationRPScratchFixScenario.cpp
     Scenarios/ProgramPipelineScenario.cpp
     Scenarios/ImageLoadStoreSsoScenario.cpp
+    Scenarios/ImageTargetKindScenario.cpp
+    Scenarios/ImageFormatQualifierScenario.cpp
     Scenarios/SsboDeclarationFormScenario.cpp
     Scenarios/Glsl420DeclarationScenario.cpp
     Scenarios/FragmentOutputArrayIndexScenario.cpp
     Scenarios/BufferTextureScenario.cpp
     Scenarios/VertexAttribBindingScenario.cpp
+    Scenarios/XfbCaptureBufferReuseScenario.cpp
+    Scenarios/VertexArrayEnableDisableScenario.cpp
+    Scenarios/CopyImageLevelRangeScenario.cpp
+    Scenarios/CopyImageLayeredScenario.cpp
+    Scenarios/LayeredAttachmentBarrierScenario.cpp
 )
 
 target_include_directories(MobileGLIntegrationTest PRIVATE
@@ -81,9 +100,20 @@ target_include_directories(MobileGLIntegrationTest PRIVATE
 # gtest, not gtest_main: Main.cpp installs the harness banner itself.
 target_link_libraries(MobileGLIntegrationTest PRIVATE
     GTest::gtest
-    MobileGL_s
+    ${MGL_ITEST_MOBILEGL_TARGET}
 )
 
+if (ANDROID)
+    find_library(MGL_ITEST_ANDROID_LIBRARY android REQUIRED)
+    find_library(MGL_ITEST_LOG_LIBRARY log REQUIRED)
+    find_library(MGL_ITEST_MEDIANDK_LIBRARY mediandk REQUIRED)
+    target_link_libraries(MobileGLIntegrationTest PRIVATE
+        ${MGL_ITEST_ANDROID_LIBRARY}
+        ${MGL_ITEST_LOG_LIBRARY}
+        ${MGL_ITEST_MEDIANDK_LIBRARY}
+    )
+endif()
+
 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
@@ -92,6 +122,10 @@ if (MSVC)
 endif()
 target_compile_definitions(MobileGLIntegrationTest PRIVATE -DNOMINMAX)
 
+if (ANDROID)
+    return()
+endif()
+
 # --- 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
@@ -238,6 +272,8 @@ 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})
+mgl_itest_join_environment(MGL_ITEST_GLES_FORCED_DS_ENVIRONMENT
+    "MOBILEGL_BACKEND_TYPE=DirectGLES" "MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION=1" ${MGL_ITEST_COMMON_ENV})
 
 # TIMEOUT on every entry: a GPU test that wedges must fail the run, not hang it.
 set(MGL_ITEST_TIMEOUT 120)
@@ -285,3 +321,21 @@ gtest_discover_tests(MobileGLIntegrationTest
         TIMEOUT ${MGL_ITEST_TIMEOUT}
         ENVIRONMENT "${MGL_ITEST_VULKAN_ASYNC_ENVIRONMENT}"
 )
+
+# A fourth registration, of the depth/stencil readback scenarios, with the ES
+# shader-sampling emulation forced on. Not paranoia - without it these scenarios are
+# UNFALSIFIABLE on the machines this suite runs on: OpenGL ES has no depth or stencil
+# readback in core, but Mesa accepts the reads anyway, so on llvmpipe every one of them
+# goes green through a native path that the Adreno device does not have. Deleting the
+# entire emulation left all of them passing. With the flag the native spellings are off
+# the table and only the path the device actually takes remains. DirectGLES only - the
+# emulation is DirectGLES's.
+gtest_discover_tests(MobileGLIntegrationTest
+    TEST_PREFIX "DirectGLES.ForcedDepthStencilEmulation."
+    TEST_FILTER "DepthStencilReadback*Scenario.*"
+    DISCOVERY_TIMEOUT 30
+    PROPERTIES
+        LABELS integration-gpu
+        TIMEOUT ${MGL_ITEST_TIMEOUT}
+        ENVIRONMENT "${MGL_ITEST_GLES_FORCED_DS_ENVIRONMENT}"
+)

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.cpp

@@ -15,6 +15,16 @@
 #include <ostream>
 #include <sstream>
 
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#elif defined(__ANDROID__)
+#include <android/hardware_buffer.h>
+#include <android/native_window.h>
+#include <media/NdkImage.h>
+#include <media/NdkImageReader.h>
+#endif
+
 // 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
@@ -32,7 +42,7 @@
 // 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>)
+#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__ANDROID__) && __has_include(<sys/wait.h>)
 #define MGITEST_HAVE_FORK_PREFLIGHT 1
 #include <csignal>
 #include <ctime>
@@ -53,6 +63,83 @@ namespace MGITest {
         constexpr int kSurfaceWidth = 128;
         constexpr int kSurfaceHeight = 96;
 
+#if defined(_WIN32)
+        HWND g_testWindow = nullptr;
+
+        HWND CreateTestWindow() {
+            static const wchar_t* const kClassName = L"MobileGLIntegrationTestWindow";
+            static bool registered = false;
+            if (!registered) {
+                WNDCLASSW windowClass{};
+                windowClass.lpfnWndProc = DefWindowProcW;
+                windowClass.hInstance = GetModuleHandleW(nullptr);
+                windowClass.lpszClassName = kClassName;
+                if (RegisterClassW(&windowClass) == 0 && GetLastError() != ERROR_CLASS_ALREADY_EXISTS) {
+                    return nullptr;
+                }
+                registered = true;
+            }
+            return CreateWindowExW(0, kClassName, L"MobileGL Integration Test", WS_OVERLAPPEDWINDOW,
+                                   CW_USEDEFAULT, CW_USEDEFAULT, kSurfaceWidth, kSurfaceHeight, nullptr, nullptr,
+                                   GetModuleHandleW(nullptr), nullptr);
+        }
+#elif defined(__ANDROID__)
+        AImageReader* g_imageReader = nullptr;
+        ANativeWindow* g_imageReaderWindow = nullptr;
+
+        void DrainImageReader(void*, AImageReader* reader) {
+            AImage* image = nullptr;
+            if (AImageReader_acquireNextImage(reader, &image) == AMEDIA_OK && image != nullptr) {
+                AImage_delete(image);
+            }
+        }
+
+        bool CreateImageReaderWindow() {
+            if (g_imageReaderWindow != nullptr) return true;
+            constexpr int kMaxImages = 4;
+            const media_status_t status = AImageReader_newWithUsage(
+                kSurfaceWidth, kSurfaceHeight, AIMAGE_FORMAT_RGBA_8888,
+                AHARDWAREBUFFER_USAGE_GPU_SAMPLED_IMAGE | AHARDWAREBUFFER_USAGE_GPU_COLOR_OUTPUT,
+                kMaxImages, &g_imageReader);
+            if (status != AMEDIA_OK || g_imageReader == nullptr) return false;
+
+            AImageReader_ImageListener listener = {nullptr, DrainImageReader};
+            AImageReader_setImageListener(g_imageReader, &listener);
+            if (AImageReader_getWindow(g_imageReader, &g_imageReaderWindow) != AMEDIA_OK ||
+                g_imageReaderWindow == nullptr) {
+                AImageReader_setImageListener(g_imageReader, nullptr);
+                AImageReader_delete(g_imageReader);
+                g_imageReader = nullptr;
+                return false;
+            }
+            ANativeWindow_acquire(g_imageReaderWindow);
+            return true;
+        }
+
+        void DestroyImageReaderWindow() {
+            if (g_imageReaderWindow != nullptr) {
+                ANativeWindow_release(g_imageReaderWindow);
+                g_imageReaderWindow = nullptr;
+            }
+            if (g_imageReader != nullptr) {
+                AImageReader_setImageListener(g_imageReader, nullptr);
+                AImageReader_delete(g_imageReader);
+                g_imageReader = nullptr;
+            }
+        }
+#endif
+
+        bool UseWindowSurface() {
+#if defined(_WIN32)
+            const char* value = std::getenv("MOBILEGL_ITEST_WINDOW_SURFACE");
+            return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0;
+#elif defined(__ANDROID__)
+            return true;
+#else
+            return false;
+#endif
+        }
+
         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);
@@ -87,10 +174,10 @@ namespace MGITest {
         // callers). surfaceless is the platform with no window-system dependency at
         // all; the surface this file then creates is still a pbuffer, which every
         // platform supports and which the amendment to this rule requires as the
-        // fallback shape. DISPLAY/WAYLAND_DISPLAY are cleared as well so that a
+        // fallback shape on desktop. Android instead supplies an AImageReader
+        // ANativeWindow. DISPLAY/WAYLAND_DISPLAY are cleared as well so that a
         // driver that consults them directly cannot reintroduce the dependency
-        // behind EGL's back. Desktop-only file: MG_IntegrationTest never builds
+        // behind EGL's back.
-        // for Android, so no device path is affected.
         void EnsureHeadlessPlatform() {
 #if defined(__linux__) && !defined(__ANDROID__)
             static bool done = false;
@@ -134,8 +221,9 @@ namespace MGITest {
                 return 3;
             }
 
+            const bool useWindowSurface = UseWindowSurface();
             const EGLint configAttribs[] = {EGL_SURFACE_TYPE,
-                                            EGL_PBUFFER_BIT,
+                                            useWindowSurface ? EGL_WINDOW_BIT : EGL_PBUFFER_BIT,
                                             EGL_RED_SIZE,
                                             8,
                                             EGL_GREEN_SIZE,
@@ -152,7 +240,9 @@ namespace MGITest {
             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");
+                outReason = WithEglError(useWindowSurface
+                                             ? "eglChooseConfig found no window-capable RGBA8/D24 config"
+                                             : "eglChooseConfig found no pbuffer-capable RGBA8/D24 config");
                 return 4;
             }
 
@@ -166,10 +256,32 @@ namespace MGITest {
                 return 5;
             }
 
+            EGLSurface surface = EGL_NO_SURFACE;
+            if (useWindowSurface) {
+#if defined(_WIN32)
+                if (g_testWindow == nullptr) g_testWindow = CreateTestWindow();
+                if (g_testWindow == nullptr) {
+                    outReason = "failed to create the Windows integration-test window";
+                    return 6;
+                }
+                surface = eglCreateWindowSurface(display, config, g_testWindow, nullptr);
+#elif defined(__ANDROID__)
+                if (!CreateImageReaderWindow()) {
+                    outReason = "failed to create the Android AImageReader integration-test window";
+                    return 6;
+                }
+                surface = eglCreateWindowSurface(display, config, g_imageReaderWindow, nullptr);
+#endif
+            } else {
                 const EGLint pbufferAttribs[] = {EGL_WIDTH, kSurfaceWidth, EGL_HEIGHT, kSurfaceHeight, EGL_NONE};
-            EGLSurface surface = eglCreatePbufferSurface(display, config, pbufferAttribs);
+                surface = eglCreatePbufferSurface(display, config, pbufferAttribs);
+            }
             if (surface == EGL_NO_SURFACE) {
-                outReason = WithEglError("eglCreatePbufferSurface failed");
+#if defined(__ANDROID__)
+                DestroyImageReaderWindow();
+#endif
+                outReason = WithEglError(useWindowSurface ? "eglCreateWindowSurface failed"
+                                                         : "eglCreatePbufferSurface failed");
                 return 6;
             }
             // The step that brings the whole backend up (DirectVulkan creates its
@@ -491,6 +603,14 @@ namespace MGITest {
         if (m_context != nullptr) eglDestroyContext(display, static_cast<EGLContext>(m_context));
         if (m_surface != nullptr) eglDestroySurface(display, static_cast<EGLSurface>(m_surface));
         eglTerminate(display);
+#if defined(_WIN32)
+        if (g_testWindow != nullptr) {
+            DestroyWindow(g_testWindow);
+            g_testWindow = nullptr;
+        }
+#elif defined(__ANDROID__)
+        DestroyImageReaderWindow();
+#endif
         m_context = nullptr;
         m_surface = nullptr;
         m_display = nullptr;

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.h

@@ -14,11 +14,11 @@
 // 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
+// Headless by construction: desktop uses an EGL pbuffer and Android uses an
-// context on a PBUFFER surface. No window, no window manager, no human. Unlike
+// AImageReader-backed ANativeWindow that needs no Activity. No window manager,
-// DriverBench the scenarios do draw to the DEFAULT framebuffer (that is where
+// no human. Unlike DriverBench the scenarios do draw to the DEFAULT framebuffer
-// the Y-flip lives) and do call eglSwapBuffers (that is the frame boundary the
+// (that is where the Y-flip lives) and do call eglSwapBuffers (that is the frame
-// cross-frame scenarios need to be real).
+// 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

MobileGL/MG_IntegrationTest/Main.cpp

@@ -31,8 +31,14 @@ namespace {
                 // silently bound to a workstation's window system is a different
                 // run from CI's and must be visible as one in the log.
                 const char* eglPlatform = std::getenv("EGL_PLATFORM");
-                std::fprintf(stderr, "  renderer: %s\n  surface:  %dx%d pbuffer (headless, EGL_PLATFORM=%s)\n",
+#if defined(__ANDROID__)
+                constexpr const char* surfaceKind = "AImageReader window";
+#else
+                constexpr const char* surfaceKind = "pbuffer";
+#endif
+                std::fprintf(stderr, "  renderer: %s\n  surface:  %dx%d %s (headless, EGL_PLATFORM=%s)\n",
                              gl.RendererString().c_str(), gl.Width(), gl.Height(),
+                             surfaceKind,
                              eglPlatform != nullptr ? eglPlatform : "<unset>");
             } else if (MGITest::RequireGpu()) {
                 std::fprintf(stderr,

MobileGL/MG_IntegrationTest/Scenarios/AdvertisedLimitsScenario.cpp

@@ -199,5 +199,61 @@ namespace MGITest {
                    "derived component limits are computed in";
         }
 
+        // ARB_viewport_array's own limits. They are advertised from three different places -
+        // GL_MAX_VIEWPORTS from the frontend's indexed state width, the bounds range and the
+        // subpixel bits from the backend caps table - and each backend fills that table from a
+        // different source, so all three are checked on both lanes.
+        //
+        // GL_VIEWPORT_BOUNDS_RANGE is the one that shipped wrong: GLES has no such query, the
+        // DirectGLES loader's glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE) therefore raised
+        // GL_INVALID_ENUM and left the probe's zero-initialized array in place, and MobileGL
+        // advertised [0, 0] - a range that admits no viewport origin at all, and the check that
+        // kept KHR-GL43.viewport_array.queries red on Espryt after the indexed-state work.
+        TEST_F(AdvertisedLimitsScenario, ViewportArrayLimitsMeetTheirGL43Floors) {
+            GLint maxViewports = -1;
+            glGetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(maxViewports, 16) << "GL 4.3 core table 23.53 sets the MAX_VIEWPORTS minimum at 16";
+            EXPECT_LE(maxViewports, 256) << "one viewport rectangle of indexed state is allocated per advertised "
+                                            "viewport, and the CTS sizes its arrays off this number";
+
+            GLfloat boundsRange[2] = {1.0f, -1.0f};
+            glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE, boundsRange);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_LE(boundsRange[0], -32768.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+            EXPECT_GE(boundsRange[1], 32767.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+
+            // KNOWN INFIDELITY, pinned here rather than hidden. MobileGL reports the driver's own
+            // VIEWPORT_SUBPIXEL_BITS (4 on llvmpipe, i.e. 1/16-pixel viewport precision), but the
+            // float viewport rectangle glViewportIndexedf stores is snapped to integers on its
+            // way to both backends (ComputeGLViewport, DirectGLES SyncRenderState). The STATE
+            // round trip is exact - which is all KHR-GL43.viewport_array.viewport_api checks, and
+            // all this cluster set out to fix - so the gap is in rasterization only: a fractional
+            // viewport origin rasterizes as if it had been rounded. Nothing in the suite or in
+            // Minecraft sets one. Only the spec floor is asserted; tightening this to EQ(0) would
+            // mean advertising no subpixel precision at all, which is a separate decision about a
+            // limit MobileGL currently passes through from the driver.
+            GLint subpixelBits = -1;
+            glGetIntegerv(GL_VIEWPORT_SUBPIXEL_BITS, &subpixelBits);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(subpixelBits, 0) << "GL 4.6 core table 23.60: VIEWPORT_SUBPIXEL_BITS has a minimum of 0, and "
+                                          "a negative value is what a sign-flipped uint32 looks like";
+
+            GLint viewportDims[2] = {-1, -1};
+            glGetIntegerv(GL_MAX_VIEWPORT_DIMS, viewportDims);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            GLint maxRenderbufferSize = -1;
+            glGetIntegerv(GL_MAX_RENDERBUFFER_SIZE, &maxRenderbufferSize);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            // GL 4.6 core 13.6.1: MAX_VIEWPORT_DIMS must be at least as large as the largest
+            // renderable surface, or a full-size framebuffer could not be fully viewported.
+            EXPECT_GE(viewportDims[0], maxRenderbufferSize);
+            EXPECT_GE(viewportDims[1], maxRenderbufferSize);
+        }
+
     } // namespace
 } // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/BufferTextureScenario.cpp

@@ -33,6 +33,7 @@
 // branch on. The driver POST's "Buffer textures" row is where that verdict is stated.
 
 #include <cstdint>
+#include <cstring>
 #include <string>
 #include <vector>
 
@@ -73,7 +74,60 @@ out vec4 o_color;
 void main() { o_color = vec4(float(vFace) / 255.0, 0.0, 0.0, 1.0); }
 )";
 
-        class BufferTextureScenario : public ScenarioTest {};
+        // A buffer texture bound as a WRITABLE image: the shader reads one texel and writes
+        // another, so a single dispatch proves the read direction (which already worked) and
+        // the write direction (which is what this exists for) apart from each other.
+        constexpr const char* kImageBufferCS = R"(#version 430 core
+layout(local_size_x = 1) in;
+layout(binding = 0, rgba8) uniform imageBuffer uImage;
+void main() {
+    vec4 read = imageLoad(uImage, 1);
+    imageStore(uImage, 0, vec4(0.0, 1.0, 0.0, 1.0));
+    imageStore(uImage, 2, read);
+}
+)";
+
+        class BufferTextureScenario : public ScenarioTest {
+        protected:
+            bool ComputeImagesAreUsable() const {
+                GLint maxImageUnits = 0;
+                GLint maxComputeImageUniforms = 0;
+                glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+                glGetIntegerv(GL_MAX_COMPUTE_IMAGE_UNIFORMS, &maxComputeImageUniforms);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return maxImageUnits >= 1 && maxComputeImageUniforms >= 1;
+            }
+
+            unsigned int MakeComputeProgram(const char* source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = GL_FALSE;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute shader did not compile: " << log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute program did not link: " << log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+        };
 
         // Draws the full-viewport quad and returns the red byte every fragment was painted with,
         // or -1 if the quad did not come out uniform (which would mean the flat varying, not the
@@ -171,4 +225,78 @@ void main() { o_color = vec4(float(vFace) / 255.0, 0.0, 0.0, 1.0); }
         EXPECT_EQ(FirstGLError(), 0u);
     }
 
+    // A shader may WRITE a buffer texture too, through an image unit, and the bytes it writes
+    // land in the backend's buffer - not in the frontend's CPU shadow, which is what MapBuffer
+    // and GetBufferSubData hand back. A storage-block write is flagged for exactly this reason
+    // and the shadow is refreshed on the next read; a buffer reached through an image unit is
+    // the same write through a different binding, and Espryt used to flag only the first, so
+    // an imageStore into a buffer texture was invisible to every CPU read that followed it -
+    // silently, with the correct value sitting in the driver's buffer the whole time.
+    //
+    // The read direction is asserted in the same dispatch (texel 2 is a copy of texel 1) so a
+    // failure here cannot be blamed on the image binding not working at all.
+    TEST_F(BufferTextureScenario, AnImageStoreIntoABufferTextureIsVisibleToTheCpu) {
+        if (!Ready()) return;
+        if (!ComputeImagesAreUsable()) GTEST_SKIP() << "no compute image units on this host";
+
+        constexpr GLuint kRed = 0x000000ffu;   // RGBA8 little-endian: r = 255
+        constexpr GLuint kGreen = 0xff00ff00u; // what the shader stores: (0, 1, 0, 1)
+        constexpr int kTexels = 16;
+
+        FirstGLError();
+
+        const unsigned int program = MakeComputeProgram(kImageBufferCS);
+        ASSERT_NE(program, 0u);
+
+        const std::vector<GLuint> texels(kTexels, kRed);
+        GLuint buffer = 0;
+        glGenBuffers(1, &buffer);
+        glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+        glBufferData(GL_TEXTURE_BUFFER, static_cast<GLsizeiptr>(texels.size() * sizeof(GLuint)), texels.data(),
+                     GL_DYNAMIC_COPY);
+
+        GLuint texture = 0;
+        glGenTextures(1, &texture);
+        glBindTexture(GL_TEXTURE_BUFFER, texture);
+        glTexBuffer(GL_TEXTURE_BUFFER, GL_RGBA8, buffer);
+        EXPECT_EQ(FirstGLError(), 0u) << "glTexBuffer(GL_RGBA8) was refused";
+
+        glBindImageTexture(0, texture, 0, GL_FALSE, 0, GL_READ_WRITE, GL_RGBA8);
+        EXPECT_EQ(FirstGLError(), 0u) << "glBindImageTexture on a buffer texture was refused";
+
+        glUseProgram(program);
+        glDispatchCompute(1, 1, 1);
+        glMemoryBarrier(GL_ALL_BARRIER_BITS);
+
+        // Both CPU read paths, because they are two entry points onto the same refresh and a
+        // fix that reaches only one of them is not a fix. Everything below is EXPECT rather than
+        // ASSERT so that a failure still reaches the cleanup at the end: the harness shares one
+        // context across every scenario in the process, and a leaked buffer or image binding
+        // here would surface as a failure somewhere else entirely.
+        std::vector<GLuint> readBack(kTexels, 0u);
+        glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+        glGetBufferSubData(GL_TEXTURE_BUFFER, 0, static_cast<GLsizeiptr>(readBack.size() * sizeof(GLuint)),
+                           readBack.data());
+        EXPECT_EQ(readBack[0], kGreen) << "glGetBufferSubData did not see the imageStore";
+        EXPECT_EQ(readBack[2], kRed) << "the imageLoad side of the same dispatch read the wrong texel";
+
+        const void* mapped = glMapBuffer(GL_TEXTURE_BUFFER, GL_READ_ONLY);
+        EXPECT_NE(mapped, nullptr) << "glMapBuffer(GL_READ_ONLY) on the texture's buffer failed";
+        if (mapped != nullptr) {
+            GLuint mappedTexel0 = 0;
+            std::memcpy(&mappedTexel0, mapped, sizeof(mappedTexel0));
+            EXPECT_EQ(mappedTexel0, kGreen) << "glMapBuffer did not see the imageStore";
+            glUnmapBuffer(GL_TEXTURE_BUFFER);
+        }
+
+        glBindImageTexture(0, 0, 0, GL_FALSE, 0, GL_READ_ONLY, GL_RGBA8);
+        glBindBuffer(GL_TEXTURE_BUFFER, 0);
+        glBindTexture(GL_TEXTURE_BUFFER, 0);
+        glUseProgram(0);
+        glDeleteProgram(program);
+        glDeleteTextures(1, &texture);
+        glDeleteBuffers(1, &buffer);
+        EXPECT_EQ(FirstGLError(), 0u);
+    }
+
 } // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ClipDistanceScenario.cpp

@@ -0,0 +1,388 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ClipDistanceScenario.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 - gl_ClipDistance ACTUALLY CLIPS, AND ONLY WHERE IT IS ENABLED.
+//
+// CapabilityInput::ClipDistance0..7 existed end to end - the GL enum converted to it, the
+// string converter named it, glEnable(GL_CLIP_DISTANCE0 + i) raised no error - and then
+// RenderState::SetCapability had no case for it and dropped it into `default: break`. Nothing
+// was stored, no version was bumped, and neither backend ever heard about it. The shader half
+// worked all along (SPIRV-Cross emits gl_ClipDistance with a
+// `#extension GL_EXT_clip_cull_distance : require` that Adreno accepts), so the distances were
+// computed and then ignored: no clipping ever happened on DirectGLES, which is the whole of
+// KHR-GLxx.clip_distance.functional. glIsEnabled lied about it too - it returned GL_FALSE
+// immediately after a successful glEnable.
+//
+// The assertions are behavioural, not query-shaped, because a query-only test passes against a
+// backend that stores the bit and never forwards it. Each case draws one full-viewport triangle
+// whose clip distance is positive on one side of the viewport and negative on the other, then
+// checks BOTH sides: the kept side proves the draw happened at all, and the clipped side is the
+// actual claim. The disabled case is the negative control - the identical shader with the
+// identical distances and the enable turned off must leave both sides painted, which is what
+// says the pixels below are being removed by clipping and not by something else.
+//
+// HONEST LIMIT OF THIS FILE IN CI. Of the four cases, only EnableIsObservableThroughIsEnabled is
+// falsifiable on the software rasterizers every automated lane runs on. llvmpipe and lavapipe
+// clip by EVERY declared gl_ClipDistance regardless of the enables, so
+// AnEnabledClipDistanceRemovesTheNegativeHalf goes green there against the broken tree as well,
+// and the two cases that need real per-distance semantics skip (see
+// DriverHonoursPerDistanceEnables). What actually pins the behaviour is Adreno, through
+// KHR-GLxx.clip_distance.functional - whose "without dynamic redeclaration" variants declare all
+// gl_MaxClipDistances slots and enable only the first N, i.e. exactly the subset semantics these
+// skipped cases assert. Read a green CI run here as "the state survives the frontend", not as
+// "clipping is correct"; the second claim is a device claim.
+//
+// Every case disables all eight distances on entry rather than assuming they start off:
+// XfbAfterClipDistanceScenario deliberately leaves one enabled for the rest of the process, and
+// forwarding the enables is what turned that leftover from inert bookkeeping into live driver
+// state.
+
+#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
+#ifndef GL_CLIP_DISTANCE1
+#define GL_CLIP_DISTANCE1 0x3001
+#endif
+
+namespace MGITest {
+    namespace {
+
+        // One clip distance per half of the viewport: distance 0 is positive on the right half
+        // (x > 0 in clip space) and distance 1 is positive on the top half. A vertex shader
+        // producing a full-screen triangle from gl_VertexID, so no buffers are needed.
+        const char* const kVertexSource = R"(#version 400 core
+out float gl_ClipDistance[2];
+void main() {
+    vec2 positions[3] = vec2[3](vec2(-1.0, -1.0), vec2(3.0, -1.0), vec2(-1.0, 3.0));
+    vec2 p = positions[gl_VertexID];
+    gl_Position = vec4(p, 0.0, 1.0);
+    gl_ClipDistance[0] = p.x;
+    gl_ClipDistance[1] = p.y;
+}
+)";
+
+        const char* const kFragmentSource = R"(#version 400 core
+out vec4 fragColor;
+void main() { fragColor = vec4(0.0, 1.0, 0.0, 1.0); }
+)";
+
+        class ClipDistanceScenario : public ScenarioTest {
+        protected:
+            GLuint BuildProgram() {
+                const GLuint vs = glCreateShader(GL_VERTEX_SHADER);
+                glShaderSource(vs, 1, &kVertexSource, nullptr);
+                glCompileShader(vs);
+                GLint compiled = 0;
+                glGetShaderiv(vs, GL_COMPILE_STATUS, &compiled);
+                if (!compiled) {
+                    m_buildLog = ShaderLog(vs);
+                    glDeleteShader(vs);
+                    return 0;
+                }
+                const GLuint fs = glCreateShader(GL_FRAGMENT_SHADER);
+                glShaderSource(fs, 1, &kFragmentSource, nullptr);
+                glCompileShader(fs);
+                glGetShaderiv(fs, GL_COMPILE_STATUS, &compiled);
+                if (!compiled) {
+                    m_buildLog = ShaderLog(fs);
+                    glDeleteShader(vs);
+                    glDeleteShader(fs);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, vs);
+                glAttachShader(program, fs);
+                glLinkProgram(program);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                glDeleteShader(vs);
+                glDeleteShader(fs);
+                if (!linked) {
+                    GLint length = 0;
+                    glGetProgramiv(program, GL_INFO_LOG_LENGTH, &length);
+                    std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                    glGetProgramInfoLog(program, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                    m_buildLog = log.data();
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            const std::string& BuildLog() const { return m_buildLog; }
+
+            // Paints the whole viewport red, then draws the clipped triangle in green.
+            void DrawClippedTriangle(GLuint program, GLuint vao) const {
+                glClearColor(1.0f, 0.0f, 0.0f, 1.0f);
+                glClear(GL_COLOR_BUFFER_BIT);
+                glUseProgram(program);
+                glBindVertexArray(vao);
+                glDrawArrays(GL_TRIANGLES, 0, 3);
+            }
+
+            static bool IsGreen(const unsigned char* px) {
+                return px[0] < 64 && px[1] > 192;
+            }
+
+            static bool IsRed(const unsigned char* px) {
+                return px[0] > 192 && px[1] < 64;
+            }
+
+            void PixelAt(int x, int y, unsigned char* out) const {
+                glReadPixels(x, y, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, out);
+            }
+
+            // Never assume the eight start disabled - see the header note about
+            // XfbAfterClipDistanceScenario leaving one on for the rest of the process.
+            static void DisableEveryClipDistance() {
+                for (int i = 0; i < 8; ++i) {
+                    glDisable(static_cast<GLenum>(GL_CLIP_DISTANCE0 + i));
+                }
+            }
+
+            // True when the driver under this backend actually implements PER-DISTANCE enable
+            // state, i.e. when a written-but-disabled gl_ClipDistance leaves its fragments
+            // alone. Not every stack does, and the difference is not MobileGL's to hide:
+            //
+            //   - Adreno's ES driver honours GL_CLIP_DISTANCE0_EXT..7_EXT, which is what makes
+            //     KHR-GLxx.clip_distance.functional pass on the device once the enables are
+            //     forwarded at all.
+            //   - Vulkan has no such state: every clip distance a shader declares is active,
+            //     always. DirectVulkan therefore clips by a disabled distance.
+            //   - Mesa's llvmpipe ES driver behaves like Vulkan here.
+            //
+            // Emulating GL's semantics on those two would mean forcing the disabled slots to a
+            // non-negative value inside the shader, which makes the enable mask part of the
+            // pipeline key - a feature, not a fix, and deliberately not attempted here. The
+            // cases that need the real semantics gate on this probe and say so when they skip,
+            // rather than being deleted or silently weakened.
+            bool DriverHonoursPerDistanceEnables(GLuint program, GLuint vao) const {
+                for (int i = 0; i < 8; ++i) {
+                    glDisable(static_cast<GLenum>(GL_CLIP_DISTANCE0 + i));
+                }
+                DrawClippedTriangle(program, vao);
+                unsigned char negativeSide[4] = {0, 0, 0, 0};
+                glReadPixels(Gl().Width() / 4, Gl().Height() / 2, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, negativeSide);
+                return IsGreen(negativeSide);
+            }
+
+        private:
+            static std::string ShaderLog(GLuint shader) {
+                GLint length = 0;
+                glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+                std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                glGetShaderInfoLog(shader, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                return log.data();
+            }
+
+            std::string m_buildLog;
+        };
+
+    } // namespace
+
+    // The state itself: glEnable must be observable through glIsEnabled. This is the cheap half
+    // of the bug - SetCapability's missing case made the query answer GL_FALSE for a capability
+    // that had just been enabled without error.
+    TEST_F(ClipDistanceScenario, EnableIsObservableThroughIsEnabled) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        DisableEveryClipDistance();
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE0), GL_FALSE)
+            << "glDisable(GL_CLIP_DISTANCE0) is not observable through glIsEnabled";
+        glEnable(GL_CLIP_DISTANCE0);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE0), GL_TRUE)
+            << "glEnable(GL_CLIP_DISTANCE0) raised no error but glIsEnabled still reports it disabled";
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE1), GL_FALSE)
+            << "enabling distance 0 must not enable distance 1 - the eight are independent";
+
+        glEnable(GL_CLIP_DISTANCE1);
+        glDisable(GL_CLIP_DISTANCE0);
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE0), GL_FALSE);
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE1), GL_TRUE);
+
+        glDisable(GL_CLIP_DISTANCE1);
+        EXPECT_EQ(FirstGLError(), 0u);
+        gl.EndFrame();
+    }
+
+    // The claim: an enabled clip distance removes the fragments where it is negative.
+    TEST_F(ClipDistanceScenario, AnEnabledClipDistanceRemovesTheNegativeHalf) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        const GLuint program = BuildProgram();
+        ASSERT_NE(program, 0u) << "the gl_ClipDistance program did not build: " << BuildLog();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_CULL_FACE);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+
+        // Distance 1 is positive by a single pixel at the sampled row, so a stray enable on it
+        // would put the "kept" probe right on the clip boundary.
+        DisableEveryClipDistance();
+        glEnable(GL_CLIP_DISTANCE0);
+        DrawClippedTriangle(program, vao);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        unsigned char right[4] = {0, 0, 0, 0};
+        unsigned char left[4] = {0, 0, 0, 0};
+        PixelAt(width - 1 - width / 4, height / 2, right);
+        PixelAt(width / 4, height / 2, left);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        EXPECT_TRUE(IsGreen(right)) << "the kept half is not painted (" << int(right[0]) << "," << int(right[1])
+                                    << "," << int(right[2]) << ") - the draw itself did not happen, so the clipped "
+                                       "half below proves nothing";
+        EXPECT_TRUE(IsRed(left)) << "gl_ClipDistance[0] is negative on the left half and GL_CLIP_DISTANCE0 is "
+                                    "enabled, so those fragments must be clipped away; found ("
+                                 << int(left[0]) << "," << int(left[1]) << "," << int(left[2]) << ")";
+
+        glDisable(GL_CLIP_DISTANCE0);
+        glUseProgram(0);
+        glBindVertexArray(0);
+        glDeleteProgram(program);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+
+    // The negative control: the same shader writing the same distances, with the enable off,
+    // must paint both halves. Without this a backend that clipped everything - or one whose
+    // draw simply failed - would pass the case above.
+    TEST_F(ClipDistanceScenario, ADisabledClipDistanceRemovesNothing) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        const GLuint program = BuildProgram();
+        ASSERT_NE(program, 0u) << "the gl_ClipDistance program did not build: " << BuildLog();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_CULL_FACE);
+        DisableEveryClipDistance();
+
+        DrawClippedTriangle(program, vao);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        unsigned char right[4] = {0, 0, 0, 0};
+        unsigned char left[4] = {0, 0, 0, 0};
+        PixelAt(width - 1 - width / 4, height / 2, right);
+        PixelAt(width / 4, height / 2, left);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        EXPECT_TRUE(IsGreen(right)) << "with every clip distance disabled the whole triangle must survive";
+        const bool driverHonoursEnables = IsGreen(left);
+
+        glUseProgram(0);
+        glBindVertexArray(0);
+        glDeleteProgram(program);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+        if (!driverHonoursEnables) {
+            GTEST_SKIP() << "renderer " << gl.RendererString()
+                         << " clips by a DISABLED gl_ClipDistance - it does not implement per-distance enable state "
+                            "(see DriverHonoursPerDistanceEnables). Emulating GL's semantics there needs shader-side "
+                            "masking keyed on the enable mask, which is a separate feature";
+        }
+    }
+
+    // The eight enables are independent: enabling only distance 1 must clip by distance 1 and
+    // leave distance 0 alone. A backend that forwarded "any clip distance enabled" as a single
+    // bit, or that always enables every declared distance (which is what Vulkan does natively),
+    // passes both cases above and fails this one.
+    TEST_F(ClipDistanceScenario, TheEnablesAreIndependentPerDistance) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        const GLuint program = BuildProgram();
+        ASSERT_NE(program, 0u) << "the gl_ClipDistance program did not build: " << BuildLog();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_CULL_FACE);
+        if (!DriverHonoursPerDistanceEnables(program, vao)) {
+            glUseProgram(0);
+            glBindVertexArray(0);
+            glDeleteProgram(program);
+            glDeleteVertexArrays(1, &vao);
+            DisableEveryClipDistance();
+            gl.EndFrame();
+            GTEST_SKIP() << "renderer " << gl.RendererString()
+                         << " clips by every declared gl_ClipDistance regardless of the enables, so per-distance "
+                            "independence is not observable here";
+        }
+
+        DisableEveryClipDistance();
+        glEnable(GL_CLIP_DISTANCE1);
+
+        DrawClippedTriangle(program, vao);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        // Distance 1 is negative on the bottom half, distance 0 on the left half. With only
+        // distance 1 enabled, the bottom-left must survive (distance 0 is off) and the bottom
+        // must not.
+        unsigned char topLeft[4] = {0, 0, 0, 0};
+        unsigned char bottomRight[4] = {0, 0, 0, 0};
+        PixelAt(width / 4, height - 1 - height / 4, topLeft);
+        PixelAt(width - 1 - width / 4, height / 4, bottomRight);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        EXPECT_TRUE(IsGreen(topLeft)) << "gl_ClipDistance[0] is negative here but GL_CLIP_DISTANCE0 is disabled, so "
+                                         "this fragment must survive";
+        EXPECT_TRUE(IsRed(bottomRight)) << "gl_ClipDistance[1] is negative here and GL_CLIP_DISTANCE1 is enabled, so "
+                                           "this fragment must be clipped";
+
+        glDisable(GL_CLIP_DISTANCE1);
+        glUseProgram(0);
+        glBindVertexArray(0);
+        glDeleteProgram(program);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/CopyImageLayeredScenario.cpp

@@ -0,0 +1,331 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/CopyImageLayeredScenario.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 - glCopyImageSubData MOVES EVERY SLICE IT WAS ASKED FOR, NOT JUST SLICE 0.
+//
+// KHR-GL43.copy_image.functional_* copies a whole 12-layer region in one call whenever both
+// endpoints are layered, i.e. for the four target pairs 2d_array->2d_array, 2d_array->3d,
+// 3d->2d_array and 3d->3d. DirectVulkan built its VkImageCopy with baseArrayLayer 0, layerCount 1
+// and srcOffset.z 0 no matter what the call asked for, so slice 0 landed correctly and slices 1..N
+// were never written - 64 conformance cases (16 compatible format pairs x those 4 pairs) failing
+// with "first mismatch at [x, y, 1]", the first texel of the first slice the copy skipped.
+//
+// The reason one hardcode covered both shapes wrongly is that GL states a layered copy ONE way -
+// srcZ/dstZ and srcDepth - while Vulkan states it two ways and picks by image type:
+//
+//   GL_TEXTURE_3D       -> VK_IMAGE_TYPE_3D: slices are z, so srcOffset.z/dstOffset.z select them
+//                          and extent.depth counts them; the layer range must stay (0, 1).
+//   GL_TEXTURE_2D_ARRAY -> VK_IMAGE_TYPE_2D: slices are array layers, so baseArrayLayer selects
+//                          them and layerCount counts them; offset.z stays 0.
+//
+// A mixed pair is legal (maintenance1, core in Vulkan 1.1) but only when the counts correspond:
+// the 3D side's extent.depth has to equal the array side's layerCount. So the four pairs below are
+// four DIFFERENT VkImageCopy shapes, not one shape with different arguments, which is why one
+// scenario per pair is the coverage that matters here.
+//
+// Every case also asserts the slices OUTSIDE the copied range still hold their fill. A backend
+// that "fixed" the miss by copying the whole image regardless of srcZ/srcDepth would pass a
+// slices-landed check and fail this one.
+//
+// The verification path is an FBO attachment per slice plus glReadPixels, not glGetTexImage: it is
+// the readback both backends share, and glFramebufferTextureLayer names an array layer and a 3D
+// slice through the same call, so the two texture kinds are read back identically.
+//
+// DirectGLES is the control - it forwards to the driver's own glCopyImageSubData - so a failure on
+// both backends means the scenario is wrong, and a failure on DirectVulkan alone means Magma is.
+
+#include <array>
+#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 int kWidth = 4;
+        constexpr int kHeight = 4;
+        // Six is enough for a copy that starts and ends away from both edges of both endpoints
+        // while still leaving untouched slices on either side to assert against.
+        constexpr int kSlices = 6;
+
+        struct Rgba8 {
+            GLubyte 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;
+            }
+        };
+
+        std::string Describe(const Rgba8& color) {
+            return "(" + std::to_string(color.r) + ", " + std::to_string(color.g) + ", " + std::to_string(color.b) +
+                   ", " + std::to_string(color.a) + ")";
+        }
+
+        // Per-slice constants, uniform within a slice. A uniform fill is deliberate: the defect is
+        // in which SLICE the copy addresses, and a value that also varied within the slice would
+        // make the assertions depend on the framebuffer row order as well.
+        Rgba8 SourceColor(int slice) {
+            return {static_cast<GLubyte>(10 + slice * 20), static_cast<GLubyte>(40 + slice * 10),
+                    static_cast<GLubyte>(200 - slice * 15), 255};
+        }
+
+        Rgba8 DestinationFill(int slice) {
+            return {static_cast<GLubyte>(3 + slice), static_cast<GLubyte>(250 - slice * 7),
+                    static_cast<GLubyte>(120 + slice * 5), 255};
+        }
+
+        class CopyImageLayeredScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                if (!CopyImageSubDataUsable()) {
+                    GTEST_SKIP() << "glCopyImageSubData is unavailable on backend " << Gl().BackendName();
+                }
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                for (const GLuint texture : m_textures) {
+                    glDeleteTextures(1, &texture);
+                }
+                m_textures.clear();
+                if (m_fbo != 0) {
+                    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    glDeleteFramebuffers(1, &m_fbo);
+                    m_fbo = 0;
+                }
+            }
+
+            // A trivial 1x1x1 array-to-array copy: it exercises the entry point without depending
+            // on any of the behaviour under test, so a driver (or a backend function table) that
+            // simply does not have the call skips instead of failing every case below.
+            bool CopyImageSubDataUsable() {
+                GLuint probe[2] = {0, 0};
+                glGenTextures(2, probe);
+                for (const GLuint texture : probe) {
+                    glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                    glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_RGBA8, 1, 1, 1);
+                }
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                glCopyImageSubData(probe[0], GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, probe[1], GL_TEXTURE_2D_ARRAY, 0, 0, 0,
+                                   0, 1, 1, 1);
+                const bool usable = glGetError() == GL_NO_ERROR;
+                glDeleteTextures(2, probe);
+                return usable;
+            }
+
+            // `target` is GL_TEXTURE_2D_ARRAY or GL_TEXTURE_3D; both take glTexStorage3D and
+            // glTexSubImage3D with the slice on the same axis, which is the whole reason GL can
+            // copy between them. `levels` > 1 puts a real mip chain behind the level the copy
+            // names, so the level's own extent - a 3D level's depth included - has to be resolved
+            // rather than assumed to be the image's.
+            GLuint MakeTexture(GLenum target, int levels, Rgba8 (*colorForSlice)(int)) {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(target, texture);
+                glTexStorage3D(target, levels, GL_RGBA8, kWidth << (levels - 1), kHeight << (levels - 1),
+                               target == GL_TEXTURE_3D ? (kSlices << (levels - 1)) : kSlices);
+                glTexParameteri(target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+
+                // Fill every level, so nothing below can pass by reading a level that was never
+                // written and happened to hold the expected bytes.
+                for (int level = 0; level < levels; ++level) {
+                    const int levelWidth = kWidth << (levels - 1 - level);
+                    const int levelHeight = kHeight << (levels - 1 - level);
+                    const int levelSlices =
+                        target == GL_TEXTURE_3D ? (kSlices << (levels - 1 - level)) : kSlices;
+                    for (int slice = 0; slice < levelSlices; ++slice) {
+                        const Rgba8 color = colorForSlice(slice % kSlices);
+                        std::vector<Rgba8> texels(static_cast<size_t>(levelWidth) * levelHeight, color);
+                        glTexSubImage3D(target, level, 0, 0, slice, levelWidth, levelHeight, 1, GL_RGBA,
+                                        GL_UNSIGNED_BYTE, texels.data());
+                    }
+                }
+                glBindTexture(target, 0);
+                return texture;
+            }
+
+            // One slice of one level, through an FBO attachment. glFramebufferTextureLayer takes an
+            // array layer and a 3D slice through the same argument, so both targets read back the
+            // same way.
+            Rgba8 ReadSlice(GLuint texture, int level, int slice, int width, int height) {
+                if (m_fbo == 0) {
+                    glGenFramebuffers(1, &m_fbo);
+                }
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, texture, level, slice);
+                EXPECT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                    << "slice " << slice << " of level " << level << " is not attachable";
+                std::vector<Rgba8> pixels(static_cast<size_t>(width) * height, Rgba8{});
+                glReadBuffer(GL_COLOR_ATTACHMENT0);
+                glPixelStorei(GL_PACK_ALIGNMENT, 1);
+                glReadPixels(0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+
+                // The fill is uniform within a slice, so any disagreement between texels is itself
+                // a failure - reported here rather than silently reduced to pixels[0].
+                for (size_t i = 1; i < pixels.size(); ++i) {
+                    EXPECT_TRUE(pixels[i] == pixels[0])
+                        << "slice " << slice << " of level " << level << " is not uniform: texel 0 is "
+                        << Describe(pixels[0]) << ", texel " << i << " is " << Describe(pixels[i]);
+                }
+                return pixels[0];
+            }
+
+            // The assertion every case ends with: slices inside [dstZ, dstZ + depth) hold the
+            // source slice they were fed, and every slice outside it still holds its own fill.
+            void ExpectCopied(GLuint destination, int level, int width, int height, int sliceCount, int srcZ,
+                              int dstZ, int depth, const char* what) {
+                for (int slice = 0; slice < sliceCount; ++slice) {
+                    const bool inRange = slice >= dstZ && slice < dstZ + depth;
+                    const Rgba8 expected =
+                        inRange ? SourceColor(srcZ + (slice - dstZ)) : DestinationFill(slice);
+                    const Rgba8 actual = ReadSlice(destination, level, slice, width, height);
+                    EXPECT_TRUE(actual == expected)
+                        << what << ": destination slice " << slice << (inRange ? " (copied)" : " (untouched)")
+                        << " is " << Describe(actual) << ", expected " << Describe(expected);
+                }
+            }
+
+            std::vector<GLuint> m_textures;
+            GLuint m_fbo = 0;
+        };
+
+        // 2d_array -> 2d_array. Both endpoints put the slices on the layer axis, so BOTH layer
+        // counts carry the depth and extent.depth must stay 1.
+        TEST_F(CopyImageLayeredScenario, ArrayToArrayCopiesEverySlice) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            glCopyImageSubData(source, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, destination, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0,
+                               kWidth, kHeight, kSlices);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, 0, 0, kSlices, "array->array, all slices");
+        }
+
+        // The same pair with the layer ranges offset differently on the two sides: the shape that
+        // separates "copies more than slice 0" from "copies the RIGHT slices". A backend that read
+        // the source range but wrote from layer 0 (or vice versa) passes the case above.
+        TEST_F(CopyImageLayeredScenario, ArrayToArrayHonoursDifferentLayerOffsets) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 3;
+            constexpr int kDstZ = 1;
+            constexpr int kDepth = 2;
+            glCopyImageSubData(source, GL_TEXTURE_2D_ARRAY, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_2D_ARRAY, 0, 0, 0,
+                               kDstZ, kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth,
+                         "array->array, offset layer ranges");
+        }
+
+        // 3d -> 3d. Neither endpoint has array layers at all: the depth travels on extent.depth and
+        // the offsets on srcOffset.z/dstOffset.z, with both layer counts pinned to 1.
+        TEST_F(CopyImageLayeredScenario, VolumeToVolumeHonoursNonZeroZ) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_3D, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_3D, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 1;
+            constexpr int kDstZ = 3;
+            constexpr int kDepth = 3;
+            glCopyImageSubData(source, GL_TEXTURE_3D, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_3D, 0, 0, 0, kDstZ,
+                               kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth, "3d->3d, non-zero z");
+        }
+
+        // The same pair one mip level down. A 3D level's DEPTH halves with its width and height, so
+        // this is the only case where the slice count the copy may name is not the image's own -
+        // the bound a layered endpoint is checked against has to come from the level.
+        TEST_F(CopyImageLayeredScenario, VolumeToVolumeAtNonZeroMipLevel) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_3D, 2, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_3D, 2, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kLevel = 1;
+            constexpr int kSrcZ = 2;
+            constexpr int kDstZ = 0;
+            constexpr int kDepth = 4;
+            glCopyImageSubData(source, GL_TEXTURE_3D, kLevel, 0, 0, kSrcZ, destination, GL_TEXTURE_3D, kLevel, 0, 0,
+                               kDstZ, kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, kLevel, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth,
+                         "3d->3d at mip level 1");
+        }
+
+        // 2d_array -> 3d. The mixed shape: the source counts its slices as layers, the destination
+        // as depth, and Vulkan requires extent.depth to equal the source's layerCount.
+        TEST_F(CopyImageLayeredScenario, ArrayToVolumeCopiesEverySlice) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_3D, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 2;
+            constexpr int kDstZ = 1;
+            constexpr int kDepth = 4;
+            glCopyImageSubData(source, GL_TEXTURE_2D_ARRAY, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_3D, 0, 0, 0, kDstZ,
+                               kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth, "2d_array->3d");
+        }
+
+        // 3d -> 2d_array, the mirror image: the depth now has to reach the DESTINATION's layerCount
+        // while the source states it as extent.depth from a z offset.
+        TEST_F(CopyImageLayeredScenario, VolumeToArrayCopiesEverySlice) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_3D, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 1;
+            constexpr int kDstZ = 2;
+            constexpr int kDepth = 4;
+            glCopyImageSubData(source, GL_TEXTURE_3D, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_2D_ARRAY, 0, 0, 0, kDstZ,
+                               kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth, "3d->2d_array");
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/CopyImageLevelRangeScenario.cpp

@@ -0,0 +1,209 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/CopyImageLevelRangeScenario.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
+//
+// KHR-GL43.copy_image.non_existent_mipmap, and what it cost.
+//
+// The CTS case is a pure negative test: two 16x16 textures that have level 0 and
+// nothing else, and a glCopyImageSubData naming level 1. The answer is
+// GL_INVALID_VALUE (GL 4.6 core 18.3.2 / ARB_copy_image: "srcLevel/dstLevel is not
+// a valid level"). MobileGL's frontend only checked the level against
+// GL_MAX_TEXTURE_SIZE, so level 1 sailed through into the backends, DirectVulkan
+// resolved it into a VkImageCopy subresource on a VkImage that was created with
+// exactly one mip level, and the Adreno driver dereferenced the level it was
+// promised - SIGSEGV inside vkCmdCopyImage, taking the whole glcts process down
+// mid-run. A negative case must never do that.
+//
+// So the level-1-on-a-one-level-texture rejection is the regression proper, and the
+// rest of this file is what keeps the fix honest. A validator that answered
+// GL_INVALID_VALUE to every level would satisfy the regression tests alone, so the
+// scenarios below pin the BOUNDARY rather than the symptom:
+//
+//   * a texture that really does have two levels must accept a copy at level 1,
+//   * the same texture must still reject level 2,
+//   * and a plain level-0 copy must move pixels, which is checked by reading the
+//     destination back rather than by trusting glGetError.
+//
+// Both backends are covered because the fix is in the shared frontend: DirectGLES
+// forwards to the ES glCopyImageSubData (whose own error lands in the ES context,
+// not in MobileGL's, so it never reached the application either) and DirectVulkan
+// records the copy itself.
+
+#include <array>
+#include <cstring>
+#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 GLsizei kSize = 16;
+
+        struct Rgba8 {
+            GLubyte r, g, b, a;
+            bool operator==(const Rgba8& other) const {
+                return r == other.r && g == other.g && b == other.b && a == other.a;
+            }
+        };
+
+        std::vector<Rgba8> SolidImage(GLsizei width, GLsizei height, Rgba8 color) {
+            return std::vector<Rgba8>(static_cast<std::size_t>(width) * static_cast<std::size_t>(height), color);
+        }
+
+        class CopyImageLevelRangeScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                DrainErrors();
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                DeleteTextures();
+                if (m_fbo != 0) {
+                    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    glDeleteFramebuffers(1, &m_fbo);
+                    m_fbo = 0;
+                }
+                DrainErrors();
+                ScenarioTest::TearDown();
+            }
+
+            static void DrainErrors() {
+                for (int i = 0; i < 16 && glGetError() != GL_NO_ERROR; ++i) {
+                }
+            }
+
+            void DeleteTextures() {
+                if (m_src != 0) glDeleteTextures(1, &m_src);
+                if (m_dst != 0) glDeleteTextures(1, &m_dst);
+                m_src = 0;
+                m_dst = 0;
+            }
+
+            // One 16x16 RGBA8 texture with `levelCount` levels defined through
+            // glTexImage2D - the same way the CTS case builds its textures, and
+            // deliberately NOT glTexStorage2D: an immutable allocation would define the
+            // whole chain up front and could not express "level 1 does not exist".
+            GLuint MakeTexture(int levelCount, Rgba8 baseColor) {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                glBindTexture(GL_TEXTURE_2D, texture);
+                for (int level = 0; level < levelCount; ++level) {
+                    const GLsizei extent = kSize >> level;
+                    const std::vector<Rgba8> pixels = SolidImage(extent, extent, baseColor);
+                    glTexImage2D(GL_TEXTURE_2D, level, GL_RGBA8, extent, extent, 0, GL_RGBA, GL_UNSIGNED_BYTE,
+                                 pixels.data());
+                }
+                // What Utils::makeTextureComplete does in the CTS case: the texture is
+                // complete for the levels it actually has, not for a chain it does not.
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, levelCount - 1);
+                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;
+            }
+
+            void MakePair(int levelCount) {
+                DeleteTextures();
+                m_src = MakeTexture(levelCount, Rgba8{11, 22, 33, 255});
+                m_dst = MakeTexture(levelCount, Rgba8{200, 100, 50, 255});
+                ASSERT_EQ(glGetError(), GL_NO_ERROR) << "texture setup with " << levelCount << " level(s)";
+            }
+
+            // The call under test, at whatever levels the caller wants, over a 1x1
+            // region so the region check can never be what rejects it.
+            GLenum CopyAt(GLint srcLevel, GLint dstLevel, GLsizei extent = 1) {
+                DrainErrors();
+                glCopyImageSubData(m_src, GL_TEXTURE_2D, srcLevel, 0, 0, 0, m_dst, GL_TEXTURE_2D, dstLevel, 0, 0, 0,
+                                   extent, extent, 1);
+                const GLenum error = glGetError();
+                // A second pending error would mean the entry point queued more than one,
+                // and the extra would be handed out at an unrelated call site later.
+                EXPECT_EQ(glGetError(), GL_NO_ERROR) << "the copy recorded more than one error";
+                return error;
+            }
+
+            Rgba8 ReadBackDestinationLevel0() {
+                if (m_fbo == 0) glGenFramebuffers(1, &m_fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, m_dst, 0);
+                const GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+                if (status != GL_FRAMEBUFFER_COMPLETE) {
+                    ADD_FAILURE() << "readback framebuffer incomplete: " << status;
+                    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    return Rgba8{0, 0, 0, 0};
+                }
+                Rgba8 texel{0, 0, 0, 0};
+                glReadPixels(0, 0, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, &texel);
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                return texel;
+            }
+
+            GLuint m_src = 0;
+            GLuint m_dst = 0;
+            GLuint m_fbo = 0;
+        };
+
+        // The regression. Level 1 of a texture that has only level 0 is not a level, and
+        // saying so is the whole job: before the fix this reached DirectVulkan, which
+        // handed mipLevel=1 to vkCmdCopyImage on a one-level VkImage and died inside the
+        // Adreno driver.
+        TEST_F(CopyImageLevelRangeScenario, LevelOneOfASingleLevelTextureIsRejected) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(1);
+
+            EXPECT_EQ(CopyAt(1, 0), static_cast<GLenum>(GL_INVALID_VALUE)) << "source level 1";
+            EXPECT_EQ(CopyAt(0, 1), static_cast<GLenum>(GL_INVALID_VALUE)) << "destination level 1";
+            EXPECT_EQ(CopyAt(1, 1), static_cast<GLenum>(GL_INVALID_VALUE)) << "both levels 1";
+        }
+
+        // The negative control that makes the test above falsifiable: the same level
+        // index, on textures that genuinely have it, must be accepted. A validator that
+        // rejected every non-zero level would pass the regression test and fail here.
+        TEST_F(CopyImageLevelRangeScenario, LevelOneOfATwoLevelTextureIsAccepted) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(2);
+
+            EXPECT_EQ(CopyAt(1, 1), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        // And the boundary from the other side: two levels means 0 and 1, not 2.
+        TEST_F(CopyImageLevelRangeScenario, LevelTwoOfATwoLevelTextureIsRejected) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(2);
+
+            EXPECT_EQ(CopyAt(2, 0), static_cast<GLenum>(GL_INVALID_VALUE)) << "source level 2";
+            EXPECT_EQ(CopyAt(0, 2), static_cast<GLenum>(GL_INVALID_VALUE)) << "destination level 2";
+        }
+
+        // Errors alone cannot tell an accepted copy from a silently dropped one, so the
+        // ordinary case is checked by reading the destination back: the copy has to move
+        // the source's texel, not merely decline to complain.
+        TEST_F(CopyImageLevelRangeScenario, AValidLevelZeroCopyStillMovesPixels) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(1);
+
+            ASSERT_EQ(ReadBackDestinationLevel0(), (Rgba8{200, 100, 50, 255})) << "destination before the copy";
+            EXPECT_EQ(CopyAt(0, 0, kSize), static_cast<GLenum>(GL_NO_ERROR));
+            EXPECT_EQ(ReadBackDestinationLevel0(), (Rgba8{11, 22, 33, 255})) << "destination after the copy";
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackAttachmentShapeScenario.cpp

@@ -0,0 +1,370 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackAttachmentShapeScenario.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 - DEPTH/STENCIL READBACK WHEN THE ATTACHMENT IS NOT A PLAIN GL_TEXTURE_2D,
+// AND THE DEFAULT FRAMEBUFFER'S ADVERTISED DEPTH/STENCIL FORMAT.
+//
+// Three shipped defects, all of them invisible to a test that only ever attaches a 2D texture
+// or only ever asks the default framebuffer for a colour value.
+//
+// (1) The ES depth/stencil readback emulation identifies the source format by binding the
+//     attachment's texture NAME to GL_TEXTURE_2D and asking that target for its internal
+//     format. A name whose target is GL_TEXTURE_2D_ARRAY (attached by
+//     glFramebufferTextureLayer) makes the bind answer GL_INVALID_OPERATION and change
+//     nothing - so the query then truthfully describes whatever texture was already on
+//     GL_TEXTURE_2D, which on that path is the emulation's own staging scratch. A wrong
+//     answer that looks like a right one: the staging blit is issued between mismatched
+//     depth formats, ES rejects it, and the read reports nothing at all.
+//
+// (2) Adreno answers GL_NONE for GL_FRAMEBUFFER_ATTACHMENT_OBJECT_TYPE on an attachment made
+//     by glFramebufferTexture (a cube map, attached layered) while still reporting its depth
+//     and stencil bits correctly. The emulation took OBJECT_TYPE as the sole witness for "is
+//     there an aspect here at all" and declined the whole read.
+//
+// (3) DirectGLES never told the frontend what its default framebuffer's depth/stencil format
+//     actually is, so the placeholder from MG_Impl/Init.cpp - GL_DEPTH32F_STENCIL8 - was what
+//     every attachment query answered, whatever the surface really had. That is not cosmetic:
+//     GL blits depth/stencil only between IDENTICAL formats, so an application that reads
+//     GL_FRAMEBUFFER_ATTACHMENT_DEPTH_SIZE, allocates the buffer it was just told about and
+//     blits gets GL_INVALID_OPERATION - and a rejected glBlitFramebuffer transfers NOTHING,
+//     colour bits included. DirectVulkan has published its real format since the swapchain
+//     work; this is the half that was missing.
+//
+// Every case poisons its destination with a value the correct answer cannot be, so "the
+// backend wrote nothing" fails loudly instead of passing on stale memory. The plain
+// GL_TEXTURE_2D case at the end is the built-in control: it shares every line of the readback
+// path with the array and cube cases, so its passing is what says a failure above is about the
+// attachment's SHAPE and not about depth readback in general.
+//
+// The scenario name starts with DepthStencilReadback on purpose - that is the filter the
+// forced-emulation ctest registration uses (MG_IntegrationTest/CMakeLists.txt), and without
+// that registration these cases are unfalsifiable on llvmpipe, which accepts the native ES
+// depth reads that the Adreno device does not have.
+
+#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 float kDepthPoison = 0.2f;
+        constexpr int kStencilPoison = 50;
+        constexpr float kDepthValue = 0.75f;
+        constexpr int kStencilValue = 7;
+        constexpr int kSize = 16;
+
+        class DepthStencilReadbackAttachmentShapeScenario : public ScenarioTest {
+        protected:
+            float ReadDepthAt(int x, int y) const {
+                float depth = kDepthPoison;
+                glReadPixels(x, y, 1, 1, GL_DEPTH_COMPONENT, GL_FLOAT, &depth);
+                return depth;
+            }
+
+            int ReadStencilAt(int x, int y) const {
+                int stencil = kStencilPoison;
+                glReadPixels(x, y, 1, 1, GL_STENCIL_INDEX, GL_INT, &stencil);
+                return stencil;
+            }
+
+            // Clears the currently bound framebuffer's depth and stencil to the shared
+            // reference values, with both write masks explicitly open (glClear honours them,
+            // and a leftover mask from another scenario in this shared context would look
+            // exactly like the bug under test).
+            void ClearDepthStencil() const {
+                glDepthMask(GL_TRUE);
+                glStencilMask(0xFFu);
+                glDisable(GL_SCISSOR_TEST);
+                glClearDepth(kDepthValue);
+                glClearStencil(kStencilValue);
+                glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+            }
+        };
+
+        // Fails the calling test if the framebuffer bound at both targets is not complete;
+        // an incomplete framebuffer would make every read below return the poison for a
+        // reason that has nothing to do with what is being tested.
+        ::testing::AssertionResult FramebufferIsComplete() {
+            const GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+            if (status == GL_FRAMEBUFFER_COMPLETE) return ::testing::AssertionSuccess();
+            return ::testing::AssertionFailure() << "framebuffer status 0x" << std::hex << status;
+        }
+
+    } // namespace
+
+    // (1) A depth slice of a 2D ARRAY texture, attached with glFramebufferTextureLayer.
+    // Pre-fix this read back the poison: the format probe answered with the staging scratch's
+    // GL_DEPTH24_STENCIL8 instead of the array's GL_DEPTH_COMPONENT24, and the mismatched
+    // staging blit was rejected.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DepthOfAnArrayLayerAttachmentReadsBack) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint depthArray = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &depthArray);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, depthArray);
+        glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_DEPTH_COMPONENT24, kSize, kSize, 4);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        // Layer 2, not layer 0: a backend that silently reads the wrong slice would still
+        // agree with a single-layer texture.
+        glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depthArray, 0, 2);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "glReadPixels(GL_DEPTH_COMPONENT) of a GL_TEXTURE_2D_ARRAY layer attachment returned " << depth
+            << (std::fabs(depth - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &depthArray);
+        gl.EndFrame();
+    }
+
+    // (2) A depth cube map, attached whole with glFramebufferTexture - a LAYERED attachment.
+    // Pre-fix the emulation declined outright, because the driver reports GL_NONE for that
+    // attachment's OBJECT_TYPE.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DepthOfALayeredCubeAttachmentReadsBack) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint depthCube = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &depthCube);
+        glBindTexture(GL_TEXTURE_CUBE_MAP, depthCube);
+        glTexStorage2D(GL_TEXTURE_CUBE_MAP, 1, GL_DEPTH_COMPONENT24, kSize, kSize);
+        glBindTexture(GL_TEXTURE_CUBE_MAP, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depthCube, 0);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "glReadPixels(GL_DEPTH_COMPONENT) of a layered GL_TEXTURE_CUBE_MAP attachment returned " << depth
+            << (std::fabs(depth - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &depthCube);
+        gl.EndFrame();
+    }
+
+    // Both aspects of a packed array attachment. The stencil half goes through a different
+    // sampling mode than the depth half, and only the depth half was covered above.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, PackedArrayLayerAttachmentReadsBackBothAspects) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint packedArray = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &packedArray);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, packedArray);
+        glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_DEPTH24_STENCIL8, kSize, kSize, 3);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, packedArray, 0, 1);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        const int stencil = ReadStencilAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "depth of a packed GL_TEXTURE_2D_ARRAY layer attachment returned " << depth;
+        EXPECT_EQ(stencil, kStencilValue)
+            << "stencil of a packed GL_TEXTURE_2D_ARRAY layer attachment returned " << stencil
+            << (stencil == kStencilPoison ? " - the destination was never written at all" : "");
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &packedArray);
+        gl.EndFrame();
+    }
+
+    // The control: the plain GL_TEXTURE_2D shape, which always worked. If this one ever fails
+    // alongside the three above, the fault is in depth readback generally rather than in how
+    // the attachment's format and presence are discovered.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DepthOfAPlainTexture2DAttachmentReadsBack) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint depthTex = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &depthTex);
+        glBindTexture(GL_TEXTURE_2D, depthTex);
+        glTexStorage2D(GL_TEXTURE_2D, 1, GL_DEPTH_COMPONENT24, kSize, kSize);
+        glBindTexture(GL_TEXTURE_2D, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, depthTex, 0);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "the control case failed: even a plain GL_TEXTURE_2D depth attachment read back " << depth;
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &depthTex);
+        gl.EndFrame();
+    }
+
+    // (3) The default framebuffer must describe its depth/stencil truthfully enough that a
+    // buffer allocated from that description is blit-compatible with it. This is the exact
+    // sequence KHR-GLxx.framebuffer_blit performs, and the exact reason 22 of its cases died
+    // on DirectGLES: the frontend answered 32-bit float depth for a 24-bit fixed-point
+    // surface, so the renderbuffer the caller allocated could never be blitted to.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DefaultFramebufferDepthStencilFormatIsBlitCompatible) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        BindDefaultFramebuffer();
+        GLint depthBits = 0;
+        GLint stencilBits = 0;
+        GLint componentType = GL_UNSIGNED_NORMALIZED;
+        glGetFramebufferAttachmentParameteriv(GL_DRAW_FRAMEBUFFER, GL_DEPTH,
+                                              GL_FRAMEBUFFER_ATTACHMENT_DEPTH_SIZE, &depthBits);
+        glGetFramebufferAttachmentParameteriv(GL_DRAW_FRAMEBUFFER, GL_STENCIL,
+                                              GL_FRAMEBUFFER_ATTACHMENT_STENCIL_SIZE, &stencilBits);
+        glGetFramebufferAttachmentParameteriv(GL_DRAW_FRAMEBUFFER, GL_DEPTH,
+                                              GL_FRAMEBUFFER_ATTACHMENT_COMPONENT_TYPE, &componentType);
+        EXPECT_EQ(FirstGLError(), 0u);
+        if (depthBits <= 0 || stencilBits <= 0) {
+            GTEST_SKIP() << "this surface has no packed depth/stencil (depth=" << depthBits
+                         << " stencil=" << stencilBits << "); the blit-compatibility contract needs both";
+        }
+
+        // The one sized format the reported description names. Getting here with the wrong
+        // answer is the bug: the two candidates are not interchangeable for a blit.
+        const GLenum reported = (componentType == GL_FLOAT || depthBits > 24) ? GL_DEPTH32F_STENCIL8
+                                                                             : GL_DEPTH24_STENCIL8;
+
+        GLuint fbo = 0;
+        GLuint colorRbo = 0;
+        GLuint depthRbo = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenRenderbuffers(1, &colorRbo);
+        glGenRenderbuffers(1, &depthRbo);
+        glBindRenderbuffer(GL_RENDERBUFFER, colorRbo);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, width, height);
+        glBindRenderbuffer(GL_RENDERBUFFER, depthRbo);
+        glRenderbufferStorage(GL_RENDERBUFFER, reported, width, height);
+        glBindRenderbuffer(GL_RENDERBUFFER, 0);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, colorRbo);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_RENDERBUFFER, depthRbo);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        // Put a known depth in the default framebuffer, then blit colour+depth+stencil out of
+        // it into the buffer that its own description asked for.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+        glClearColor(0.0f, 1.0f, 0.0f, 1.0f);
+        glClear(GL_COLOR_BUFFER_BIT);
+        ClearDepthStencil();
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, 0);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, fbo);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height,
+                          GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT, GL_NEAREST);
+        EXPECT_EQ(FirstGLError(), 0u)
+            << "blitting depth/stencil out of the default framebuffer into a buffer allocated from the format "
+               "the default framebuffer itself reported was rejected - the report and the storage disagree";
+
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, fbo);
+        unsigned char color[4] = {0, 0, 0, 0};
+        glReadPixels(width / 2, height / 2, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, color);
+        const float depth = ReadDepthAt(width / 2, height / 2);
+        const int stencil = ReadStencilAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        // The colour bit is the precondition, not the claim: it says this stack can blit out of
+        // its default framebuffer at all, which has nothing to do with depth/stencil formats.
+        // DirectVulkan on a surfaceless pbuffer cannot - the whole call, colour included, is a
+        // no-op there, while the same blit works on a real surface (KHR-GLxx.framebuffer_blit
+        // exercises exactly it and Magma passes 33/33 on device). Skipping keeps the
+        // depth/stencil claim below falsifiable instead of drowning it in an unrelated
+        // harness limitation.
+        if (int(color[1]) <= 192) {
+            // GTEST_SKIP() expands to a return, so the teardown below it would never run and this
+            // scenario would hand the next one a foreign framebuffer plus three leaked objects -
+            // and this is the path DirectVulkan takes on every headless run, not a rare one.
+            BindDefaultFramebuffer();
+            glDeleteFramebuffers(1, &fbo);
+            glDeleteRenderbuffers(1, &colorRbo);
+            glDeleteRenderbuffers(1, &depthRbo);
+            gl.EndFrame();
+            GTEST_SKIP() << "backend " << gl.BackendName() << " on this surface transferred no colour either (green="
+                         << int(color[1])
+                         << "): it cannot blit out of the default framebuffer here, so the depth/stencil half proves "
+                            "nothing. The GL-error assertion above still ran, and it is the format contract";
+        }
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "depth blitted out of the default framebuffer read back " << depth
+            << (std::fabs(depth - kDepthPoison) < 1e-6f ? " - the blit transferred nothing" : "");
+        EXPECT_EQ(stencil, kStencilValue) << "stencil blitted out of the default framebuffer read back " << stencil;
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteRenderbuffers(1, &colorRbo);
+        glDeleteRenderbuffers(1, &depthRbo);
+        gl.EndFrame();
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackMatrixScenario.cpp

@@ -0,0 +1,784 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackMatrixScenario.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 - THE DEPTH/STENCIL READBACK MATRIX: every verb, every source kind.
+//
+// DepthStencilReadbackScenario pins the default framebuffer. This file pins the rest of
+// the surface a depth/stencil read has to cover, because the three verbs and the four
+// source kinds do NOT share a code path by accident - they share one on purpose, and a
+// change that quietly serves only one of them is exactly what these assertions catch:
+//
+//   verbs         glReadPixels(GL_DEPTH_COMPONENT | GL_STENCIL_INDEX | GL_DEPTH_STENCIL),
+//                 glGetTexImage(GL_DEPTH_STENCIL), glCopyTexImage2D followed by a read
+//   source kinds  depth(-stencil) TEXTURE, RENDERBUFFER (not samplable at all),
+//                 MULTISAMPLE renderbuffer (needs a resolve first), default framebuffer
+//   formats       DEPTH24_STENCIL8, DEPTH32F_STENCIL8, DEPTH_COMPONENT16/24/32F,
+//                 STENCIL_INDEX8
+//   client types  GL_FLOAT / GL_UNSIGNED_INT / GL_UNSIGNED_SHORT depth, GL_INT /
+//                 GL_UNSIGNED_BYTE stencil, both packed GL_DEPTH_STENCIL layouts
+//
+// On DirectGLES none of this exists natively - ES has no depth or stencil readback in
+// core - so every assertion here is really an assertion about the shader-sampling
+// emulation. The catch is that some ES drivers accept the reads anyway (Mesa does,
+// Adreno does not), which would make the emulation dead code on the very stack the
+// headless suite runs on. That is what the second ctest registration is for: the same
+// scenarios run again with MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION=1, which takes the
+// native spellings off the table and leaves only the path the device actually uses.
+//
+// Every destination is poisoned with a value the correct answer cannot be, so "the
+// backend wrote nothing" fails loudly instead of passing on a coincidence - a test that
+// only checked "no GL error" would pass against a readback that never touched the buffer,
+// which is precisely how this whole cluster hid for so long.
+
+#include <cmath>
+#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 float kDepthPoison = 0.2f;
+        constexpr int kStencilPoison = 50;
+        constexpr int kWidth = 64;
+        constexpr int kHeight = 48;
+
+        // A depth-stencil pair no clear in these tests produces, packed both ways.
+        constexpr unsigned int kPacked24_8Poison = 0xAAAAAA33u;
+
+        struct D32fS8 {
+            float depth;
+            unsigned int stencil;
+        };
+
+        // Everything a source needs to be read: the framebuffer to bind, plus the objects
+        // to delete afterwards.
+        struct DepthSource {
+            GLuint fbo = 0;
+            GLuint colorTexture = 0;
+            GLuint depthTexture = 0;
+            GLuint depthRenderbuffer = 0;
+            GLuint colorRenderbuffer = 0;
+        };
+
+        void DestroySource(DepthSource& source) {
+            if (source.fbo != 0) glDeleteFramebuffers(1, &source.fbo);
+            if (source.colorTexture != 0) glDeleteTextures(1, &source.colorTexture);
+            if (source.depthTexture != 0) glDeleteTextures(1, &source.depthTexture);
+            if (source.depthRenderbuffer != 0) glDeleteRenderbuffers(1, &source.depthRenderbuffer);
+            if (source.colorRenderbuffer != 0) glDeleteRenderbuffers(1, &source.colorRenderbuffer);
+            source = DepthSource{};
+        }
+
+        GLenum AttachmentPointFor(GLenum internalFormat) {
+            switch (internalFormat) {
+            case GL_DEPTH24_STENCIL8:
+            case GL_DEPTH32F_STENCIL8: return GL_DEPTH_STENCIL_ATTACHMENT;
+            case GL_STENCIL_INDEX8: return GL_STENCIL_ATTACHMENT;
+            default: return GL_DEPTH_ATTACHMENT;
+            }
+        }
+
+        bool FormatHasDepth(GLenum internalFormat) { return internalFormat != GL_STENCIL_INDEX8; }
+        bool FormatHasStencil(GLenum internalFormat) {
+            return internalFormat == GL_DEPTH24_STENCIL8 || internalFormat == GL_DEPTH32F_STENCIL8 ||
+                   internalFormat == GL_STENCIL_INDEX8;
+        }
+
+        // A framebuffer whose depth/stencil lives in a TEXTURE. The colour attachment is
+        // there so a stencil-only or depth-only framebuffer still has something to size it.
+        DepthSource MakeTextureSource(GLenum internalFormat) {
+            DepthSource source;
+            glGenFramebuffers(1, &source.fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, source.fbo);
+            glGenTextures(1, &source.colorTexture);
+            glBindTexture(GL_TEXTURE_2D, source.colorTexture);
+            glTexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, kWidth, kHeight);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, source.colorTexture, 0);
+            glGenTextures(1, &source.depthTexture);
+            glBindTexture(GL_TEXTURE_2D, source.depthTexture);
+            glTexStorage2D(GL_TEXTURE_2D, 1, internalFormat, kWidth, kHeight);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, AttachmentPointFor(internalFormat), GL_TEXTURE_2D,
+                                   source.depthTexture, 0);
+            return source;
+        }
+
+        // The same, with the depth/stencil in a RENDERBUFFER - which cannot be sampled at
+        // all, so the readback has no choice but to copy it somewhere samplable first.
+        // `samples` > 0 makes it multisample, which additionally needs a resolve.
+        DepthSource MakeRenderbufferSource(GLenum internalFormat, int samples) {
+            DepthSource source;
+            glGenFramebuffers(1, &source.fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, source.fbo);
+            glGenRenderbuffers(1, &source.colorRenderbuffer);
+            glBindRenderbuffer(GL_RENDERBUFFER, source.colorRenderbuffer);
+            if (samples > 0) {
+                glRenderbufferStorageMultisample(GL_RENDERBUFFER, samples, GL_RGBA8, kWidth, kHeight);
+            } else {
+                glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, kWidth, kHeight);
+            }
+            glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, source.colorRenderbuffer);
+            glGenRenderbuffers(1, &source.depthRenderbuffer);
+            glBindRenderbuffer(GL_RENDERBUFFER, source.depthRenderbuffer);
+            if (samples > 0) {
+                glRenderbufferStorageMultisample(GL_RENDERBUFFER, samples, internalFormat, kWidth, kHeight);
+            } else {
+                glRenderbufferStorage(GL_RENDERBUFFER, internalFormat, kWidth, kHeight);
+            }
+            glFramebufferRenderbuffer(GL_FRAMEBUFFER, AttachmentPointFor(internalFormat), GL_RENDERBUFFER,
+                                      source.depthRenderbuffer);
+            glBindRenderbuffer(GL_RENDERBUFFER, 0);
+            return source;
+        }
+
+        // Clears the bound framebuffer's depth and stencil to known values, with the masks
+        // and the scissor explicitly out of the way (a leaked scissor from an earlier
+        // scenario would clip the clear and every assertion after it).
+        void ClearDepthStencil(GLenum internalFormat, float depth, int stencil) {
+            glDisable(GL_SCISSOR_TEST);
+            glViewport(0, 0, kWidth, kHeight);
+            GLbitfield mask = 0;
+            if (FormatHasDepth(internalFormat)) {
+                glDepthMask(GL_TRUE);
+                glClearDepth(depth);
+                mask |= GL_DEPTH_BUFFER_BIT;
+            }
+            if (FormatHasStencil(internalFormat)) {
+                glStencilMask(0xFFu);
+                glClearStencil(stencil);
+                mask |= GL_STENCIL_BUFFER_BIT;
+            }
+            glClear(mask);
+        }
+
+        class DepthStencilReadbackMatrixScenario : public ScenarioTest {
+        protected:
+            // Not every ES driver can render to every depth format (DEPTH_COMPONENT32F and
+            // the multisample counts in particular), and an incomplete framebuffer would
+            // turn a legitimate "this machine cannot host the source" into a spurious
+            // failure about the readback.
+            static bool SourceIsUsable() {
+                return glCheckFramebufferStatus(GL_FRAMEBUFFER) == GLenum(GL_FRAMEBUFFER_COMPLETE);
+            }
+
+            static std::vector<float> ReadDepthFloat(int x, int y, int width, int height) {
+                std::vector<float> depth(static_cast<size_t>(width) * height, kDepthPoison);
+                glReadPixels(x, y, width, height, GL_DEPTH_COMPONENT, GL_FLOAT, depth.data());
+                return depth;
+            }
+
+            static std::vector<int> ReadStencilInt(int x, int y, int width, int height) {
+                std::vector<int> stencil(static_cast<size_t>(width) * height, kStencilPoison);
+                glReadPixels(x, y, width, height, GL_STENCIL_INDEX, GL_INT, stencil.data());
+                return stencil;
+            }
+
+            // "every value in the region is `expected`" rather than "the middle pixel is":
+            // a staging blit that lands the wrong rectangle, or a conversion pass with a
+            // half-texel offset, still gets the centre right.
+            static void ExpectAllDepth(const std::vector<float>& values, float expected, const char* what) {
+                size_t bad = 0;
+                float worst = expected;
+                for (float value : values) {
+                    if (std::fabs(value - expected) > 1.0f / 4096.0f) {
+                        if (bad == 0) worst = value;
+                        ++bad;
+                    }
+                }
+                EXPECT_EQ(bad, 0u) << what << ": " << bad << " of " << values.size()
+                                   << " depth values differ from " << expected << "; first bad value " << worst
+                                   << (std::fabs(worst - kDepthPoison) < 1e-6f
+                                           ? " - which is the poison value, so nothing was written at all"
+                                           : "");
+            }
+
+            static void ExpectAllStencil(const std::vector<int>& values, int expected, const char* what) {
+                size_t bad = 0;
+                int worst = expected;
+                for (int value : values) {
+                    if (value != expected) {
+                        if (bad == 0) worst = value;
+                        ++bad;
+                    }
+                }
+                EXPECT_EQ(bad, 0u) << what << ": " << bad << " of " << values.size()
+                                   << " stencil values differ from " << expected << "; first bad value " << worst
+                                   << (worst == kStencilPoison
+                                           ? " - which is the poison value, so nothing was written at all"
+                                           : "");
+            }
+        };
+
+        // ---- glReadPixels across the source kinds -----------------------------------
+
+        struct SourceCase {
+            const char* name;
+            GLenum internalFormat;
+            int samples;
+            bool renderbuffer;
+        };
+
+        const SourceCase kSourceCases[] = {
+            {"texture depth24_stencil8", GL_DEPTH24_STENCIL8, 0, false},
+            {"texture depth32f_stencil8", GL_DEPTH32F_STENCIL8, 0, false},
+            {"texture depth_component16", GL_DEPTH_COMPONENT16, 0, false},
+            {"texture depth_component24", GL_DEPTH_COMPONENT24, 0, false},
+            {"texture depth_component32f", GL_DEPTH_COMPONENT32F, 0, false},
+            {"renderbuffer depth24_stencil8", GL_DEPTH24_STENCIL8, 0, true},
+            {"renderbuffer depth_component24", GL_DEPTH_COMPONENT24, 0, true},
+            {"renderbuffer stencil_index8", GL_STENCIL_INDEX8, 0, true},
+        };
+
+    } // namespace
+
+    TEST_F(DepthStencilReadbackMatrixScenario, EverySourceKindReadsItsClearBack) {
+        if (!Ready()) return;
+        int exercised = 0;
+        for (const SourceCase& testCase : kSourceCases) {
+            SCOPED_TRACE(testCase.name);
+            DepthSource source = testCase.renderbuffer
+                                     ? MakeRenderbufferSource(testCase.internalFormat, testCase.samples)
+                                     : MakeTextureSource(testCase.internalFormat);
+            if (!SourceIsUsable()) {
+                DestroySource(source);
+                continue;
+            }
+            FirstGLError(); // the storage calls above may have probed an unsupported combination
+            ClearDepthStencil(testCase.internalFormat, 0.625f, 9);
+            EXPECT_EQ(FirstGLError(), 0u) << "clearing the source";
+
+            if (FormatHasDepth(testCase.internalFormat)) {
+                const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+                EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_DEPTH_COMPONENT, GL_FLOAT)";
+                ExpectAllDepth(depth, 0.625f, testCase.name);
+            }
+            if (FormatHasStencil(testCase.internalFormat)) {
+                const std::vector<int> stencil = ReadStencilInt(0, 0, kWidth, kHeight);
+                EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_STENCIL_INDEX, GL_INT)";
+                ExpectAllStencil(stencil, 9, testCase.name);
+            }
+            ++exercised;
+            DestroySource(source);
+        }
+        // A machine that hosted none of the sources would report a vacuous pass.
+        EXPECT_GE(exercised, 4) << "too few depth/stencil source kinds were usable to call this a matrix";
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // Depth and stencil in two SEPARATE objects, with two different formats, on the same
+    // framebuffer. Legal GL, and the shape KHR-GL3x.framebuffer_blit builds when its depth
+    // config and its stencil config are configured independently - so a readback that
+    // describes "the" depth/stencil source as one thing serves whichever aspect it happened
+    // to find first and silently abandons the other. Each aspect has to be staged from its
+    // own attachment, in its own format.
+    TEST_F(DepthStencilReadbackMatrixScenario, SeparateDepthAndStencilAttachmentsAreBothReadable) {
+        if (!Ready()) return;
+        DepthSource source;
+        glGenFramebuffers(1, &source.fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, source.fbo);
+        glGenRenderbuffers(1, &source.colorRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, source.colorRenderbuffer);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, kWidth, kHeight);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, source.colorRenderbuffer);
+        // Depth in a DEPTH_COMPONENT24 renderbuffer...
+        glGenRenderbuffers(1, &source.depthRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, source.depthRenderbuffer);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, kWidth, kHeight);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, source.depthRenderbuffer);
+        // ...and stencil in a STENCIL_INDEX8 one of its own.
+        GLuint stencilRenderbuffer = 0;
+        glGenRenderbuffers(1, &stencilRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, stencilRenderbuffer);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_STENCIL_INDEX8, kWidth, kHeight);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_RENDERBUFFER, stencilRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, 0);
+        if (!SourceIsUsable()) {
+            // Separate depth and stencil images are legal GL but many stacks answer
+            // GL_FRAMEBUFFER_UNSUPPORTED for them; say which, so a skip here is a fact about
+            // the driver rather than an unexplained hole in the matrix.
+            const GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+            glDeleteRenderbuffers(1, &stencilRenderbuffer);
+            DestroySource(source);
+            GTEST_SKIP() << "this driver cannot host separate DEPTH_COMPONENT24 and STENCIL_INDEX8 attachments: "
+                         << "glCheckFramebufferStatus = 0x" << std::hex << status;
+        }
+        FirstGLError();
+
+        glDisable(GL_SCISSOR_TEST);
+        glViewport(0, 0, kWidth, kHeight);
+        glDepthMask(GL_TRUE);
+        glStencilMask(0xFFu);
+        glClearDepth(0.3125);
+        glClearStencil(17);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u) << "reading depth from a separately-attached DEPTH_COMPONENT24";
+        ExpectAllDepth(depth, 0.3125f, "separate depth attachment");
+
+        const std::vector<int> stencil = ReadStencilInt(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u) << "reading stencil from a separately-attached STENCIL_INDEX8";
+        ExpectAllStencil(stencil, 17, "separate stencil attachment");
+
+        glDeleteRenderbuffers(1, &stencilRenderbuffer);
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // A multisample source is never read directly - glReadPixels on a multisampled
+    // framebuffer is INVALID_OPERATION in GL as much as in ES, and the state layer says so.
+    // The way multisample depth reaches a reader is a resolve blit into a single-sampled
+    // framebuffer, which is then read; that pair is
+    // KHR-GL3x.framebuffer_blit.multisampled_to_singlesampled_blit_depth_config_test, and
+    // the assertion here is that the resolved depth arrives intact rather than as the
+    // destination's own clear value.
+    TEST_F(DepthStencilReadbackMatrixScenario, AResolvedMultisampleDepthReadsBackFromTheDestination) {
+        if (!Ready()) return;
+        DepthSource multisampled = MakeRenderbufferSource(GL_DEPTH24_STENCIL8, 4);
+        if (!SourceIsUsable()) {
+            DestroySource(multisampled);
+            GTEST_SKIP() << "this driver cannot host a 4x multisample DEPTH24_STENCIL8 renderbuffer";
+        }
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.875f, 0);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // The destination starts at a depth the resolve must overwrite everywhere.
+        DepthSource resolved = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.125f, 0);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, multisampled.fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, resolved.fbo);
+        glDisable(GL_SCISSOR_TEST);
+        glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
+        EXPECT_EQ(FirstGLError(), 0u) << "resolving a multisample depth buffer into a single-sampled one";
+
+        glBindFramebuffer(GL_FRAMEBUFFER, resolved.fbo);
+        const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(depth, 0.875f, "resolved multisample depth");
+
+        DestroySource(resolved);
+        DestroySource(multisampled);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // A read whose rectangle is NOT the whole attachment. The staging copy has to carry
+    // the requested rect (not the origin) and hand back its rows bottom-up, which a
+    // full-extent uniform read is a fixed point of and therefore cannot see.
+    TEST_F(DepthStencilReadbackMatrixScenario, ASubRectangleReadsTheRightBandInTheRightOrder) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+
+        // Bottom half 0.25, top half 0.75, and the stencil banded the other way round so a
+        // mix-up between the two aspects cannot pass either.
+        glDisable(GL_SCISSOR_TEST);
+        glViewport(0, 0, kWidth, kHeight);
+        glDepthMask(GL_TRUE);
+        glStencilMask(0xFFu);
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(0, 0, kWidth, kHeight / 2);
+        glClearDepth(0.25);
+        glClearStencil(11);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+        glScissor(0, kHeight / 2, kWidth, kHeight - kHeight / 2);
+        glClearDepth(0.75);
+        glClearStencil(22);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+        glDisable(GL_SCISSOR_TEST);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // A rect wholly inside the bottom band, offset from the origin in both axes.
+        const int rectWidth = 8;
+        const int rectHeight = 4;
+        const std::vector<float> bottom = ReadDepthFloat(16, 4, rectWidth, rectHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(bottom, 0.25f, "sub-rect inside the bottom depth band");
+        const std::vector<int> bottomStencil = ReadStencilInt(16, 4, rectWidth, rectHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllStencil(bottomStencil, 11, "sub-rect inside the bottom stencil band");
+
+        // And one wholly inside the top band. Reading the mirrored row would answer 0.25.
+        const std::vector<float> top = ReadDepthFloat(16, kHeight - 4 - rectHeight, rectWidth, rectHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(top, 0.75f, "sub-rect inside the top depth band");
+
+        // A rect that STRADDLES the boundary pins the row order itself: its first rows must
+        // be the bottom band and its last rows the top one.
+        const int straddleHeight = 8;
+        const std::vector<float> straddle =
+            ReadDepthFloat(16, kHeight / 2 - straddleHeight / 2, rectWidth, straddleHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_EQ(straddle.size(), static_cast<size_t>(rectWidth) * straddleHeight);
+        EXPECT_NEAR(straddle[0], 0.25f, 1.0f / 4096.0f)
+            << "the first row of the returned rect must be its BOTTOM row (GL order), which is in the 0.25 band";
+        EXPECT_NEAR(straddle[straddle.size() - 1], 0.75f, 1.0f / 4096.0f)
+            << "the last row of the returned rect must be its TOP row, which is in the 0.75 band";
+
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // The packed layouts the packed_depth_stencil family reads its gradients with.
+    TEST_F(DepthStencilReadbackMatrixScenario, PackedDepthStencilReadPixelsCarriesBothAspects) {
+        if (!Ready()) return;
+        struct PackedCase {
+            const char* name;
+            GLenum internalFormat;
+            GLenum type;
+        };
+        const PackedCase cases[] = {
+            {"depth24_stencil8 / GL_UNSIGNED_INT_24_8", GL_DEPTH24_STENCIL8, GL_UNSIGNED_INT_24_8},
+            {"depth32f_stencil8 / GL_FLOAT_32_UNSIGNED_INT_24_8_REV", GL_DEPTH32F_STENCIL8,
+             GL_FLOAT_32_UNSIGNED_INT_24_8_REV},
+        };
+        int exercised = 0;
+        for (const PackedCase& testCase : cases) {
+            SCOPED_TRACE(testCase.name);
+            DepthSource source = MakeTextureSource(testCase.internalFormat);
+            if (!SourceIsUsable()) {
+                DestroySource(source);
+                continue;
+            }
+            FirstGLError();
+            ClearDepthStencil(testCase.internalFormat, 0.5f, 3);
+            ASSERT_EQ(FirstGLError(), 0u);
+
+            const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+            if (testCase.type == GL_UNSIGNED_INT_24_8) {
+                std::vector<unsigned int> packed(pixels, kPacked24_8Poison);
+                glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_STENCIL, testCase.type, packed.data());
+                EXPECT_EQ(FirstGLError(), 0u);
+                size_t bad = 0;
+                for (unsigned int value : packed) {
+                    const float depth = static_cast<float>(value >> 8) / 16777215.0f;
+                    const int stencil = static_cast<int>(value & 0xFFu);
+                    if (std::fabs(depth - 0.5f) > 0.01f || stencil != 3) ++bad;
+                }
+                EXPECT_EQ(bad, 0u) << testCase.name << ": " << bad << " of " << pixels
+                                   << " packed words carry the wrong depth or stencil (first word 0x" << std::hex
+                                   << packed[0] << std::dec << ")";
+            } else {
+                std::vector<D32fS8> packed(pixels, D32fS8{kDepthPoison, static_cast<unsigned int>(kStencilPoison)});
+                glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_STENCIL, testCase.type, packed.data());
+                EXPECT_EQ(FirstGLError(), 0u);
+                size_t bad = 0;
+                for (const D32fS8& value : packed) {
+                    if (std::fabs(value.depth - 0.5f) > 0.01f || (value.stencil & 0xFFu) != 3u) ++bad;
+                }
+                EXPECT_EQ(bad, 0u) << testCase.name << ": " << bad << " of " << pixels
+                                   << " packed pairs carry the wrong depth or stencil (first pair depth "
+                                   << packed[0].depth << " stencil " << (packed[0].stencil & 0xFFu) << ")";
+            }
+            ++exercised;
+            DestroySource(source);
+        }
+        EXPECT_GE(exercised, 1) << "neither packed depth/stencil format was renderable";
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // glGetTexImage reads a TEXTURE, not the bound framebuffer - a different entry point
+    // that has to reach the same machinery. This is verify_get_tex_image's shape.
+    TEST_F(DepthStencilReadbackMatrixScenario, GetTexImageReadsAPackedDepthStencilTexture) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.375f, 5);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // Read it back through the texture, with the framebuffer that owns it unbound so a
+        // path that secretly read the framebuffer instead would answer from somewhere else.
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        glBindTexture(GL_TEXTURE_2D, source.depthTexture);
+        const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+        std::vector<unsigned int> packed(pixels, kPacked24_8Poison);
+        glGetTexImage(GL_TEXTURE_2D, 0, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, packed.data());
+        EXPECT_EQ(FirstGLError(), 0u);
+        size_t bad = 0;
+        for (unsigned int value : packed) {
+            const float depth = static_cast<float>(value >> 8) / 16777215.0f;
+            if (std::fabs(depth - 0.375f) > 0.01f || (value & 0xFFu) != 5u) ++bad;
+        }
+        EXPECT_EQ(bad, 0u) << bad << " of " << pixels
+                           << " words from glGetTexImage(GL_DEPTH_STENCIL) are wrong (first word 0x" << std::hex
+                           << packed[0] << std::dec << ")";
+
+        glBindTexture(GL_TEXTURE_2D, 0);
+        DestroySource(source);
+        Gl().EndFrame();
+    }
+
+    // glCopyTexImage2D out of a depth attachment, then read the copy - verify_copy_tex_image.
+    TEST_F(DepthStencilReadbackMatrixScenario, CopyTexImageFromADepthAttachmentSurvivesAReadBack) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.75f, 6);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        GLuint copy = 0;
+        glGenTextures(1, &copy);
+        glBindTexture(GL_TEXTURE_2D, copy);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, kWidth, kHeight, 0, GL_DEPTH_STENCIL,
+                     GL_UNSIGNED_INT_24_8, nullptr);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        glCopyTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, 0, 0, kWidth, kHeight, 0);
+        EXPECT_EQ(FirstGLError(), 0u) << "glCopyTexImage2D from a depth/stencil attachment";
+
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+        std::vector<unsigned int> packed(pixels, kPacked24_8Poison);
+        glGetTexImage(GL_TEXTURE_2D, 0, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, packed.data());
+        EXPECT_EQ(FirstGLError(), 0u);
+        size_t bad = 0;
+        for (unsigned int value : packed) {
+            const float depth = static_cast<float>(value >> 8) / 16777215.0f;
+            if (std::fabs(depth - 0.75f) > 0.01f) ++bad;
+        }
+        EXPECT_EQ(bad, 0u) << bad << " of " << pixels << " copied depth values are wrong (first word 0x" << std::hex
+                           << packed[0] << std::dec << ")";
+
+        glBindTexture(GL_TEXTURE_2D, 0);
+        glDeleteTextures(1, &copy);
+        DestroySource(source);
+        Gl().EndFrame();
+    }
+
+    // The integer client widths, which are a separate conversion each.
+    TEST_F(DepthStencilReadbackMatrixScenario, DepthAndStencilConvertIntoEveryClientWidth) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.5f, 200);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+
+        std::vector<unsigned int> depthUint(pixels, 0xDEADBEEFu);
+        glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_COMPONENT, GL_UNSIGNED_INT, depthUint.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_DEPTH_COMPONENT, GL_UNSIGNED_INT)";
+        // 0.5 of the full 32-bit range, with room for the source's 24-bit quantisation.
+        EXPECT_NEAR(static_cast<double>(depthUint[0]) / 4294967295.0, 0.5, 0.01)
+            << "GL_UNSIGNED_INT depth came back as " << depthUint[0];
+
+        std::vector<unsigned short> depthUshort(pixels, 0xBEEFu);
+        glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT, depthUshort.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT)";
+        EXPECT_NEAR(static_cast<double>(depthUshort[0]) / 65535.0, 0.5, 0.01)
+            << "GL_UNSIGNED_SHORT depth came back as " << depthUshort[0];
+
+        // A stencil index is written unconverted into whichever width was asked for, so 200
+        // must survive intact in all of them - it is also large enough that a signed byte
+        // would wrap, which is the point of choosing it.
+        std::vector<unsigned char> stencilByte(pixels, static_cast<unsigned char>(kStencilPoison));
+        glReadPixels(0, 0, kWidth, kHeight, GL_STENCIL_INDEX, GL_UNSIGNED_BYTE, stencilByte.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_STENCIL_INDEX, GL_UNSIGNED_BYTE)";
+        EXPECT_EQ(static_cast<int>(stencilByte[0]), 200);
+
+        std::vector<int> stencilInt(pixels, kStencilPoison);
+        glReadPixels(0, 0, kWidth, kHeight, GL_STENCIL_INDEX, GL_INT, stencilInt.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_STENCIL_INDEX, GL_INT)";
+        EXPECT_EQ(stencilInt[0], 200);
+
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // The PACK pixel-store parameters apply to a depth read exactly as they do to a colour
+    // one, and the gap regions they create must be left alone.
+    TEST_F(DepthStencilReadbackMatrixScenario, DepthReadbackHonoursThePackPixelStoreParameters) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH_COMPONENT24);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH_COMPONENT24, 0.5f, 0);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const int rectWidth = 4;
+        const int rectHeight = 3;
+        const int rowLength = 8;
+        const int skipPixels = 2;
+        const int skipRows = 1;
+        constexpr float kGap = -7.0f;
+        std::vector<float> destination(static_cast<size_t>(rowLength) * (skipRows + rectHeight) + 16, kGap);
+
+        glPixelStorei(GL_PACK_ROW_LENGTH, rowLength);
+        glPixelStorei(GL_PACK_SKIP_PIXELS, skipPixels);
+        glPixelStorei(GL_PACK_SKIP_ROWS, skipRows);
+        glPixelStorei(GL_PACK_ALIGNMENT, 4);
+        glReadPixels(0, 0, rectWidth, rectHeight, GL_DEPTH_COMPONENT, GL_FLOAT, destination.data());
+        const unsigned int readError = FirstGLError();
+        glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+        glPixelStorei(GL_PACK_SKIP_PIXELS, 0);
+        glPixelStorei(GL_PACK_SKIP_ROWS, 0);
+        glPixelStorei(GL_PACK_ALIGNMENT, 4);
+        EXPECT_EQ(readError, 0u);
+
+        size_t written = 0;
+        size_t gapsTouched = 0;
+        for (size_t index = 0; index < destination.size(); ++index) {
+            const long row = static_cast<long>(index) / rowLength - skipRows;
+            const long column = static_cast<long>(index) % rowLength - skipPixels;
+            const bool inRect = row >= 0 && row < rectHeight && column >= 0 && column < rectWidth;
+            if (inRect) {
+                if (std::fabs(destination[index] - 0.5f) <= 1.0f / 4096.0f) ++written;
+            } else if (destination[index] != kGap) {
+                ++gapsTouched;
+            }
+        }
+        EXPECT_EQ(written, static_cast<size_t>(rectWidth) * rectHeight)
+            << "only " << written << " of " << (rectWidth * rectHeight)
+            << " destination pixels landed where GL_PACK_ROW_LENGTH/SKIP_* put them";
+        EXPECT_EQ(gapsTouched, 0u) << gapsTouched << " bytes outside the packed rectangle were overwritten";
+
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // The readback borrows the application's context for a full-screen pass. Everything it
+    // touches has to come back, or the next draw inherits it - which is how an emulation
+    // that "works" takes the rest of the renderer down with it.
+    TEST_F(DepthStencilReadbackMatrixScenario, ReadbackLeavesNoGLStateBehind) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.5f, 4);
+
+        // A deliberately awkward state: nothing here is what an emulation pass would want,
+        // so anything it forgets to put back shows up below.
+        GLuint scratchTexture = 0;
+        glGenTextures(1, &scratchTexture);
+        glBindTexture(GL_TEXTURE_2D, scratchTexture);
+        glActiveTexture(GL_TEXTURE3);
+        glBindTexture(GL_TEXTURE_2D, scratchTexture);
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(3, 5, 7, 11);
+        glEnable(GL_CULL_FACE);
+        glEnable(GL_BLEND);
+        glEnable(GL_DEPTH_TEST);
+        glDepthFunc(GL_GEQUAL);
+        glDepthMask(GL_FALSE);
+        glEnable(GL_STENCIL_TEST);
+        glStencilFunc(GL_NOTEQUAL, 0x5, 0x0Fu);
+        glStencilOp(GL_INCR, GL_DECR, GL_INVERT);
+        glStencilMask(0x3Cu);
+        glColorMask(GL_FALSE, GL_TRUE, GL_FALSE, GL_TRUE);
+        glViewport(2, 3, 5, 7);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+        const std::vector<int> stencil = ReadStencilInt(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(depth, 0.5f, "state-preservation case depth");
+        ExpectAllStencil(stencil, 4, "state-preservation case stencil");
+
+        GLint viewport[4] = {0, 0, 0, 0};
+        GLint scissorBox[4] = {0, 0, 0, 0};
+        GLboolean colorMask[4] = {GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE};
+        GLint depthFunc = 0;
+        GLboolean depthMask = GL_TRUE;
+        GLint stencilFunc = 0, stencilRef = 0, stencilValueMask = 0, stencilWriteMask = 0;
+        GLint stencilFail = 0, stencilPassDepthFail = 0, stencilPassDepthPass = 0;
+        GLint activeTexture = 0, boundTexture = 0;
+        glGetIntegerv(GL_VIEWPORT, viewport);
+        glGetIntegerv(GL_SCISSOR_BOX, scissorBox);
+        glGetBooleanv(GL_COLOR_WRITEMASK, colorMask);
+        glGetIntegerv(GL_DEPTH_FUNC, &depthFunc);
+        glGetBooleanv(GL_DEPTH_WRITEMASK, &depthMask);
+        glGetIntegerv(GL_STENCIL_FUNC, &stencilFunc);
+        glGetIntegerv(GL_STENCIL_REF, &stencilRef);
+        glGetIntegerv(GL_STENCIL_VALUE_MASK, &stencilValueMask);
+        glGetIntegerv(GL_STENCIL_WRITEMASK, &stencilWriteMask);
+        glGetIntegerv(GL_STENCIL_FAIL, &stencilFail);
+        glGetIntegerv(GL_STENCIL_PASS_DEPTH_FAIL, &stencilPassDepthFail);
+        glGetIntegerv(GL_STENCIL_PASS_DEPTH_PASS, &stencilPassDepthPass);
+        glGetIntegerv(GL_ACTIVE_TEXTURE, &activeTexture);
+        glGetIntegerv(GL_TEXTURE_BINDING_2D, &boundTexture);
+
+        EXPECT_EQ(viewport[0], 2);
+        EXPECT_EQ(viewport[1], 3);
+        EXPECT_EQ(viewport[2], 5);
+        EXPECT_EQ(viewport[3], 7);
+        EXPECT_EQ(scissorBox[0], 3);
+        EXPECT_EQ(scissorBox[1], 5);
+        EXPECT_EQ(scissorBox[2], 7);
+        EXPECT_EQ(scissorBox[3], 11);
+        EXPECT_EQ(glIsEnabled(GL_SCISSOR_TEST), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_CULL_FACE), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_BLEND), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_DEPTH_TEST), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_STENCIL_TEST), GLboolean(GL_TRUE));
+        EXPECT_EQ(colorMask[0], GLboolean(GL_FALSE));
+        EXPECT_EQ(colorMask[1], GLboolean(GL_TRUE));
+        EXPECT_EQ(colorMask[2], GLboolean(GL_FALSE));
+        EXPECT_EQ(colorMask[3], GLboolean(GL_TRUE));
+        EXPECT_EQ(depthFunc, GLint(GL_GEQUAL));
+        EXPECT_EQ(depthMask, GLboolean(GL_FALSE));
+        EXPECT_EQ(stencilFunc, GLint(GL_NOTEQUAL));
+        EXPECT_EQ(stencilRef, 0x5);
+        EXPECT_EQ(stencilValueMask, 0x0F);
+        EXPECT_EQ(stencilWriteMask, 0x3C);
+        EXPECT_EQ(stencilFail, GLint(GL_INCR));
+        EXPECT_EQ(stencilPassDepthFail, GLint(GL_DECR));
+        EXPECT_EQ(stencilPassDepthPass, GLint(GL_INVERT));
+        EXPECT_EQ(activeTexture, GLint(GL_TEXTURE3));
+        EXPECT_EQ(boundTexture, GLint(scratchTexture))
+            << "the readback left a scratch texture on the application's texture unit";
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        // Put the awkward state back so the next scenario in this process starts clean.
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_CULL_FACE);
+        glDisable(GL_BLEND);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_STENCIL_TEST);
+        glDepthFunc(GL_LESS);
+        glDepthMask(GL_TRUE);
+        glStencilFunc(GL_ALWAYS, 0, 0xFFFFFFFFu);
+        glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);
+        glStencilMask(0xFFFFFFFFu);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+        glBindTexture(GL_TEXTURE_2D, 0);
+        glActiveTexture(GL_TEXTURE0);
+        glDeleteTextures(1, &scratchTexture);
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        glViewport(0, 0, Gl().Width(), Gl().Height());
+        Gl().EndFrame();
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackScenario.cpp

@@ -58,12 +58,12 @@ namespace MGITest {
 
         class DepthStencilReadbackScenario : public ScenarioTest {
         protected:
-            // DirectGLES reads depth and stencil back through the ES driver, which has no
+            // Both backends now answer these reads. DirectGLES has no native ES path for
-            // guaranteed path for either (GL_NV_read_depth / GL_NV_read_stencil are optional and
+            // either aspect (GL_NV_read_depth / GL_NV_read_stencil are optional and absent on
-            // absent on both the Adreno device and Mesa's ES). That gap is tracked separately as
+            // both the Adreno device and Mesa's ES), so it stages the attachment into a
-            // the packed_depth_stencil cluster and needs a shader-sampling emulation, not this
+            // scratch depth texture and samples it into a colour target; the assertions below
-            // change; asserting it here would only pin a known-missing feature.
+            // are the same either way, which is the point.
-            bool BackendReadsDepthStencil() const { return Gl().BackendName() == "DirectVulkan"; }
+            bool BackendReadsDepthStencil() const { return true; }
 
             float ReadDepthAt(int x, int y) const {
                 float depth = kDepthPoison;

MobileGL/MG_IntegrationTest/Scenarios/DoublePrecisionScenario.cpp

@@ -99,6 +99,8 @@ void main() {
 
             void TearDown() override {
                 if (!Ready()) return;
+                if (m_shapeOutput != 0) glDeleteBuffers(1, &m_shapeOutput);
+                if (m_shapeProgram != 0) glDeleteProgram(m_shapeProgram);
                 if (m_output != 0) glDeleteBuffers(1, &m_output);
                 if (m_program != 0) glDeleteProgram(m_program);
             }
@@ -146,9 +148,147 @@ void main() {
 
             unsigned int m_program = 0;
             unsigned int m_output = 0;
+            unsigned int m_shapeProgram = 0;
+            unsigned int m_shapeOutput = 0;
             std::string m_buildLog;
         };
 
+        // Every double-typed uniform shape GLSL has, all thirteen of them, in one program - the
+        // shape of KHR-GL43.compute_shader.fp64-case2. The scalar and the square matrices are
+        // covered by the cases above; what only a set like this reaches is the NON-SQUARE
+        // matrices, whose column stride and total size both change when the demotion turns a
+        // 64-bit column into a 32-bit one, and whose members therefore move every uniform
+        // declared after them.
+        //
+        // The shader reports every component separately rather than one pass/fail flag, because
+        // "the readback is wrong" is not a diagnosis: a wrong column stride, a wrong member
+        // offset and a wrong narrowing all fail the same single comparison, and only the
+        // component map says which.
+        // No #version here on purpose: it is handed over as a separate source string, the way
+        // the CTS case hands it over.
+        constexpr const char* kAllDoubleShapesSource = R"(
+layout(local_size_x = 1) in;
+uniform double  g_0;
+uniform dvec2   g_1;
+uniform dvec3   g_2;
+uniform dvec4   g_3;
+uniform dmat2   g_4;
+uniform dmat2x3 g_5;
+uniform dmat2x4 g_6;
+uniform dmat3x2 g_7;
+uniform dmat3   g_8;
+uniform dmat3x4 g_9;
+uniform dmat4x2 g_10;
+uniform dmat4x3 g_11;
+uniform dmat4   g_12;
+layout(std430, binding = 0) buffer Output {
+    float g_out[];
+};
+void main() {
+    g_out[0] = float(g_0);
+    for (int i = 0; i < 2; ++i) g_out[1 + i] = float(g_1[i]);
+    for (int i = 0; i < 3; ++i) g_out[3 + i] = float(g_2[i]);
+    for (int i = 0; i < 4; ++i) g_out[6 + i] = float(g_3[i]);
+    for (int c = 0; c < 2; ++c) for (int r = 0; r < 2; ++r) g_out[10 + c * 2 + r] = float(g_4[c][r]);
+    for (int c = 0; c < 2; ++c) for (int r = 0; r < 3; ++r) g_out[14 + c * 3 + r] = float(g_5[c][r]);
+    for (int c = 0; c < 2; ++c) for (int r = 0; r < 4; ++r) g_out[20 + c * 4 + r] = float(g_6[c][r]);
+    for (int c = 0; c < 3; ++c) for (int r = 0; r < 2; ++r) g_out[28 + c * 2 + r] = float(g_7[c][r]);
+    for (int c = 0; c < 3; ++c) for (int r = 0; r < 3; ++r) g_out[34 + c * 3 + r] = float(g_8[c][r]);
+    for (int c = 0; c < 3; ++c) for (int r = 0; r < 4; ++r) g_out[43 + c * 4 + r] = float(g_9[c][r]);
+    for (int c = 0; c < 4; ++c) for (int r = 0; r < 2; ++r) g_out[55 + c * 2 + r] = float(g_10[c][r]);
+    for (int c = 0; c < 4; ++c) for (int r = 0; r < 3; ++r) g_out[63 + c * 3 + r] = float(g_11[c][r]);
+    for (int c = 0; c < 4; ++c) for (int r = 0; r < 4; ++r) g_out[75 + c * 4 + r] = float(g_12[c][r]);
+}
+)";
+
+        // The values the CTS case sets, spelled the way it spells them - column-major, and small
+        // enough that every one is exact in a float. Nothing here is a precision question; a
+        // component that comes back wrong came back from the wrong bytes.
+        constexpr double kG0 = 1.0;
+        constexpr double kG1[2] = {2.0, 3.0};
+        constexpr double kG2[3] = {4.0, 5.0, 6.0};
+        constexpr double kG3[4] = {7.0, 8.0, 9.0, 10.0};
+        constexpr double kG4[4] = {11.0, 12.0, 13.0, 14.0};
+        constexpr double kG5[6] = {15.0, 16.0, 17.0, 18.0, 19.0, 20.0};
+        constexpr double kG6[8] = {21.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0, 28.0};
+        constexpr double kG7[6] = {29.0, 30.0, 31.0, 32.0, 33.0, 34.0};
+        constexpr double kG8[9] = {35.0, 36.0, 37.0, 38.0, 39.0, 40.0, 41.0, 42.0, 43.0};
+        constexpr double kG9[12] = {44.0, 45.0, 46.0, 47.0, 48.0, 49.0, 50.0, 51.0, 52.0, 53.0, 54.0, 55.0};
+        constexpr double kG10[8] = {56.0, 57.0, 58.0, 59.0, 60.0, 61.0, 62.0, 63.0};
+        constexpr double kG11[12] = {63.0, 64.0, 65.0, 66.0, 67.0, 68.0, 69.0, 70.0, 71.0, 27.0, 73.0, 74.0};
+        constexpr double kG12[16] = {75.0, 76.0, 77.0, 78.0, 79.0, 80.0, 81.0, 82.0,
+                                     83.0, 84.0, 85.0, 86.0, 87.0, 88.0, 89.0, 90.0};
+
+        struct DoubleShape {
+            const char* name;
+            int base;
+            int columns; // 1 for the scalar and the vectors
+            int rows;    // component count for the scalar and the vectors
+            const double* values;
+        };
+
+        constexpr DoubleShape kDoubleShapes[] = {
+            {"g_0 double", 0, 1, 1, &kG0},      {"g_1 dvec2", 1, 1, 2, kG1},
+            {"g_2 dvec3", 3, 1, 3, kG2},        {"g_3 dvec4", 6, 1, 4, kG3},
+            {"g_4 dmat2", 10, 2, 2, kG4},       {"g_5 dmat2x3", 14, 2, 3, kG5},
+            {"g_6 dmat2x4", 20, 2, 4, kG6},     {"g_7 dmat3x2", 28, 3, 2, kG7},
+            {"g_8 dmat3", 34, 3, 3, kG8},       {"g_9 dmat3x4", 43, 3, 4, kG9},
+            {"g_10 dmat4x2", 55, 4, 2, kG10},   {"g_11 dmat4x3", 63, 4, 3, kG11},
+            {"g_12 dmat4", 75, 4, 4, kG12},
+        };
+
+        constexpr int kAllShapeSlots = 91;
+
+        // The conformance case's own shader, kept verbatim down to the literal suffixes and the
+        // unnamed, unqualified storage block - except that each comparison sets its OWN bit
+        // instead of collapsing all thirteen into one flag. That single flag is the whole reason
+        // the case was unexplained for a wave: it says "something is wrong" and nothing else.
+        //
+        // Verbatim matters here. Reading the components out one at a time (the case above)
+        // passes; whatever fails does so through the shape the conformance case actually
+        // writes - whole-matrix comparison against a constructor, a storage block with no
+        // layout qualifier and no instance name, values reached with constant indices.
+        constexpr const char* kCtsShapedSource = R"(
+layout(local_size_x = 1) in;
+buffer Result {
+  int g_result;
+};
+uniform double g_0;
+uniform dvec2 g_1;
+uniform dvec3 g_2;
+uniform dvec4 g_3;
+uniform dmat2 g_4;
+uniform dmat2x3 g_5;
+uniform dmat2x4 g_6;
+uniform dmat3x2 g_7;
+uniform dmat3 g_8;
+uniform dmat3x4 g_9;
+uniform dmat4x2 g_10;
+uniform dmat4x3 g_11;
+uniform dmat4 g_12;
+
+void main() {
+  g_result = 0;
+
+  if (g_0 != 1.0LF) g_result |= 1;
+  if (g_1 != dvec2(2.0LF, 3.0LF)) g_result |= 2;
+  if (g_2 != dvec3(4.0LF, 5.0LF, 6.0LF)) g_result |= 4;
+  if (g_3 != dvec4(7.0LF, 8.0LF, 9.0LF, 10.0LF)) g_result |= 8;
+
+  if (g_4 != dmat2(11.0LF, 12.0LF, 13.0LF, 14.0LF)) g_result |= 16;
+  if (g_5 != dmat2x3(15.0LF, 16.0LF, 17.0LF, 18.0LF, 19.0LF, 20.0LF)) g_result |= 32;
+  if (g_6 != dmat2x4(21.0LF, 22.0LF, 23.0LF, 24.0LF, 25.0LF, 26.0LF, 27.0LF, 28.0LF)) g_result |= 64;
+
+  if (g_7 != dmat3x2(29.0LF, 30.0LF, 31.0LF, 32.0LF, 33.0LF, 34.0LF)) g_result |= 128;
+  if (g_8 != dmat3(35.0LF, 36.0LF, 37.0LF, 38.0LF, 39.0LF, 40.0LF, 41.0LF, 42.0LF, 43.0LF)) g_result |= 256;
+  if (g_9 != dmat3x4(44.0LF, 45.0LF, 46.0LF, 47.0LF, 48.0LF, 49.0LF, 50.0LF, 51.0LF, 52.0LF, 53.0LF, 54.0LF, 55.0LF)) g_result |= 512;
+
+  if (g_10 != dmat4x2(56.0, 57.0, 58.0, 59.0, 60.0, 61.0, 62.0, 63.0)) g_result |= 1024;
+  if (g_11 != dmat4x3(63.0, 64.0, 65.0, 66.0, 67.0, 68.0, 69.0, 70.0, 71.0, 27.0, 73, 74.0)) g_result |= 2048;
+  if (g_12 != dmat4(75.0, 76.0, 77.0, 78.0, 79.0, 80.0, 81.0, 82.0, 83.0, 84.0, 85.0, 86.0, 87.0, 88.0, 89.0, 90.0)) g_result |= 4096;
+}
+)";
+
         TEST_F(DoublePrecisionScenario, ADoubleUniformReachesTheShaderAtFloatPrecision) {
             if (!Ready()) return;
             glUseProgram(m_program);
@@ -353,6 +493,190 @@ void main() {
             EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
         }
 
+        TEST_F(DoublePrecisionScenario, EveryDoubleUniformShapeArrivesWhereTheShaderReadsIt) {
+            if (!Ready()) return;
+            // Built the way the CTS case builds it, because every step of that build has been a
+            // bug here at least once: the source arrives as TWO strings (the version directive
+            // and the body), the shader is attached before it has a source and deleted while
+            // still attached, and the program is linked twice.
+            m_shapeProgram = glCreateProgram();
+            ASSERT_NE(m_shapeProgram, 0u);
+            {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glAttachShader(m_shapeProgram, shader);
+                glDeleteShader(shader);
+                const char* const sources[2] = {"#version 430 core\n", kAllDoubleShapesSource};
+                glShaderSource(shader, 2, sources, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    FAIL() << "compute shader did not compile: " << log;
+                }
+            }
+            glLinkProgram(m_shapeProgram);
+            {
+                GLint linkedOnce = 0;
+                glGetProgramiv(m_shapeProgram, GL_LINK_STATUS, &linkedOnce);
+                if (linkedOnce == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(m_shapeProgram, sizeof(log) - 1, nullptr, log);
+                    FAIL() << "compute program did not link: " << log;
+                }
+            }
+
+            glGenBuffers(1, &m_shapeOutput);
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_shapeOutput);
+            const std::vector<float> zeroes(kAllShapeSlots, 0.0f);
+            glBufferData(GL_SHADER_STORAGE_BUFFER, kAllShapeSlots * sizeof(float), zeroes.data(), GL_DYNAMIC_DRAW);
+            glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_shapeOutput);
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+            const auto location = [&](const char* name) { return glGetUniformLocation(m_shapeProgram, name); };
+
+            // Pass one sets through glProgramUniform*, pass two through glUniform* after a
+            // re-link - the two entry-point families the CTS case exercises, and two different
+            // routes into the same uniform storage.
+            const auto setWithProgramUniform = [&]() {
+                glProgramUniform1d(m_shapeProgram, location("g_0"), kG0);
+                glProgramUniform2d(m_shapeProgram, location("g_1"), kG1[0], kG1[1]);
+                glProgramUniform3d(m_shapeProgram, location("g_2"), kG2[0], kG2[1], kG2[2]);
+                glProgramUniform4d(m_shapeProgram, location("g_3"), kG3[0], kG3[1], kG3[2], kG3[3]);
+                glProgramUniformMatrix2dv(m_shapeProgram, location("g_4"), 1, GL_FALSE, kG4);
+                glProgramUniformMatrix2x3dv(m_shapeProgram, location("g_5"), 1, GL_FALSE, kG5);
+                glProgramUniformMatrix2x4dv(m_shapeProgram, location("g_6"), 1, GL_FALSE, kG6);
+                glProgramUniformMatrix3x2dv(m_shapeProgram, location("g_7"), 1, GL_FALSE, kG7);
+                glProgramUniformMatrix3dv(m_shapeProgram, location("g_8"), 1, GL_FALSE, kG8);
+                glProgramUniformMatrix3x4dv(m_shapeProgram, location("g_9"), 1, GL_FALSE, kG9);
+                glProgramUniformMatrix4x2dv(m_shapeProgram, location("g_10"), 1, GL_FALSE, kG10);
+                glProgramUniformMatrix4x3dv(m_shapeProgram, location("g_11"), 1, GL_FALSE, kG11);
+                glProgramUniformMatrix4dv(m_shapeProgram, location("g_12"), 1, GL_FALSE, kG12);
+            };
+            // Deliberately does NOT re-issue glUseProgram: the CTS case leaves the program
+            // current across the re-link and writes into it from there, so this is the path
+            // where a re-link has to keep the current program's uniform storage addressable.
+            const auto setWithUniform = [&]() {
+                glUniform1d(location("g_0"), kG0);
+                glUniform2d(location("g_1"), kG1[0], kG1[1]);
+                glUniform3d(location("g_2"), kG2[0], kG2[1], kG2[2]);
+                glUniform4d(location("g_3"), kG3[0], kG3[1], kG3[2], kG3[3]);
+                glUniformMatrix2dv(location("g_4"), 1, GL_FALSE, kG4);
+                glUniformMatrix2x3dv(location("g_5"), 1, GL_FALSE, kG5);
+                glUniformMatrix2x4dv(location("g_6"), 1, GL_FALSE, kG6);
+                glUniformMatrix3x2dv(location("g_7"), 1, GL_FALSE, kG7);
+                glUniformMatrix3dv(location("g_8"), 1, GL_FALSE, kG8);
+                glUniformMatrix3x4dv(location("g_9"), 1, GL_FALSE, kG9);
+                glUniformMatrix4x2dv(location("g_10"), 1, GL_FALSE, kG10);
+                glUniformMatrix4x3dv(location("g_11"), 1, GL_FALSE, kG11);
+                glUniformMatrix4dv(location("g_12"), 1, GL_FALSE, kG12);
+            };
+
+            const auto dispatchAndRead = [&]() {
+                glUseProgram(m_shapeProgram);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+                std::vector<float> values(kAllShapeSlots, -1.0f);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_shapeOutput);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, kAllShapeSlots * sizeof(float), values.data());
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                // The program stays current on purpose - see setWithUniform.
+                return values;
+            };
+
+            const auto expectEverything = [](const std::vector<float>& values, const char* pass) {
+                for (const DoubleShape& shape : kDoubleShapes) {
+                    for (int c = 0; c < shape.columns; ++c) {
+                        for (int r = 0; r < shape.rows; ++r) {
+                            const int component = c * shape.rows + r;
+                            EXPECT_FLOAT_EQ(values[shape.base + component],
+                                            static_cast<float>(shape.values[component]))
+                                << pass << ": " << shape.name << " column " << c << " row " << r;
+                        }
+                    }
+                }
+            };
+
+            setWithProgramUniform();
+            expectEverything(dispatchAndRead(), "glProgramUniform*");
+
+            // A re-link zeroes every uniform, so pass two proves its own writes rather than
+            // reading pass one's bytes back.
+            glLinkProgram(m_shapeProgram);
+            GLint linked = 0;
+            glGetProgramiv(m_shapeProgram, GL_LINK_STATUS, &linked);
+            ASSERT_EQ(linked, GL_TRUE);
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_shapeOutput);
+            glBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, kAllShapeSlots * sizeof(float), zeroes.data());
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+            setWithUniform();
+            expectEverything(dispatchAndRead(), "glUniform* after re-link");
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        TEST_F(DoublePrecisionScenario, TheConformanceUniformShaderAgreesWithEveryValueItWasGiven) {
+            if (!Ready()) return;
+            m_shapeProgram = glCreateProgram();
+            ASSERT_NE(m_shapeProgram, 0u);
+            {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glAttachShader(m_shapeProgram, shader);
+                glDeleteShader(shader);
+                const char* const sources[2] = {"#version 430 core\n", kCtsShapedSource};
+                glShaderSource(shader, 2, sources, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    FAIL() << "compute shader did not compile: " << log;
+                }
+            }
+            glLinkProgram(m_shapeProgram);
+            GLint linked = 0;
+            glGetProgramiv(m_shapeProgram, GL_LINK_STATUS, &linked);
+            if (linked == GL_FALSE) {
+                char log[2048] = {};
+                glGetProgramInfoLog(m_shapeProgram, sizeof(log) - 1, nullptr, log);
+                FAIL() << "compute program did not link: " << log;
+            }
+
+            glGenBuffers(1, &m_shapeOutput);
+            const GLint seed = 123;
+            glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_shapeOutput);
+            glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(seed), &seed, GL_STATIC_DRAW);
+
+            const auto location = [&](const char* name) { return glGetUniformLocation(m_shapeProgram, name); };
+            glProgramUniform1d(m_shapeProgram, location("g_0"), kG0);
+            glProgramUniform2d(m_shapeProgram, location("g_1"), kG1[0], kG1[1]);
+            glProgramUniform3d(m_shapeProgram, location("g_2"), kG2[0], kG2[1], kG2[2]);
+            glProgramUniform4d(m_shapeProgram, location("g_3"), kG3[0], kG3[1], kG3[2], kG3[3]);
+            glProgramUniformMatrix2dv(m_shapeProgram, location("g_4"), 1, GL_FALSE, kG4);
+            glProgramUniformMatrix2x3dv(m_shapeProgram, location("g_5"), 1, GL_FALSE, kG5);
+            glProgramUniformMatrix2x4dv(m_shapeProgram, location("g_6"), 1, GL_FALSE, kG6);
+            glProgramUniformMatrix3x2dv(m_shapeProgram, location("g_7"), 1, GL_FALSE, kG7);
+            glProgramUniformMatrix3dv(m_shapeProgram, location("g_8"), 1, GL_FALSE, kG8);
+            glProgramUniformMatrix3x4dv(m_shapeProgram, location("g_9"), 1, GL_FALSE, kG9);
+            glProgramUniformMatrix4x2dv(m_shapeProgram, location("g_10"), 1, GL_FALSE, kG10);
+            glProgramUniformMatrix4x3dv(m_shapeProgram, location("g_11"), 1, GL_FALSE, kG11);
+            glProgramUniformMatrix4dv(m_shapeProgram, location("g_12"), 1, GL_FALSE, kG12);
+
+            glUseProgram(m_shapeProgram);
+            glDispatchCompute(1, 1, 1);
+            glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+
+            GLint disagreements = -1;
+            glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(disagreements), &disagreements);
+            for (int bit = 0; bit < 13; ++bit) {
+                EXPECT_EQ(disagreements & (1 << bit), 0)
+                    << kDoubleShapes[bit].name << " did not compare equal to the value it was given";
+            }
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
         TEST_F(DoublePrecisionScenario, TheFp64ExtensionIsNotAdvertised) {
             if (!Ready()) return;
             // The shader above compiled, linked and ran without the extension string, which is

MobileGL/MG_IntegrationTest/Scenarios/ImageFormatQualifierScenario.cpp

@@ -0,0 +1,314 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ImageFormatQualifierScenario.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 - AN IMAGE UNIFORM THAT DECLARES NO FORMAT.
+//
+// Desktop GLSL 4.2 lets a writeonly image declaration omit its format layout qualifier:
+//
+//     writeonly uniform uimage2D uni_image;      // legal desktop GLSL
+//
+// GLSL ES has no such relaxation; every image uniform must carry one, and Adreno says so as "all
+// images have to define layout format", which fails the whole program. That is what took the
+// compute half of KHR-GL4x.packed_depth_stencil.stencil_texturing.
+//
+// The only qualifier that is CORRECT to substitute is whatever glBindImageTexture named for the
+// unit that uniform addresses - GL requires the qualifier, the bind format and the texture's
+// internal format to belong to one format class - so the format is not knowable when the shader
+// is compiled, only when it is drawn with. Espryt therefore BAKES it into the program it
+// generates and keys that program on the (unit, format) pairs it baked
+// (BackendProgramObjectImpl::ImageUnitFormatsStillMatch, MG_Backend/DirectGLES).
+//
+// Three separate things follow from "the program is built against live binding state", and each
+// one is a case below:
+//
+//   1. the format reaches the shader at all, so the store lands where the texture is (Writes);
+//   2. binding a DIFFERENT format to the same unit rebuilds the program, rather than reusing one
+//      compiled against the old format (RebindToADifferentFormatRebuilds);
+//   3. an image bound for the FIRST time after the link works, i.e. the program built against
+//      "nothing bound yet" is not the one the dispatch runs (FirstBindAfterLinkRebuilds).
+//
+// Magma needs none of this - Vulkan takes an Unknown-format storage image given
+// shaderStorageImageWriteWithoutFormat, and the view format is resolved from the same bind state
+// at descriptor time - so every case here runs on both backends and must agree, which is what
+// makes the ES-only machinery falsifiable rather than merely exercised.
+
+#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 int kExtent = 4;
+        // The image unit is deliberately NOT 0 and the uniform declares no binding, so the unit
+        // has to travel through glUniform1i and be baked into the ESSL alongside the format -
+        // the two bakes share a rebuild key and a bug in either shows up as the wrong texel.
+        constexpr GLint kImageUnit = 1;
+
+        // KHR-GL4x.packed_depth_stencil.stencil_texturing's own image declaration, verbatim.
+        const char* kStoreSource = R"(#version 430 core
+
+layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+
+writeonly uniform uimage2D uni_image;
+
+void main()
+{
+    imageStore(uni_image, ivec2(gl_GlobalInvocationID.xy), uvec4(gl_GlobalInvocationID.x + 100u, 0u, 0u, 0u));
+}
+)";
+
+        class ImageFormatQualifierScenario : public ScenarioTest {
+        protected:
+            void TearDown() override {
+                if (!Ready()) return;
+                glUseProgram(0);
+                for (GLuint p : m_programs) glDeleteProgram(p);
+                for (GLuint t : m_textures) glDeleteTextures(1, &t);
+                m_programs.clear();
+                m_textures.clear();
+                GLint maxImageUnits = 0;
+                glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+                for (GLint unit = 0; unit < maxImageUnits; ++unit) {
+                    glBindImageTexture(static_cast<GLuint>(unit), 0, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R32UI);
+                }
+                while (glGetError() != GL_NO_ERROR) {
+                }
+            }
+
+            bool ImagesAreUsable() const {
+                GLint maxImageUnits = 0;
+                glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+                GLint maxComputeImageUniforms = 0;
+                glGetIntegerv(GL_MAX_COMPUTE_IMAGE_UNIFORMS, &maxComputeImageUniforms);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return maxImageUnits > kImageUnit && maxComputeImageUniforms >= 1;
+            }
+
+            GLuint MakeComputeProgram(const std::string& source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                const char* text = source.c_str();
+                glShaderSource(shader, 1, &text, nullptr);
+                glCompileShader(shader);
+                GLint compiled = GL_FALSE;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute shader did not compile: " << log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                m_programs.push_back(program);
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute program did not link: " << log;
+                    return 0;
+                }
+                return program;
+            }
+
+            GLuint MakeTexture(GLenum internalFormat) {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(GL_TEXTURE_2D, texture);
+                glTexStorage2D(GL_TEXTURE_2D, 1, internalFormat, kExtent, kExtent);
+                if (const GLenum error = FirstGLError()) {
+                    ADD_FAILURE() << "allocating storage errored with " << GLErrorName(error);
+                    return 0;
+                }
+                // Seeded to a value no dispatch writes, so "the store never happened" and "the
+                // store wrote the right thing" cannot be confused.
+                const std::vector<GLuint> zeros(static_cast<std::size_t>(kExtent) * kExtent * 4u, 0u);
+                glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, kExtent, kExtent,
+                                internalFormat == GL_RGBA32UI ? GL_RGBA_INTEGER : GL_RED_INTEGER, GL_UNSIGNED_INT,
+                                zeros.data());
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return texture;
+            }
+
+            // Texel (x, 0) of the texture's red channel, read back through the GL frontend rather
+            // than through a second image uniform: a defect in the format bake would be shared by
+            // a reader declared the same way and could cancel itself out.
+            GLuint ReadRedTexel(GLuint texture, GLenum internalFormat, int x) {
+                const bool rgba = internalFormat == GL_RGBA32UI;
+                std::vector<GLuint> texels(static_cast<std::size_t>(kExtent) * kExtent * (rgba ? 4u : 1u),
+                                           0xFFFFFFFFu);
+                glBindTexture(GL_TEXTURE_2D, texture);
+                glGetTexImage(GL_TEXTURE_2D, 0, rgba ? GL_RGBA_INTEGER : GL_RED_INTEGER, GL_UNSIGNED_INT,
+                              texels.data());
+                if (const GLenum error = FirstGLError()) {
+                    ADD_FAILURE() << "reading the image back errored with " << GLErrorName(error);
+                    return 0xFFFFFFFFu;
+                }
+                return texels[static_cast<std::size_t>(x) * (rgba ? 4u : 1u)];
+            }
+
+            void DispatchStore(GLuint program, GLuint texture, GLenum internalFormat) {
+                glBindImageTexture(static_cast<GLuint>(kImageUnit), texture, 0, GL_FALSE, 0, GL_WRITE_ONLY,
+                                   internalFormat);
+                ASSERT_EQ(FirstGLError(), 0u) << "glBindImageTexture errored";
+                glUseProgram(program);
+                const GLint location = glGetUniformLocation(program, "uni_image");
+                ASSERT_GE(location, 0) << "the image uniform was not reflected";
+                glUniform1i(location, kImageUnit);
+                ASSERT_EQ(FirstGLError(), 0u) << "assigning the image unit errored";
+                glDispatchCompute(kExtent, 1, 1);
+                glMemoryBarrier(GL_ALL_BARRIER_BITS);
+                EXPECT_EQ(FirstGLError(), 0u) << "the dispatch leaked a GL error";
+                glUseProgram(0);
+            }
+
+            std::vector<GLuint> m_programs;
+            std::vector<GLuint> m_textures;
+        };
+
+        // The defect itself. Without the bake the ES driver refuses the program outright and the
+        // texture keeps its seed - which is also exactly what a silently no-op dispatch looks
+        // like, and why the seed is a value no store writes.
+        TEST_F(ImageFormatQualifierScenario, AFormatlessWriteonlyImageWrites) {
+            if (!Ready()) GTEST_SKIP() << "no GL context";
+            if (!ImagesAreUsable()) GTEST_SKIP() << "no image load/store on this driver";
+
+            const GLuint program = MakeComputeProgram(kStoreSource);
+            const GLuint texture = MakeTexture(GL_R32UI);
+            if (program == 0 || texture == 0) return;
+
+            DispatchStore(program, texture, GL_R32UI);
+            for (int x = 0; x < kExtent; ++x) {
+                EXPECT_EQ(ReadRedTexel(texture, GL_R32UI, x), static_cast<GLuint>(x) + 100u)
+                    << "texel " << x << " of a format-less writeonly image did not take the store";
+            }
+        }
+
+        // The rebuild key. The SAME program is dispatched twice with a different format bound to
+        // its unit; a build keyed only on the link (or only on the image UNIT) would reuse the
+        // r32ui program for the rgba32ui texture, and the second half would come back seeded.
+        //
+        // What the SOFTWARE lanes cannot falsify: with the key disabled this case still passes on
+        // Mesa, because the reused r32ui declaration writes the red channel of an RGBA32UI image
+        // anyway - a format-class mismatch GL leaves undefined and that driver happens to absorb.
+        // FirstBindAfterLinkRebuilds below is the case that fails there, because the reused
+        // program was built with no format at all and never compiled. Both are kept: this one is
+        // the shape a strict driver is entitled to reject, and it is the shape the device runs.
+        TEST_F(ImageFormatQualifierScenario, RebindToADifferentFormatRebuilds) {
+            if (!Ready()) GTEST_SKIP() << "no GL context";
+            if (!ImagesAreUsable()) GTEST_SKIP() << "no image load/store on this driver";
+
+            const GLuint program = MakeComputeProgram(kStoreSource);
+            const GLuint first = MakeTexture(GL_R32UI);
+            const GLuint second = MakeTexture(GL_RGBA32UI);
+            if (program == 0 || first == 0 || second == 0) return;
+
+            DispatchStore(program, first, GL_R32UI);
+            for (int x = 0; x < kExtent; ++x) {
+                ASSERT_EQ(ReadRedTexel(first, GL_R32UI, x), static_cast<GLuint>(x) + 100u)
+                    << "the first format must work before the rebind can be blamed for anything";
+            }
+
+            DispatchStore(program, second, GL_RGBA32UI);
+            for (int x = 0; x < kExtent; ++x) {
+                EXPECT_EQ(ReadRedTexel(second, GL_RGBA32UI, x), static_cast<GLuint>(x) + 100u)
+                    << "texel " << x << ": the program was not rebuilt for the newly bound format";
+            }
+
+            // ...and back, so the rebuild is not a one-way door: returning to a format the
+            // program was once built against must build for it again, not resurrect a cache row.
+            const GLuint third = MakeTexture(GL_R32UI);
+            if (third == 0) return;
+            DispatchStore(program, third, GL_R32UI);
+            for (int x = 0; x < kExtent; ++x) {
+                EXPECT_EQ(ReadRedTexel(third, GL_R32UI, x), static_cast<GLuint>(x) + 100u)
+                    << "texel " << x << ": going back to the first format did not rebuild";
+            }
+        }
+
+        // Nothing is bound to the unit when the program links, so whatever the first build sees
+        // is not the format the dispatch needs. glBindImageTexture must not itself trigger a
+        // build - it is an entry point, and building there is the constraint
+        // glShaderStorageBlockBinding is held to as well - so the rebuild has to happen at the
+        // next dispatch preparation instead. This case fails either way round: no rebuild, or a
+        // build attempted from the entry point before the state settles.
+        TEST_F(ImageFormatQualifierScenario, FirstBindAfterLinkRebuilds) {
+            if (!Ready()) GTEST_SKIP() << "no GL context";
+            if (!ImagesAreUsable()) GTEST_SKIP() << "no image load/store on this driver";
+
+            const GLuint program = MakeComputeProgram(kStoreSource);
+            if (program == 0) return;
+
+            // Use it once with NOTHING bound to the unit, which is what makes the backend build
+            // against an empty binding. The dispatch writes nowhere and must not error.
+            glUseProgram(program);
+            const GLint location = glGetUniformLocation(program, "uni_image");
+            ASSERT_GE(location, 0);
+            glUniform1i(location, kImageUnit);
+            glDispatchCompute(kExtent, 1, 1);
+            glMemoryBarrier(GL_ALL_BARRIER_BITS);
+            EXPECT_EQ(FirstGLError(), 0u) << "dispatching with an unbound image unit must not error";
+            glUseProgram(0);
+
+            const GLuint texture = MakeTexture(GL_R32UI);
+            if (texture == 0) return;
+            DispatchStore(program, texture, GL_R32UI);
+            for (int x = 0; x < kExtent; ++x) {
+                EXPECT_EQ(ReadRedTexel(texture, GL_R32UI, x), static_cast<GLuint>(x) + 100u)
+                    << "texel " << x << ": the first bind after the link did not reach the shader";
+            }
+        }
+
+        // A DECLARED format is authoritative and the bake must never touch it - including when
+        // the texture behind the unit has a different (but class-compatible) internal format,
+        // which GL explicitly allows. If the bake ever overrode a declaration, this is the case
+        // that would go wrong while every other one stayed green.
+        TEST_F(ImageFormatQualifierScenario, ADeclaredFormatStillWins) {
+            if (!Ready()) GTEST_SKIP() << "no GL context";
+            if (!ImagesAreUsable()) GTEST_SKIP() << "no image load/store on this driver";
+
+            const GLuint program = MakeComputeProgram(R"(#version 430 core
+
+layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+
+layout (r32ui) writeonly uniform uimage2D uni_image;
+
+void main()
+{
+    imageStore(uni_image, ivec2(gl_GlobalInvocationID.xy), uvec4(gl_GlobalInvocationID.x + 100u, 0u, 0u, 0u));
+}
+)");
+            const GLuint texture = MakeTexture(GL_R32UI);
+            if (program == 0 || texture == 0) return;
+
+            DispatchStore(program, texture, GL_R32UI);
+            for (int x = 0; x < kExtent; ++x) {
+                EXPECT_EQ(ReadRedTexel(texture, GL_R32UI, x), static_cast<GLuint>(x) + 100u)
+                    << "texel " << x << ": a declared format stopped working";
+            }
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ImageTargetKindScenario.cpp

@@ -0,0 +1,550 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ImageTargetKindScenario.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 - ONE IMAGE TARGET KIND AT A TIME, THROUGH A COMPUTE DISPATCH.
+//
+// KHR-GL44.multi_bind.dispatch_bind_image_textures decomposed. That conformance case declares
+// ELEVEN image uniforms of eleven different target kinds in one compute shader, binds a texture
+// of the matching kind to each unit, sums one texel from every one of them and compares the sum
+// against N*(N-1)/2. It is a single pass/fail bit over eleven independent mechanisms: if any one
+// of them is wrong - or merely fails to compile - the case fails and says nothing about which.
+// That is what it did here, on both backends, for two waves.
+//
+// So the eleven are pulled apart into one case each. Each case declares ONE image uniform, binds
+// ONE texture and checks the value that comes back, so a failure names the target kind and the
+// direction. What the conformance case does with eleven at once, AllKindsInOneProgram at the
+// bottom still does - a defect that only appears when several kinds share a program is invisible
+// to the single-kind cases by construction.
+//
+// The shape is deliberately the conformance case's own, not a cleaner equivalent:
+//
+//   * r32ui / GL_R32UI throughout, 6x6x6 storage, one level, texel (0,0,0) read;
+//   * `layout (location = N, r32ui) readonly uniform` - an explicit uniform LOCATION, not a
+//     binding, with the image unit then assigned by glUniform1i. That combination is the one ES
+//     cannot express directly, because ES forbids glUniform1i on an image uniform and the unit
+//     has to be baked into the generated ESSL (RebindImageUniformsToFrontendUnits);
+//   * `layout (std140, ...) buffer` for the result block - legal, but unusual enough that a
+//     frontend could plausibly mishandle it. Mirroring it means a green scenario cannot be green
+//     for a reason the conformance case excludes;
+//   * glBindImageTexture with layered = GL_TRUE, which is what glBindImageTextures is specified
+//     to pass, and which is where a target kind whose layeredness a backend does not recognise
+//     goes wrong.
+//
+// MULTISAMPLE is the one kind that is not merely an emulation problem, and the conformance case
+// already knows it: it reads GL_MAX_IMAGE_SAMPLES and, when that is zero, substitutes a plain 2D
+// texture and a plain uimage2D for both multisample entries. MobileGL reports zero, so the
+// conformance case never asks it for a multisample image at all. The two cases below are kept
+// and skip on that same query, so the coverage is already written the day a backend advertises
+// them - and so the skip is a standing record of WHY the conformance case passes without them.
+
+#include <algorithm>
+#include <cstddef>
+#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 {
+
+        // The conformance case's own dimensions: one level, 6 on every axis (which is also
+        // exactly one cube's worth for a cube array), and a single texel read at the origin.
+        constexpr int kExtent = 6;
+        constexpr GLuint kFilledValue = 7u;
+        constexpr GLuint kStoredValue = 13u;
+
+        // Everything that differs between the eleven kinds, in one row.
+        struct TargetKind {
+            const char* name;      // this scenario's name for it, which failure messages carry
+            GLenum target;         // the GL texture target
+            const char* imageType; // the GLSL image uniform type
+            const char* coord;     // the coordinate expression imageLoad/imageStore takes
+            bool multisample;      // needs GL_MAX_IMAGE_SAMPLES > 0
+            bool buffer;           // storage comes from a buffer object, not TexStorage
+        };
+
+        constexpr TargetKind kKind1D{"1D", GL_TEXTURE_1D, "uimage1D", "0", false, false};
+        constexpr TargetKind kKind1DArray{"1DArray", GL_TEXTURE_1D_ARRAY, "uimage1DArray", "ivec2(0, 0)", false,
+                                          false};
+        constexpr TargetKind kKind2D{"2D", GL_TEXTURE_2D, "uimage2D", "ivec2(0, 0)", false, false};
+        constexpr TargetKind kKind2DArray{"2DArray", GL_TEXTURE_2D_ARRAY, "uimage2DArray", "ivec3(0, 0, 0)", false,
+                                          false};
+        constexpr TargetKind kKind3D{"3D", GL_TEXTURE_3D, "uimage3D", "ivec3(0, 0, 0)", false, false};
+        constexpr TargetKind kKindBuffer{"Buffer", GL_TEXTURE_BUFFER, "uimageBuffer", "0", false, true};
+        constexpr TargetKind kKindCube{"Cube", GL_TEXTURE_CUBE_MAP, "uimageCube", "ivec3(0, 0, 0)", false, false};
+        constexpr TargetKind kKindCubeArray{"CubeArray", GL_TEXTURE_CUBE_MAP_ARRAY, "uimageCubeArray",
+                                            "ivec3(0, 0, 0)", false, false};
+        constexpr TargetKind kKindRect{"Rect", GL_TEXTURE_RECTANGLE, "uimage2DRect", "ivec2(0, 0)", false, false};
+        constexpr TargetKind kKind2DMS{"2DMS", GL_TEXTURE_2D_MULTISAMPLE, "uimage2DMS", "ivec2(0, 0)", true, false};
+        constexpr TargetKind kKind2DMSArray{"2DMSArray", GL_TEXTURE_2D_MULTISAMPLE_ARRAY, "uimage2DMSArray",
+                                            "ivec3(0, 0, 0)", true, false};
+
+        // A multisample image load/store takes the sample index as an extra argument; no other
+        // kind does. Keeping that in one place stops the two spellings drifting apart.
+        std::string LoadExpression(const TargetKind& kind, const std::string& name) {
+            return "imageLoad(" + name + ", " + kind.coord + (kind.multisample ? ", 0)" : ")");
+        }
+
+        std::string StoreStatement(const TargetKind& kind, const std::string& name, const char* value) {
+            return "imageStore(" + name + ", " + kind.coord + (kind.multisample ? ", 0, uvec4(" : ", uvec4(") +
+                   value + ", 0, 0, 0));";
+        }
+
+        const char* kComputePrologue = "#version 440 core\n"
+                                       "\n"
+                                       "layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;\n"
+                                       "\n";
+
+        const char* kResultBlock = "layout (std140, binding = 0) buffer SSB {\n"
+                                   "    uint sum;\n"
+                                   "} ssb;\n"
+                                   "\n";
+
+        // The conformance case's shader, narrowed to a single image.
+        std::string SingleLoadSource(const TargetKind& kind) {
+            return std::string(kComputePrologue) + "layout (location = 0, r32ui) readonly uniform " + kind.imageType +
+                   " i0;\n" + kResultBlock + "void main()\n{\n    uvec4 v = " + LoadExpression(kind, "i0") +
+                   ";\n    ssb.sum = v.r;\n}\n";
+        }
+
+        // The other direction. Written as its own program rather than a read-write one so that a
+        // backend which gets the store right and the load wrong (or the reverse) is not able to
+        // cancel its own defect out.
+        std::string SingleStoreSource(const TargetKind& kind) {
+            return std::string(kComputePrologue) + "layout (location = 0, r32ui) writeonly uniform " +
+                   kind.imageType + " i0;\n\nvoid main()\n{\n    " + StoreStatement(kind, "i0", "13u") + "\n}\n";
+        }
+
+        class ImageTargetKindScenario : public ScenarioTest {
+        protected:
+            void TearDown() override {
+                if (!Ready()) return;
+                glUseProgram(0);
+                for (GLuint p : m_programs) glDeleteProgram(p);
+                for (GLuint t : m_textures) glDeleteTextures(1, &t);
+                for (GLuint b : m_buffers) glDeleteBuffers(1, &b);
+                m_programs.clear();
+                m_textures.clear();
+                m_buffers.clear();
+                // Leave no image unit bound. These scenarios share one context, and a stale image
+                // binding is exactly the kind of state that makes the NEXT scenario's failure
+                // impossible to reproduce on its own.
+                GLint maxImageUnits = 0;
+                glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+                for (GLint unit = 0; unit < maxImageUnits; ++unit) {
+                    glBindImageTexture(static_cast<GLuint>(unit), 0, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R32UI);
+                }
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, 0);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+            }
+
+            bool ImagesAreUsable() const {
+                GLint maxImageUnits = 0;
+                glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+                GLint maxComputeImageUniforms = 0;
+                glGetIntegerv(GL_MAX_COMPUTE_IMAGE_UNIFORMS, &maxComputeImageUniforms);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return maxImageUnits >= 1 && maxComputeImageUniforms >= 1;
+            }
+
+            // The conformance case's own multisample gate, asked the same way it asks it.
+            bool MultisampleImagesAreUsable() const {
+                GLint maxImageSamples = 0;
+                glGetIntegerv(GL_MAX_IMAGE_SAMPLES, &maxImageSamples);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return maxImageSamples > 0;
+            }
+
+            GLuint MakeComputeProgram(const std::string& source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                const char* text = source.c_str();
+                glShaderSource(shader, 1, &text, nullptr);
+                glCompileShader(shader);
+                GLint compiled = GL_FALSE;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute shader did not compile: " << log << "\nsource:\n" << source;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                m_programs.push_back(program);
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute program did not link: " << log << "\nsource:\n" << source;
+                    return 0;
+                }
+                return program;
+            }
+
+            // Storage plus a full fill with `value`, in the spelling each target kind needs.
+            // Returns 0 - having already reported - when the target could not be created.
+            GLuint MakeTexture(const TargetKind& kind, bool fill, GLuint value = kFilledValue) {
+                const std::vector<GLuint> texels(static_cast<std::size_t>(kExtent) * kExtent * kExtent, value);
+
+                if (kind.buffer) {
+                    GLuint buffer = 0;
+                    glGenBuffers(1, &buffer);
+                    m_buffers.push_back(buffer);
+                    glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+                    glBufferData(GL_TEXTURE_BUFFER, static_cast<GLsizeiptr>(texels.size() * sizeof(GLuint)),
+                                 fill ? texels.data() : nullptr, GL_DYNAMIC_COPY);
+                    GLuint texture = 0;
+                    glGenTextures(1, &texture);
+                    m_textures.push_back(texture);
+                    glBindTexture(GL_TEXTURE_BUFFER, texture);
+                    glTexBuffer(GL_TEXTURE_BUFFER, GL_R32UI, buffer);
+                    if (const GLenum error = FirstGLError()) {
+                        ADD_FAILURE() << kind.name << ": creating the texture buffer errored with "
+                                      << GLErrorName(error);
+                        return 0;
+                    }
+                    return texture;
+                }
+
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(kind.target, texture);
+
+                switch (kind.target) {
+                case GL_TEXTURE_1D:
+                    glTexStorage1D(kind.target, 1, GL_R32UI, kExtent);
+                    break;
+                case GL_TEXTURE_2D:
+                case GL_TEXTURE_RECTANGLE:
+                case GL_TEXTURE_1D_ARRAY:
+                case GL_TEXTURE_CUBE_MAP:
+                    glTexStorage2D(kind.target, 1, GL_R32UI, kExtent, kExtent);
+                    break;
+                case GL_TEXTURE_2D_ARRAY:
+                case GL_TEXTURE_3D:
+                case GL_TEXTURE_CUBE_MAP_ARRAY:
+                    glTexStorage3D(kind.target, 1, GL_R32UI, kExtent, kExtent, kExtent);
+                    break;
+                case GL_TEXTURE_2D_MULTISAMPLE:
+                    glTexStorage2DMultisample(kind.target, 1, GL_R32UI, kExtent, kExtent, GL_FALSE);
+                    break;
+                case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+                    glTexStorage3DMultisample(kind.target, 1, GL_R32UI, kExtent, kExtent, kExtent, GL_FALSE);
+                    break;
+                default:
+                    ADD_FAILURE() << kind.name << ": no storage spelling for target 0x" << std::hex << kind.target;
+                    return 0;
+                }
+                if (const GLenum error = FirstGLError()) {
+                    ADD_FAILURE() << kind.name << ": allocating storage errored with " << GLErrorName(error);
+                    return 0;
+                }
+
+                // A multisample texture has no TexSubImage - the conformance case fills it with a
+                // compute pass, which is what the store cases below do.
+                if (!fill || kind.multisample) return texture;
+
+                switch (kind.target) {
+                case GL_TEXTURE_1D:
+                    glTexSubImage1D(kind.target, 0, 0, kExtent, GL_RED_INTEGER, GL_UNSIGNED_INT, texels.data());
+                    break;
+                case GL_TEXTURE_2D:
+                case GL_TEXTURE_RECTANGLE:
+                case GL_TEXTURE_1D_ARRAY:
+                    glTexSubImage2D(kind.target, 0, 0, 0, kExtent, kExtent, GL_RED_INTEGER, GL_UNSIGNED_INT,
+                                    texels.data());
+                    break;
+                case GL_TEXTURE_CUBE_MAP:
+                    for (int face = 0; face < 6; ++face) {
+                        glTexSubImage2D(static_cast<GLenum>(GL_TEXTURE_CUBE_MAP_POSITIVE_X + face), 0, 0, 0, kExtent,
+                                        kExtent, GL_RED_INTEGER, GL_UNSIGNED_INT, texels.data());
+                    }
+                    break;
+                case GL_TEXTURE_2D_ARRAY:
+                case GL_TEXTURE_3D:
+                case GL_TEXTURE_CUBE_MAP_ARRAY:
+                    glTexSubImage3D(kind.target, 0, 0, 0, 0, kExtent, kExtent, kExtent, GL_RED_INTEGER,
+                                    GL_UNSIGNED_INT, texels.data());
+                    break;
+                default:
+                    break;
+                }
+                if (const GLenum error = FirstGLError()) {
+                    ADD_FAILURE() << kind.name << ": uploading texels errored with " << GLErrorName(error);
+                    return 0;
+                }
+                return texture;
+            }
+
+            // A 4-byte `buffer` block bound to base 0, which is where every case puts its answer.
+            GLuint MakeResultBuffer() {
+                GLuint ssbo = 0;
+                glGenBuffers(1, &ssbo);
+                m_buffers.push_back(ssbo);
+                const GLuint zero = 0u;
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
+                glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(GLuint), &zero, GL_DYNAMIC_COPY);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, ssbo);
+                return ssbo;
+            }
+
+            GLuint ReadResult(GLuint ssbo) {
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
+                GLuint value = 0xFFFFFFFFu;
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(GLuint), &value);
+                return value;
+            }
+
+            // Fill a texture of `kind`, read texel (0,0,0) of it through an image uniform in a
+            // compute dispatch, and require the value back.
+            void RunLoadCase(const TargetKind& kind) {
+                const GLuint program = MakeComputeProgram(SingleLoadSource(kind));
+                if (program == 0) return;
+                const GLuint texture = MakeTexture(kind, true);
+                if (texture == 0) return;
+                const GLuint ssbo = MakeResultBuffer();
+
+                glBindImageTexture(0, texture, 0, GL_TRUE, 0, GL_READ_ONLY, GL_R32UI);
+                ASSERT_EQ(FirstGLError(), 0u) << kind.name << ": glBindImageTexture errored";
+
+                glUseProgram(program);
+                // The unit, by LOCATION - the conformance case's own redundant-but-legal
+                // assignment, and the one ES cannot take at the API level.
+                glUniform1i(0, 0);
+                ASSERT_EQ(FirstGLError(), 0u) << kind.name << ": assigning the image unit errored";
+
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_ALL_BARRIER_BITS);
+                EXPECT_EQ(FirstGLError(), 0u) << kind.name << ": the dispatch leaked a GL error";
+
+                EXPECT_EQ(ReadResult(ssbo), kFilledValue)
+                    << kind.name << ": the compute dispatch did not read the value the texture was filled with";
+                glUseProgram(0);
+            }
+
+            // The other direction: store through an image uniform, then read the same texel back
+            // through a SECOND program, so a defect cannot cancel itself out.
+            void RunStoreCase(const TargetKind& kind) {
+                const GLuint storeProgram = MakeComputeProgram(SingleStoreSource(kind));
+                const GLuint loadProgram = MakeComputeProgram(SingleLoadSource(kind));
+                if (storeProgram == 0 || loadProgram == 0) return;
+                const GLuint texture = MakeTexture(kind, false);
+                if (texture == 0) return;
+                const GLuint ssbo = MakeResultBuffer();
+
+                glBindImageTexture(0, texture, 0, GL_TRUE, 0, GL_READ_WRITE, GL_R32UI);
+                ASSERT_EQ(FirstGLError(), 0u) << kind.name << ": glBindImageTexture errored";
+
+                glUseProgram(storeProgram);
+                glUniform1i(0, 0);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_ALL_BARRIER_BITS);
+                EXPECT_EQ(FirstGLError(), 0u) << kind.name << ": the storing dispatch leaked a GL error";
+
+                glUseProgram(loadProgram);
+                glUniform1i(0, 0);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_ALL_BARRIER_BITS);
+                EXPECT_EQ(FirstGLError(), 0u) << kind.name << ": the loading dispatch leaked a GL error";
+
+                EXPECT_EQ(ReadResult(ssbo), kStoredValue)
+                    << kind.name << ": the value stored through the image did not come back";
+                glUseProgram(0);
+            }
+
+            std::vector<GLuint> m_programs;
+            std::vector<GLuint> m_textures;
+            std::vector<GLuint> m_buffers;
+        };
+
+    } // namespace
+
+    // ---- the load direction, one target kind per case -----------------------
+    //
+    // Exactly what the conformance case does with each of its eleven uniforms, but alone, so a
+    // failure names the kind.
+
+#define MGL_DEFINE_LOAD_CASE(CaseName, Kind)                                                                    \
+    TEST_F(ImageTargetKindScenario, Loads##CaseName) {                                                          \
+        if (!Ready()) return;                                                                                   \
+        if (!ImagesAreUsable()) GTEST_SKIP() << "no compute image uniforms";                                    \
+        if ((Kind).multisample && !MultisampleImagesAreUsable()) {                                              \
+            GTEST_SKIP() << "GL_MAX_IMAGE_SAMPLES is 0, so the conformance case substitutes a plain 2D image "   \
+                            "here and never asks for a multisample one";                                        \
+        }                                                                                                       \
+        RunLoadCase(Kind);                                                                                      \
+    }
+
+#define MGL_DEFINE_STORE_CASE(CaseName, Kind)                                                                   \
+    TEST_F(ImageTargetKindScenario, Stores##CaseName) {                                                         \
+        if (!Ready()) return;                                                                                   \
+        if (!ImagesAreUsable()) GTEST_SKIP() << "no compute image uniforms";                                    \
+        if ((Kind).multisample && !MultisampleImagesAreUsable()) {                                              \
+            GTEST_SKIP() << "GL_MAX_IMAGE_SAMPLES is 0, so the conformance case substitutes a plain 2D image "   \
+                            "here and never asks for a multisample one";                                        \
+        }                                                                                                       \
+        RunStoreCase(Kind);                                                                                      \
+    }
+
+    MGL_DEFINE_LOAD_CASE(Texture1D, kKind1D)
+    MGL_DEFINE_LOAD_CASE(Texture1DArray, kKind1DArray)
+    MGL_DEFINE_LOAD_CASE(Texture2D, kKind2D)
+    MGL_DEFINE_LOAD_CASE(Texture2DArray, kKind2DArray)
+    MGL_DEFINE_LOAD_CASE(Texture3D, kKind3D)
+    MGL_DEFINE_LOAD_CASE(TextureBuffer, kKindBuffer)
+    MGL_DEFINE_LOAD_CASE(TextureCube, kKindCube)
+    MGL_DEFINE_LOAD_CASE(TextureCubeArray, kKindCubeArray)
+    MGL_DEFINE_LOAD_CASE(TextureRectangle, kKindRect)
+    MGL_DEFINE_LOAD_CASE(Texture2DMultisample, kKind2DMS)
+    MGL_DEFINE_LOAD_CASE(Texture2DMultisampleArray, kKind2DMSArray)
+
+    MGL_DEFINE_STORE_CASE(Texture1D, kKind1D)
+    MGL_DEFINE_STORE_CASE(Texture1DArray, kKind1DArray)
+    MGL_DEFINE_STORE_CASE(Texture2D, kKind2D)
+    MGL_DEFINE_STORE_CASE(Texture2DArray, kKind2DArray)
+    MGL_DEFINE_STORE_CASE(Texture3D, kKind3D)
+    MGL_DEFINE_STORE_CASE(TextureBuffer, kKindBuffer)
+    MGL_DEFINE_STORE_CASE(TextureCube, kKindCube)
+    MGL_DEFINE_STORE_CASE(TextureCubeArray, kKindCubeArray)
+    MGL_DEFINE_STORE_CASE(TextureRectangle, kKindRect)
+    MGL_DEFINE_STORE_CASE(Texture2DMultisample, kKind2DMS)
+    MGL_DEFINE_STORE_CASE(Texture2DMultisampleArray, kKind2DMSArray)
+
+#undef MGL_DEFINE_LOAD_CASE
+#undef MGL_DEFINE_STORE_CASE
+
+    // ---- and all of them at once -------------------------------------------
+    //
+    // The conformance case's actual shape. The single-kind cases above cannot see a defect that
+    // needs several kinds in one program - a binding remap that only collides when two image
+    // types share a descriptor set, a per-kind rewrite that is not idempotent across declarations
+    // - and that class of defect is precisely what "each kind passes alone but the case still
+    // fails" would mean.
+    //
+    // Each unit is filled with its own DISTINCT value rather than a shared one, so a shortfall
+    // names WHICH kind is missing rather than merely how many are: with one shared value, "three
+    // kinds read zero" and "one kind read zero" differ only by a multiple, and any two kinds are
+    // interchangeable in the total. A sum still cannot see two kinds SWAPPING - addition is
+    // commutative, and the conformance case has exactly the same blind spot - but the single-kind
+    // cases above pin each kind to its own texture already, so a swap cannot hide there.
+    TEST_F(ImageTargetKindScenario, AllKindsInOneProgram) {
+        if (!Ready()) return;
+        if (!ImagesAreUsable()) GTEST_SKIP() << "no compute image uniforms";
+
+        // The two kinds this whole scenario file exists for come FIRST, and that ordering is
+        // load-bearing rather than cosmetic. The list has to be truncated to the device's image
+        // unit count, and the guaranteed minimum is small - ES 3.1 promises only four compute
+        // image uniforms - so a list in the conformance case's own order would put imageBuffer
+        // at index five and drop it on exactly the devices most likely to get it wrong. A test
+        // that quietly stops covering its own subject is worse than one that fails.
+        const bool multisample = MultisampleImagesAreUsable();
+        std::vector<TargetKind> kinds{kKind1DArray, kKindBuffer, kKind2D,   kKind1D,      kKind2DArray,
+                                      kKind3D,      kKindCube,   kKindRect, kKindCubeArray};
+        if (multisample) {
+            kinds.push_back(kKind2DMS);
+            kinds.push_back(kKind2DMSArray);
+        }
+
+        GLint maxComputeImageUniforms = 0;
+        glGetIntegerv(GL_MAX_COMPUTE_IMAGE_UNIFORMS, &maxComputeImageUniforms);
+        GLint maxImageUnits = 0;
+        glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+        while (glGetError() != GL_NO_ERROR) {
+        }
+        const std::size_t count =
+            std::min<std::size_t>(kinds.size(), static_cast<std::size_t>(std::max(0, std::min(maxComputeImageUniforms,
+                                                                                              maxImageUnits))));
+        if (count == 0) GTEST_SKIP() << "no image units";
+        // Named, not silently dropped: `expected` is computed over whatever survives, so a
+        // truncated run is self-consistently green and would otherwise never say what it stopped
+        // covering.
+        if (count < kinds.size()) {
+            std::string dropped;
+            for (std::size_t i = count; i < kinds.size(); ++i) {
+                if (!dropped.empty()) dropped += ", ";
+                dropped += kinds[i].name;
+            }
+            RecordProperty("dropped_image_target_kinds", dropped);
+            GTEST_LOG_(INFO) << "only " << count << " image units, so these kinds are not covered by the "
+                             << "combined case: " << dropped;
+        }
+        kinds.resize(count);
+
+        std::string declarations;
+        std::string sum;
+        for (std::size_t i = 0; i < kinds.size(); ++i) {
+            const std::string name = "i" + std::to_string(i);
+            declarations += "layout (location = " + std::to_string(i) + ", r32ui) readonly uniform " +
+                            kinds[i].imageType + " " + name + ";\n";
+            if (!sum.empty()) sum += " + ";
+            sum += LoadExpression(kinds[i], name);
+        }
+        const std::string source = std::string(kComputePrologue) + declarations + kResultBlock +
+                                   "void main()\n{\n    uvec4 v = " + sum + ";\n    ssb.sum = v.r;\n}\n";
+
+        const GLuint program = MakeComputeProgram(source);
+        if (program == 0) return;
+
+        // Powers of two, so the shortfall's bit pattern names exactly which kinds read zero -
+        // no other subset of the values can sum to the same total. Eleven kinds at most, so the
+        // largest is 1 << 10 and the sum cannot approach a uint's range.
+        GLuint expected = 0;
+        for (std::size_t i = 0; i < kinds.size(); ++i) {
+            const GLuint value = 1u << i;
+            const GLuint texture = MakeTexture(kinds[i], true, value);
+            if (texture == 0) return;
+            expected += value;
+            glBindImageTexture(static_cast<GLuint>(i), texture, 0, GL_TRUE, 0, GL_READ_ONLY, GL_R32UI);
+            ASSERT_EQ(FirstGLError(), 0u) << kinds[i].name << ": glBindImageTexture errored";
+        }
+        const GLuint ssbo = MakeResultBuffer();
+
+        glUseProgram(program);
+        for (std::size_t i = 0; i < kinds.size(); ++i) {
+            glUniform1i(static_cast<GLint>(i), static_cast<GLint>(i));
+        }
+        ASSERT_EQ(FirstGLError(), 0u) << "assigning the image units errored";
+
+        glDispatchCompute(1, 1, 1);
+        glMemoryBarrier(GL_ALL_BARRIER_BITS);
+        EXPECT_EQ(FirstGLError(), 0u) << "the dispatch leaked a GL error";
+
+        const GLuint actual = ReadResult(ssbo);
+        std::string missing;
+        for (std::size_t i = 0; i < kinds.size(); ++i) {
+            if ((actual & (1u << i)) == 0u) {
+                if (!missing.empty()) missing += ", ";
+                missing += kinds[i].name;
+            }
+        }
+        EXPECT_EQ(actual, expected)
+            << "the sum over " << kinds.size()
+            << " image target kinds is wrong; each kind contributes its own bit, and these read "
+               "zero: "
+            << (missing.empty() ? "(none - so some kind read a value it was never given)" : missing);
+        glUseProgram(0);
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/IterationRPFirstReductionScenario.cpp

@@ -0,0 +1,898 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/IterationRPFirstReductionScenario.cpp
+// Copyright (c) 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 - ITERATIONRP'S FIRST SUBGROUP REDUCTION.
+//
+// iterationRP reduces a 32 x 16 exposure tile with a vector subgroup inclusive add,
+// then a shared-memory scan of subgroup totals.  The source assumes that every
+// subgroup has a last lane, that there are 2..32 subgroups, and that local index
+// 511 belongs to the last subgroup and its last lane.  Those are source assumptions,
+// not API contracts.  This probe intentionally does not repair them: it records the
+// observed topology and makes each handoff independently observable.
+
+#include <algorithm>
+#include <array>
+#include <bit>
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+#include <iomanip>
+#include <iostream>
+#include <limits>
+#include <sstream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#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 std::size_t kInvocationCount = 512;
+        constexpr std::size_t kScanStageCount = 6;
+        constexpr std::uint32_t kQuietNanBits = 0x7fc00000u;
+        constexpr std::size_t kNoSlot = std::numeric_limits<std::size_t>::max();
+
+        struct UVec4 {
+            std::uint32_t x;
+            std::uint32_t y;
+            std::uint32_t z;
+            std::uint32_t w;
+        };
+
+        struct Vec4 {
+            float x;
+            float y;
+            float z;
+            float w;
+        };
+
+        // Matches the std430 block exactly. uvec4/vec4 arrays have a 16-byte
+        // stride, floats are a dense scalar array, and the outer scan array is
+        // stage-major in both GLSL and C++.
+        struct ProbeOutput {
+            std::array<UVec4, kInvocationCount> invocation;
+            std::array<UVec4, kInvocationCount> subgroup;
+            std::array<Vec4, kInvocationCount> reduction;
+            std::array<float, kInvocationCount> finalAverage;
+            std::array<std::array<float, kInvocationCount>, kScanStageCount> scanAfter;
+        };
+
+        static_assert(sizeof(UVec4) == 16);
+        static_assert(sizeof(Vec4) == 16);
+        static_assert(std::is_standard_layout_v<ProbeOutput>);
+        static_assert(offsetof(ProbeOutput, invocation) == 0);
+        static_assert(offsetof(ProbeOutput, subgroup) == 8192);
+        static_assert(offsetof(ProbeOutput, reduction) == 16384);
+        static_assert(offsetof(ProbeOutput, finalAverage) == 24576);
+        static_assert(offsetof(ProbeOutput, scanAfter) == 26624);
+        static_assert(sizeof(ProbeOutput) == 38912);
+
+        enum class InputMode {
+            SampledRgba32f,
+            IndexedSsbo,
+        };
+
+        const char* InputModeName(InputMode mode) {
+            return mode == InputMode::SampledRgba32f ? "sampled RGBA32F" : "indexed SSBO";
+        }
+
+        std::uint32_t FloatBits(float value) {
+            return std::bit_cast<std::uint32_t>(value);
+        }
+
+        bool SameBits(float lhs, float rhs) {
+            return FloatBits(lhs) == FloatBits(rhs);
+        }
+
+        bool IsQuietNanSentinel(float value) {
+            return FloatBits(value) == kQuietNanBits;
+        }
+
+        bool DrainGlErrors() {
+            bool hadError = false;
+            while (glGetError() != GL_NO_ERROR) hadError = true;
+            return hadError;
+        }
+
+        bool HasExtension(const char* wanted) {
+            GLint extensionCount = 0;
+            glGetIntegerv(GL_NUM_EXTENSIONS, &extensionCount);
+            for (GLint i = 0; i < extensionCount; ++i) {
+                const auto* extension = reinterpret_cast<const char*>(glGetStringi(GL_EXTENSIONS, static_cast<GLuint>(i)));
+                if (extension != nullptr && std::string(extension) == wanted) return true;
+            }
+            return false;
+        }
+
+        struct CapabilityInfo {
+            bool subgroupExtension = false;
+            GLint subgroupSize = 0;
+            GLint supportedStages = 0;
+            GLint supportedFeatures = 0;
+            GLint maxComputeStorageBlocks = 0;
+            GLint maxStorageBindings = 0;
+            GLint maxWorkGroupInvocations = 0;
+            std::array<GLint, 3> maxWorkGroupSize{};
+            bool queryHadError = false;
+
+            // iterationRP's source contract needs gl_NumSubgroups in [2, 32] for its 512
+            // invocations, i.e. an advertised subgroup width in [16, 256]. A device
+            // outside that window (lavapipe's 8-lane subgroups give 64 subgroups) cannot
+            // run the fixture's verbatim reduction at all, so the scenario SKIPS there -
+            // the pack itself replays through the FixIterationRPSubgroupScratch patch, which
+            // this probe deliberately does not model. The width only gates the domain;
+            // lane placement and group counts still come from observed values alone.
+            bool SubgroupWidthInSourceDomain() const {
+                return subgroupSize >= 16 && subgroupSize <= 256;
+            }
+
+            bool SupportsProbe() const {
+                const auto stages = static_cast<GLbitfield>(supportedStages);
+                const auto features = static_cast<GLbitfield>(supportedFeatures);
+                return !queryHadError && subgroupExtension &&
+                       (stages & GL_COMPUTE_SHADER_BIT) != 0 &&
+                       (features & (GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR)) ==
+                           (GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR) &&
+                       SubgroupWidthInSourceDomain() &&
+                       maxComputeStorageBlocks >= 2 && maxStorageBindings >= 2 &&
+                       maxWorkGroupInvocations >= static_cast<GLint>(kInvocationCount) && maxWorkGroupSize[0] >= 32 &&
+                       maxWorkGroupSize[1] >= 16 && maxWorkGroupSize[2] >= 1;
+            }
+
+            std::string MissingRequirements() const {
+                std::vector<std::string> missing;
+                const auto stages = static_cast<GLbitfield>(supportedStages);
+                const auto features = static_cast<GLbitfield>(supportedFeatures);
+                if (queryHadError) missing.emplace_back("a subgroup/compute capability query generated GL error");
+                if (!subgroupExtension) missing.emplace_back("GL_KHR_shader_subgroup");
+                if ((stages & GL_COMPUTE_SHADER_BIT) == 0) {
+                    missing.emplace_back("GL_COMPUTE_SHADER_BIT in GL_SUBGROUP_SUPPORTED_STAGES_KHR");
+                }
+                const auto requiredFeatures =
+                    GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR;
+                if ((features & requiredFeatures) != requiredFeatures) {
+                    missing.emplace_back("basic|arithmetic in GL_SUBGROUP_SUPPORTED_FEATURES_KHR");
+                }
+                if (!SubgroupWidthInSourceDomain()) {
+                    missing.emplace_back(
+                        "GL_SUBGROUP_SIZE_KHR in [16, 256] (iterationRP's source contract needs "
+                        "gl_NumSubgroups in [2, 32] for 512 invocations; width " +
+                        std::to_string(subgroupSize) + " is outside the fixture's domain)");
+                }
+                if (maxComputeStorageBlocks < 2 || maxStorageBindings < 2) {
+                    missing.emplace_back("two compute SSBO bindings");
+                }
+                if (maxWorkGroupInvocations < static_cast<GLint>(kInvocationCount) || maxWorkGroupSize[0] < 32 ||
+                    maxWorkGroupSize[1] < 16 || maxWorkGroupSize[2] < 1) {
+                    missing.emplace_back("a 32x16x1 / 512-invocation compute workgroup");
+                }
+
+                std::ostringstream message;
+                for (std::size_t i = 0; i < missing.size(); ++i) {
+                    if (i != 0) message << ", ";
+                    message << missing[i];
+                }
+                return message.str();
+            }
+        };
+
+        CapabilityInfo QueryCapabilities() {
+            CapabilityInfo info;
+            DrainGlErrors();
+            info.subgroupExtension = HasExtension("GL_KHR_shader_subgroup");
+            glGetIntegerv(GL_SUBGROUP_SIZE_KHR, &info.subgroupSize);
+            glGetIntegerv(GL_SUBGROUP_SUPPORTED_STAGES_KHR, &info.supportedStages);
+            glGetIntegerv(GL_SUBGROUP_SUPPORTED_FEATURES_KHR, &info.supportedFeatures);
+            glGetIntegerv(GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS, &info.maxComputeStorageBlocks);
+            glGetIntegerv(GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS, &info.maxStorageBindings);
+            glGetIntegerv(GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, &info.maxWorkGroupInvocations);
+            for (GLuint axis = 0; axis < info.maxWorkGroupSize.size(); ++axis) {
+                glGetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_SIZE, axis, &info.maxWorkGroupSize[axis]);
+            }
+            info.queryHadError = DrainGlErrors();
+            return info;
+        }
+
+        void PrintMetadata(const CapabilityInfo& info, std::ostream& output) {
+            output << "IterationRPFirstReductionScenario metadata: "
+                   << "GL_SUBGROUP_SIZE_KHR=" << info.subgroupSize
+                   << ", GL_SUBGROUP_SUPPORTED_STAGES_KHR=0x" << std::hex
+                   << static_cast<GLbitfield>(info.supportedStages)
+                   << ", GL_SUBGROUP_SUPPORTED_FEATURES_KHR=0x"
+                   << static_cast<GLbitfield>(info.supportedFeatures) << std::dec
+                   << ", subgroupExtension=" << info.subgroupExtension
+                   << ", GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS=" << info.maxComputeStorageBlocks
+                   << ", GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS=" << info.maxStorageBindings
+                   << ", GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS=" << info.maxWorkGroupInvocations
+                   << ", GL_MAX_COMPUTE_WORK_GROUP_SIZE=" << info.maxWorkGroupSize[0] << 'x'
+                   << info.maxWorkGroupSize[1] << 'x' << info.maxWorkGroupSize[2]
+                   << ", queryHadError=" << info.queryHadError << '\n';
+        }
+
+        bool DumpRequested() {
+            const char* value = std::getenv("MOBILEGL_ITEST_SUBGROUP_PROBE_DUMP");
+            return value != nullptr && std::string(value) == "1";
+        }
+
+        constexpr const char* kShaderPreamble = R"(#version 430 core
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x = 32, local_size_y = 16, local_size_z = 1) in;
+
+layout(std430, binding = 1) buffer SubgroupProbeOutput {
+    uvec4 invocation[512];
+    uvec4 subgroup[512];
+    vec4 reduction[512];
+    float finalAverage[512];
+    float scanAfter[6][512];
+} outProbe;
+
+shared vec2 prefixSumCache[32];
+)";
+
+        constexpr const char* kSampledInput = R"(
+uniform sampler2D colortex2;
+uniform vec2 pixelSize;
+)";
+
+        constexpr const char* kIndexedInput = R"(
+layout(std430, binding = 0) readonly buffer Input {
+    float value[512];
+} inputData;
+)";
+
+        // Only the expression producing tileExposure differs between the two
+        // tests. The remainder is the iterationRP first reduction, with stores
+        // placed after its existing barriers to expose each handoff.
+        constexpr const char* kSampledTileExposure = R"(
+    vec2 texCoord = (vec2(gl_GlobalInvocationID.xy) + 0.5) *
+                    vec2(1.0 / 32.0, 1.0 / 16.0);
+    vec2 sampleCoord = texCoord * (1.0 / 64.0);
+    sampleCoord.x += (15.0 / 32.0) + pixelSize.x * 12.0;
+
+    float tileExposure = dot(
+        textureLod(colortex2, sampleCoord, 0.0).rgb,
+        vec3(0.2125, 0.7154, 0.0721));
+)";
+
+        constexpr const char* kIndexedTileExposure = R"(
+    float tileExposure = inputData.value[gl_LocalInvocationIndex];
+)";
+
+        constexpr const char* kReductionBody = R"(
+    vec2 sampleLuminance = vec2(tileExposure, 0.0);
+    sampleLuminance = subgroupInclusiveAdd(sampleLuminance);
+    float nativeInclusive = sampleLuminance.x;
+
+    // This is a uniform, safety-only branch: it leaves an invalid source
+    // contract visible without indexing past the 32-entry cache or underflowing
+    // loopLength - 1. It is deliberately a failure on the CPU, not a skip.
+    bool sourceDomain = gl_NumSubgroups >= 2u && gl_NumSubgroups <= 32u;
+    if (!sourceDomain) {
+        float qNaN = uintBitsToFloat(0x7fc00000u);
+        uint localIndex = gl_LocalInvocationIndex;
+        outProbe.invocation[localIndex] = uvec4(localIndex, gl_LocalInvocationID);
+        outProbe.subgroup[localIndex] = uvec4(gl_SubgroupSize, gl_NumSubgroups, gl_SubgroupID,
+                                               gl_SubgroupInvocationID);
+        outProbe.reduction[localIndex] = vec4(tileExposure, nativeInclusive, qNaN, qNaN);
+        outProbe.finalAverage[localIndex] = qNaN;
+        for (uint stage = 0u; stage < 6u; ++stage)
+            outProbe.scanAfter[stage][localIndex] = qNaN;
+        return;
+    }
+
+    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+        prefixSumCache[gl_SubgroupID] = sampleLuminance;
+    barrier();
+
+    float sourceRawSubtotal = prefixSumCache[gl_SubgroupID].x;
+
+    uint loopLength = uint(findMSB(gl_NumSubgroups));
+    loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
+
+    for (uint scanStage = 0u; scanStage < loopLength; ++scanStage) {
+        if ((gl_SubgroupID & (1u << scanStage)) > 0u) {
+            sampleLuminance += prefixSumCache[(gl_SubgroupID >> scanStage << scanStage) - 1u];
+            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+                prefixSumCache[gl_SubgroupID] = sampleLuminance;
+        }
+        barrier();
+        outProbe.scanAfter[scanStage][gl_LocalInvocationIndex] = sampleLuminance.x;
+    }
+
+    float sourceMergedPrefix = sampleLuminance.x;
+
+    if (gl_LocalInvocationIndex == 511u)
+        prefixSumCache[0] = sampleLuminance / 512.0;
+    barrier();
+
+    float avg = prefixSumCache[0].x;
+
+    uint localIndex = gl_LocalInvocationIndex;
+    outProbe.invocation[localIndex] = uvec4(localIndex, gl_LocalInvocationID);
+    outProbe.subgroup[localIndex] = uvec4(gl_SubgroupSize, gl_NumSubgroups, gl_SubgroupID,
+                                           gl_SubgroupInvocationID);
+    outProbe.reduction[localIndex] = vec4(tileExposure, nativeInclusive, sourceRawSubtotal, sourceMergedPrefix);
+    outProbe.finalAverage[localIndex] = avg;
+}
+)";
+
+        std::string BuildProbeShader(InputMode mode) {
+            std::string source = kShaderPreamble;
+            source += mode == InputMode::SampledRgba32f ? kSampledInput : kIndexedInput;
+            source += "\nvoid main() {\n";
+            source += mode == InputMode::SampledRgba32f ? kSampledTileExposure : kIndexedTileExposure;
+            source += kReductionBody;
+            return source;
+        }
+
+        std::string FormatFloat(float value) {
+            std::ostringstream text;
+            text << std::hexfloat << value;
+            return text.str();
+        }
+
+        struct ValidationResult {
+            bool ok = true;
+            std::string phase;
+            std::string message;
+            bool scanStageMismatch = false;
+            int scanStage = -1;
+            bool ownerEvaluated = false;
+            bool index511IsSourceLastLaneWriter = false;
+            bool index511IsHighestSubgroupMember = false;
+            std::uint32_t highestObservedSubgroup = 0;
+        };
+
+        ValidationResult Failure(std::string phase, std::string message) {
+            ValidationResult result;
+            result.ok = false;
+            result.phase = std::move(phase);
+            result.message = std::move(message);
+            return result;
+        }
+
+        constexpr float kSampledLuminance = 0.2125f + 0.7154f + 0.0721f;
+
+        float ExpectedInput(InputMode mode, std::uint32_t localIndex) {
+            return mode == InputMode::SampledRgba32f ? kSampledLuminance : static_cast<float>(localIndex + 1u);
+        }
+
+        ValidationResult ValidateProbe(const ProbeOutput& output, InputMode mode) {
+            std::array<std::size_t, kInvocationCount> slotForLocal{};
+            slotForLocal.fill(kNoSlot);
+
+            // 1. Record identity. Slots are only used to locate each reported
+            // local index; all subgroup behavior below groups recorded IDs/lanes.
+            for (std::size_t slot = 0; slot < kInvocationCount; ++slot) {
+                const std::uint32_t localIndex = output.invocation[slot].x;
+                if (localIndex >= kInvocationCount) {
+                    std::ostringstream message;
+                    message << "output slot " << slot << " reports localIndex " << localIndex << " outside [0, 511]";
+                    return Failure("record identity", message.str());
+                }
+                if (slotForLocal[localIndex] != kNoSlot) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " appears in output slots " << slotForLocal[localIndex]
+                            << " and " << slot;
+                    return Failure("record identity", message.str());
+                }
+                slotForLocal[localIndex] = slot;
+            }
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                if (slotForLocal[localIndex] == kNoSlot) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " is missing from all 512 records";
+                    return Failure("record identity", message.str());
+                }
+            }
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                const std::size_t slot = slotForLocal[localIndex];
+                const UVec4& invocation = output.invocation[slot];
+                const std::uint32_t expectedX = static_cast<std::uint32_t>(localIndex % 32u);
+                const std::uint32_t expectedY = static_cast<std::uint32_t>(localIndex / 32u);
+                if (invocation.y != expectedX || invocation.z != expectedY || invocation.w != 0u) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reports local invocation (" << invocation.y << ','
+                            << invocation.z << ',' << invocation.w << "), expected (" << expectedX << ',' << expectedY
+                            << ",0)";
+                    return Failure("record identity", message.str());
+                }
+                const float expectedInput = ExpectedInput(mode, static_cast<std::uint32_t>(localIndex));
+                const float actualInput = output.reduction[slot].x;
+                if (!SameBits(actualInput, expectedInput)) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " input was " << FormatFloat(actualInput) << ", expected "
+                            << FormatFloat(expectedInput);
+                    return Failure("input", message.str());
+                }
+            }
+
+            // 2. Observed topology. Do not derive lanes or subgroup membership
+            // from local invocation indices: only the values the shader recorded
+            // participate in grouping.
+            const std::uint32_t reportedNumSubgroups = output.subgroup[slotForLocal[0]].y;
+            if (reportedNumSubgroups == 0u) {
+                return Failure("observed topology", "localIndex 0 reported gl_NumSubgroups == 0");
+            }
+            if (reportedNumSubgroups > kInvocationCount) {
+                std::ostringstream message;
+                message << "reported gl_NumSubgroups=" << reportedNumSubgroups
+                        << " exceeds the 512 recorded invocations, so at least one subgroup ID is missing";
+                return Failure("observed topology", message.str());
+            }
+            std::vector<std::vector<std::size_t>> subgroupSlots(reportedNumSubgroups);
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                const std::size_t slot = slotForLocal[localIndex];
+                const UVec4& subgroup = output.subgroup[slot];
+                if (subgroup.x == 0u || subgroup.y == 0u) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported subgroupSize=" << subgroup.x
+                            << ", numSubgroups=" << subgroup.y;
+                    return Failure("observed topology", message.str());
+                }
+                if (subgroup.y != reportedNumSubgroups) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported numSubgroups=" << subgroup.y
+                            << ", while localIndex 0 reported " << reportedNumSubgroups;
+                    return Failure("observed topology", message.str());
+                }
+                if (subgroup.z >= reportedNumSubgroups) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported subgroupID=" << subgroup.z
+                            << " outside [0, " << (reportedNumSubgroups - 1u) << ']';
+                    return Failure("observed topology", message.str());
+                }
+                if (subgroup.w >= subgroup.x) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported laneID=" << subgroup.w
+                            << " outside its subgroupSize=" << subgroup.x;
+                    return Failure("observed topology", message.str());
+                }
+                subgroupSlots[subgroup.z].push_back(slot);
+            }
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                if (subgroupSlots[subgroupID].empty()) {
+                    std::ostringstream message;
+                    message << "reported gl_NumSubgroups=" << reportedNumSubgroups
+                            << " but subgroupID " << subgroupID << " has no recorded members";
+                    return Failure("observed topology", message.str());
+                }
+                auto& members = subgroupSlots[subgroupID];
+                std::sort(members.begin(), members.end(), [&output](std::size_t lhs, std::size_t rhs) {
+                    return output.subgroup[lhs].w < output.subgroup[rhs].w;
+                });
+                for (std::size_t i = 1; i < members.size(); ++i) {
+                    if (output.subgroup[members[i - 1]].w == output.subgroup[members[i]].w) {
+                        std::ostringstream message;
+                        message << "subgroupID " << subgroupID << " contains duplicate laneID "
+                                << output.subgroup[members[i]].w;
+                        return Failure("observed topology", message.str());
+                    }
+                }
+            }
+
+            // 3. Native subgroup arithmetic, in the actual lane ordering emitted
+            // by the driver. The fixture values and all partial sums are exactly
+            // representable binary32 values, so compare representation, not epsilon.
+            std::array<float, kInvocationCount> nativePrefix{};
+            std::vector<float> nativeSubtotal(reportedNumSubgroups, 0.0f);
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                float inclusive = 0.0f;
+                for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                    const std::uint32_t localIndex = output.invocation[slot].x;
+                    inclusive += ExpectedInput(mode, localIndex);
+                    nativePrefix[slot] = inclusive;
+                    const float actualNative = output.reduction[slot].y;
+                    if (!SameBits(actualNative, inclusive)) {
+                        std::ostringstream message;
+                        message << "subgroupID " << subgroupID << ", laneID " << output.subgroup[slot].w
+                                << ", localIndex " << localIndex << " nativeInclusive was " << FormatFloat(actualNative)
+                                << ", expected " << FormatFloat(inclusive);
+                        return Failure("native subgroup arithmetic", message.str());
+                    }
+                }
+                nativeSubtotal[subgroupID] = inclusive;
+            }
+
+            // sourceDomain is the narrow source-side safety branch. It is checked
+            // after native arithmetic so an unsupported source topology still
+            // reports native subgroup behavior before failing explicitly.
+            if (reportedNumSubgroups < 2u || reportedNumSubgroups > 32u) {
+                for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                    const std::size_t slot = slotForLocal[localIndex];
+                    const Vec4& reduction = output.reduction[slot];
+                    if (!IsQuietNanSentinel(reduction.z) || !IsQuietNanSentinel(reduction.w) ||
+                        !IsQuietNanSentinel(output.finalAverage[slot])) {
+                        std::ostringstream message;
+                        message << "iterationRP source reduction has no valid contract for gl_NumSubgroups="
+                                << reportedNumSubgroups << "; localIndex " << localIndex
+                                << " did not preserve its qNaN source-reduction sentinel";
+                        return Failure("source domain", message.str());
+                    }
+                    for (std::size_t stage = 0; stage < kScanStageCount; ++stage) {
+                        if (!IsQuietNanSentinel(output.scanAfter[stage][slot])) {
+                            std::ostringstream message;
+                            message << "iterationRP source reduction has no valid contract for gl_NumSubgroups="
+                                    << reportedNumSubgroups << "; localIndex " << localIndex << ", scan stage " << stage
+                                    << " did not preserve its qNaN source-reduction sentinel";
+                            return Failure("source domain", message.str());
+                        }
+                    }
+                }
+                std::ostringstream message;
+                message << "iterationRP source reduction has no valid contract for observed gl_NumSubgroups="
+                        << reportedNumSubgroups << " (requires 2..32); native subgroup results were recorded";
+                return Failure("source domain", message.str());
+            }
+
+            // 4. iterationRP source writer and first shared-memory handoff.
+            std::vector<std::size_t> sourceWriter(reportedNumSubgroups, kNoSlot);
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                std::size_t writerCount = 0;
+                for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                    const UVec4& subgroup = output.subgroup[slot];
+                    if (subgroup.w == subgroup.x - 1u) {
+                        sourceWriter[subgroupID] = slot;
+                        ++writerCount;
+                    }
+                }
+                if (writerCount != 1u) {
+                    std::ostringstream message;
+                    message << "subgroupID " << subgroupID << " has " << writerCount
+                            << " recorded lane(s) where laneID == subgroupSize - 1; iterationRP leaves that "
+                               "shared-cache entry unwritten";
+                    return Failure("source writer", message.str());
+                }
+                for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                    const float actualRawSubtotal = output.reduction[slot].z;
+                    if (!SameBits(actualRawSubtotal, nativeSubtotal[subgroupID])) {
+                        std::ostringstream message;
+                        message << "subgroupID " << subgroupID << ", localIndex " << output.invocation[slot].x
+                                << " sourceRawSubtotal was " << FormatFloat(actualRawSubtotal) << ", expected "
+                                << FormatFloat(nativeSubtotal[subgroupID]);
+                        return Failure("source raw subtotal", message.str());
+                    }
+                }
+            }
+
+            // 5. Reproduce the source loop exactly, including the redundant final
+            // scan iteration on power-of-two subgroup counts. Reads and writes in
+            // one iteration target disjoint cache entries, so update the cache at
+            // the CPU equivalent of the source barrier.
+            std::array<float, kInvocationCount> mergedPrefix = nativePrefix;
+            std::vector<float> cache = nativeSubtotal;
+            std::uint32_t loopLength = std::bit_width(reportedNumSubgroups) - 1u;
+            loopLength +=
+                static_cast<std::uint32_t>(reportedNumSubgroups - (1u << (loopLength - 1u)) > 0u);
+            for (std::uint32_t scanStage = 0u; scanStage < loopLength; ++scanStage) {
+                std::vector<float> cacheAfterStage = cache;
+                for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                    if ((subgroupID & (1u << scanStage)) == 0u) continue;
+                    const std::uint32_t sourceCacheIndex = (subgroupID >> scanStage << scanStage) - 1u;
+                    const float sourcePrefix = cache[sourceCacheIndex];
+                    for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                        mergedPrefix[slot] += sourcePrefix;
+                    }
+                    cacheAfterStage[subgroupID] = mergedPrefix[sourceWriter[subgroupID]];
+                }
+                cache.swap(cacheAfterStage);
+                for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                    const std::size_t slot = slotForLocal[localIndex];
+                    const float actualAfterStage = output.scanAfter[scanStage][slot];
+                    if (!SameBits(actualAfterStage, mergedPrefix[slot])) {
+                        std::ostringstream message;
+                        message << "scanStage " << scanStage << ", subgroupID " << output.subgroup[slot].z
+                                << ", laneID " << output.subgroup[slot].w << ", localIndex " << localIndex
+                                << " scanAfter was " << FormatFloat(actualAfterStage) << ", expected "
+                                << FormatFloat(mergedPrefix[slot]);
+                        ValidationResult result = Failure("source scan", message.str());
+                        result.scanStageMismatch = true;
+                        result.scanStage = static_cast<int>(scanStage);
+                        return result;
+                    }
+                }
+            }
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                const std::size_t slot = slotForLocal[localIndex];
+                const float actualMergedPrefix = output.reduction[slot].w;
+                if (!SameBits(actualMergedPrefix, mergedPrefix[slot])) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " sourceMergedPrefix was "
+                            << FormatFloat(actualMergedPrefix) << ", expected " << FormatFloat(mergedPrefix[slot]);
+                    return Failure("source scan", message.str());
+                }
+            }
+
+            // 6. Final owner and average. The uniformity check is intentionally
+            // separate from the source's topology contract at local index 511.
+            const float firstAverage = output.finalAverage[slotForLocal[0]];
+            for (std::size_t localIndex = 1; localIndex < kInvocationCount; ++localIndex) {
+                const float actualAverage = output.finalAverage[slotForLocal[localIndex]];
+                if (!SameBits(actualAverage, firstAverage)) {
+                    std::ostringstream message;
+                    message << "finalAverage differs: localIndex 0 has " << FormatFloat(firstAverage)
+                            << ", localIndex " << localIndex << " has " << FormatFloat(actualAverage);
+                    return Failure("final average", message.str());
+                }
+            }
+
+            ValidationResult ownerResult;
+            ownerResult.ownerEvaluated = true;
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                if (!subgroupSlots[subgroupID].empty()) {
+                    ownerResult.highestObservedSubgroup = std::max(ownerResult.highestObservedSubgroup, subgroupID);
+                }
+            }
+            const std::size_t index511Slot = slotForLocal[kInvocationCount - 1u];
+            const UVec4& index511Subgroup = output.subgroup[index511Slot];
+            ownerResult.index511IsSourceLastLaneWriter =
+                index511Subgroup.w == index511Subgroup.x - 1u;
+            ownerResult.index511IsHighestSubgroupMember =
+                index511Subgroup.z == ownerResult.highestObservedSubgroup;
+            if (!ownerResult.index511IsSourceLastLaneWriter || !ownerResult.index511IsHighestSubgroupMember) {
+                std::ostringstream message;
+                message << "iterationRP topology incompatibility: localIndex 511 is sourceLastLaneWriter="
+                        << ownerResult.index511IsSourceLastLaneWriter << ", highestSubgroupMember="
+                        << ownerResult.index511IsHighestSubgroupMember << " (subgroupID=" << index511Subgroup.z
+                        << ", highest observed subgroupID=" << ownerResult.highestObservedSubgroup << ')';
+                ownerResult.ok = false;
+                ownerResult.phase = "final average";
+                ownerResult.message = message.str();
+                return ownerResult;
+            }
+
+            float total = 0.0f;
+            for (const float subtotal : nativeSubtotal) total += subtotal;
+            float sampledExpectedTotal = 0.0f;
+            for (std::size_t i = 0; i < kInvocationCount; ++i) sampledExpectedTotal += kSampledLuminance;
+            const float expectedTotal = mode == InputMode::IndexedSsbo ? 131328.0f : sampledExpectedTotal;
+            if (!SameBits(total, expectedTotal) || !SameBits(mergedPrefix[index511Slot], expectedTotal)) {
+                std::ostringstream message;
+                message << "iterationRP source total was " << FormatFloat(mergedPrefix[index511Slot])
+                        << " (native total " << FormatFloat(total) << "), expected " << FormatFloat(expectedTotal);
+                ownerResult.ok = false;
+                ownerResult.phase = "final average";
+                ownerResult.message = message.str();
+                return ownerResult;
+            }
+
+            const float expectedAverage = mode == InputMode::IndexedSsbo ? 256.5f : sampledExpectedTotal / 512.0f;
+            if (!SameBits(firstAverage, expectedAverage)) {
+                std::ostringstream message;
+                message << "finalAverage was " << FormatFloat(firstAverage) << ", expected "
+                        << FormatFloat(expectedAverage);
+                ownerResult.ok = false;
+                ownerResult.phase = "final average";
+                ownerResult.message = message.str();
+                return ownerResult;
+            }
+            return ownerResult;
+        }
+
+        void DumpProbe(const ProbeOutput& output, const CapabilityInfo& capabilities, const ValidationResult& validation,
+                       bool includeScanStages) {
+            PrintMetadata(capabilities, std::cout);
+            if (validation.ok) {
+                std::cout << "IterationRPFirstReductionScenario firstFailure=none\n";
+            } else {
+                std::cout << "IterationRPFirstReductionScenario firstFailure=" << validation.phase << ": "
+                          << validation.message << '\n';
+            }
+            std::cout << "localIndex,localX,localY,localZ,subgroupSize,numSubgroups,subgroupID,laneID,input,"
+                         "nativeInclusive,subgroupSubtotal,mergedPrefix,finalAverage\n";
+            for (std::size_t slot = 0; slot < kInvocationCount; ++slot) {
+                const UVec4& invocation = output.invocation[slot];
+                const UVec4& subgroup = output.subgroup[slot];
+                const Vec4& reduction = output.reduction[slot];
+                std::cout << invocation.x << ',' << invocation.y << ',' << invocation.z << ',' << invocation.w << ','
+                          << subgroup.x << ',' << subgroup.y << ',' << subgroup.z << ',' << subgroup.w << ','
+                          << std::hexfloat << reduction.x << ',' << reduction.y << ',' << reduction.z << ','
+                          << reduction.w << ',' << output.finalAverage[slot] << std::defaultfloat << '\n';
+            }
+            if (includeScanStages) {
+                std::cout << "scanStage,localIndex,scanAfter\n";
+                for (std::size_t scanStage = 0; scanStage < kScanStageCount; ++scanStage) {
+                    for (std::size_t slot = 0; slot < kInvocationCount; ++slot) {
+                        std::cout << scanStage << ',' << output.invocation[slot].x << ',' << std::hexfloat
+                                  << output.scanAfter[scanStage][slot] << std::defaultfloat << '\n';
+                    }
+                }
+            }
+        }
+
+        class IterationRPFirstReductionScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                m_capabilities = QueryCapabilities();
+                // GL_SUBGROUP_SIZE_KHR gates only whether the fixture's source contract
+                // can hold on this device (SubgroupWidthInSourceDomain); it is
+                // deliberately never used to infer lane placement or an expected group
+                // count - those come from observed values alone.
+                PrintMetadata(m_capabilities, std::cout);
+                RecordProperty("iterationrp_gl_subgroup_size_khr", std::to_string(m_capabilities.subgroupSize));
+                if (!m_capabilities.SupportsProbe()) {
+                    GTEST_SKIP() << "subgroup probe requires " << m_capabilities.MissingRequirements();
+                }
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glUseProgram(0);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, 0);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, 0);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                glActiveTexture(GL_TEXTURE3);
+                glBindTexture(GL_TEXTURE_2D, 0);
+                glActiveTexture(GL_TEXTURE0);
+                if (m_texture != 0) glDeleteTextures(1, &m_texture);
+                if (m_inputBuffer != 0) glDeleteBuffers(1, &m_inputBuffer);
+                if (m_outputBuffer != 0) glDeleteBuffers(1, &m_outputBuffer);
+                if (m_program != 0) glDeleteProgram(m_program);
+                m_texture = 0;
+                m_inputBuffer = 0;
+                m_outputBuffer = 0;
+                m_program = 0;
+            }
+
+            GLuint CompileComputeProgram(const std::string& source, std::string* outError) {
+                const char* text = source.c_str();
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                if (shader == 0) {
+                    *outError = "glCreateShader(GL_COMPUTE_SHADER) returned 0";
+                    return 0;
+                }
+                glShaderSource(shader, 1, &text, nullptr);
+                glCompileShader(shader);
+                GLint compiled = GL_FALSE;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[8192] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    *outError = std::string("the subgroup probe compute shader did not compile: ") + log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[8192] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    *outError = std::string("the subgroup probe compute program did not link: ") + log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            bool RunProbe(InputMode mode, ProbeOutput* output, std::string* outError) {
+                m_program = CompileComputeProgram(BuildProbeShader(mode), outError);
+                if (m_program == 0) return false;
+
+                ProbeOutput poison{};
+                std::memset(&poison, 0xa5, sizeof(poison));
+                glGenBuffers(1, &m_outputBuffer);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_outputBuffer);
+                glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(ProbeOutput), &poison, GL_DYNAMIC_COPY);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, m_outputBuffer);
+
+                if (mode == InputMode::IndexedSsbo) {
+                    std::array<float, kInvocationCount> values{};
+                    for (std::size_t i = 0; i < values.size(); ++i) values[i] = static_cast<float>(i + 1u);
+                    glGenBuffers(1, &m_inputBuffer);
+                    glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_inputBuffer);
+                    glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(values), values.data(), GL_STATIC_DRAW);
+                    glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_inputBuffer);
+                } else {
+                    constexpr std::array<float, 4> kOneTexel = {1.0f, 1.0f, 1.0f, 1.0f};
+                    glGenTextures(1, &m_texture);
+                    glActiveTexture(GL_TEXTURE3);
+                    glBindTexture(GL_TEXTURE_2D, m_texture);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+                    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA32F, 1, 1, 0, GL_RGBA, GL_FLOAT, kOneTexel.data());
+                }
+
+                if (const GLenum error = FirstGLError(); error != GL_NO_ERROR) {
+                    std::ostringstream message;
+                    message << "subgroup probe resource setup left " << GLErrorName(error);
+                    *outError = message.str();
+                    return false;
+                }
+
+                glUseProgram(m_program);
+                if (mode == InputMode::SampledRgba32f) {
+                    const GLint sampler = glGetUniformLocation(m_program, "colortex2");
+                    const GLint pixelSize = glGetUniformLocation(m_program, "pixelSize");
+                    if (sampler == -1 || pixelSize == -1) {
+                        *outError = "the sampled probe uniforms were optimized away or not reflected";
+                        return false;
+                    }
+                    glUniform1i(sampler, 3);
+                    glUniform2f(pixelSize, 1.0f / 854.0f, 1.0f / 480.0f);
+                }
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_ALL_BARRIER_BITS);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_outputBuffer);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(ProbeOutput), output);
+                if (const GLenum error = FirstGLError(); error != GL_NO_ERROR) {
+                    std::ostringstream message;
+                    message << "subgroup probe dispatch/readback left " << GLErrorName(error);
+                    *outError = message.str();
+                    return false;
+                }
+                return true;
+            }
+
+            void RunAndValidate(InputMode mode) {
+                ProbeOutput output{};
+                std::string error;
+                ASSERT_TRUE(RunProbe(mode, &output, &error)) << InputModeName(mode) << ": " << error;
+
+                const ValidationResult validation = ValidateProbe(output, mode);
+                if (validation.ownerEvaluated) {
+                    RecordProperty("iterationrp_index511_source_last_lane_writer",
+                                   validation.index511IsSourceLastLaneWriter ? "true" : "false");
+                    RecordProperty("iterationrp_index511_highest_subgroup_member",
+                                   validation.index511IsHighestSubgroupMember ? "true" : "false");
+                    RecordProperty("iterationrp_highest_observed_subgroup",
+                                   std::to_string(validation.highestObservedSubgroup));
+                    std::cout << "IterationRPFirstReductionScenario owner: localIndex511 sourceLastLaneWriter="
+                              << validation.index511IsSourceLastLaneWriter << ", highestSubgroupMember="
+                              << validation.index511IsHighestSubgroupMember << ", highestObservedSubgroup="
+                              << validation.highestObservedSubgroup << '\n';
+                }
+                if (!validation.ok || DumpRequested()) {
+                    DumpProbe(output, m_capabilities, validation, validation.scanStageMismatch || DumpRequested());
+                }
+                EXPECT_TRUE(validation.ok) << validation.phase << ": " << validation.message;
+            }
+
+            CapabilityInfo m_capabilities;
+            GLuint m_program = 0;
+            GLuint m_inputBuffer = 0;
+            GLuint m_outputBuffer = 0;
+            GLuint m_texture = 0;
+        };
+
+    } // namespace
+
+    TEST_F(IterationRPFirstReductionScenario, SampledRgba32fFirstAverage) {
+        if (!Ready() || IsSkipped()) return;
+        RunAndValidate(InputMode::SampledRgba32f);
+    }
+
+    TEST_F(IterationRPFirstReductionScenario, IndexedInputTopologyAndReduction) {
+        if (!Ready() || IsSkipped()) return;
+        RunAndValidate(InputMode::IndexedSsbo);
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/IterationRPProgram203Scenario.cpp

@@ -0,0 +1,379 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/IterationRPProgram203Scenario.cpp
+// Copyright (c) 2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Full iterationRP Program 203 golden input/output fixture. The original shader
+// consumes deterministic complete textures and uniforms, then its complete
+// 512x513 RG16F output image is compared against fixed half-float golden bits.
+// This catches both a wrong exposure slot and collateral scratch corruption.
+
+#include <array>
+#include <bit>
+#include <cmath>
+#include <cstddef>
+#include <cstdint>
+#include <iostream>
+#include <sstream>
+#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 int kSceneWidth = 854;
+        constexpr int kSceneHeight = 480;
+        constexpr int kPixelDataWidth = 512;
+        constexpr int kPixelDataHeight = 513;
+        constexpr std::size_t kSceneTexelCount =
+            static_cast<std::size_t>(kSceneWidth) * kSceneHeight;
+        constexpr std::size_t kPixelDataTexelCount =
+            static_cast<std::size_t>(kPixelDataWidth) * kPixelDataHeight;
+
+        struct Rgba32f {
+            float r, g, b, a;
+        };
+
+        struct Rg16 {
+            std::uint16_t r, g;
+        };
+
+        static_assert(sizeof(Rgba32f) == 16);
+        static_assert(sizeof(Rg16) == 4);
+
+        // Captured from the fixed fixture on Adreno 830. These are the exact
+        // RG16F storage bits for (0.806640625, 8.2578125), not rounded decimal
+        // comparisons performed by the test.
+        constexpr Rg16 kGoldenExposure = {0x3a74u, 0x4821u};
+
+        constexpr const char* kCommonSource = R"glsl(
+#version 430 core
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+uniform int frameCounter;
+uniform float frameTime;
+uniform float aspectRatio;
+uniform vec2 pixelSize;
+uniform float nightVision;
+uniform float darknessLightFactor;
+uniform sampler2D colortex2;
+uniform sampler2D pixelData2D;
+layout(rg16f) uniform image2D img_pixelData2D;
+
+float remapSaturate(float x, float e0, float e1) {
+    return clamp((x - e0) / (e1 - e0), 0.0f, 1.0f);
+}
+
+float GetExposureValue(float luminance) {
+    float aeCurve = 0.65f;
+    aeCurve = mix(aeCurve, clamp(aeCurve * 1.2f, 0.0f, 1.0f), nightVision);
+    aeCurve *= remapSaturate(luminance, 2.0f, 1.0f) * 0.6f + 0.4f;
+    float ae = pow(luminance, -aeCurve);
+    ae *= 1.0f - min(darknessLightFactor * 2.0f, 0.9f);
+    ae *= 8.5f;
+    return ae;
+}
+)glsl";
+
+        constexpr const char* kOriginalMain = R"glsl(
+layout(local_size_x = 32, local_size_y = 16) in;
+shared vec2 prefixSumCache[32];
+
+void main() {
+    vec2 texCoord = (vec2(gl_GlobalInvocationID.xy) + 0.5f) * vec2(1.0f / 32.0f, 1.0f / 16.0f);
+    vec2 sampleCoord = texCoord * (1.0f / 64.0f);
+    sampleCoord.x += (15.0f / 32.0f) + pixelSize.x * 12.0f;
+    float tileExposure = dot(textureLod(colortex2, sampleCoord, 0.0f).rgb,
+                             vec3(0.2125f, 0.7154f, 0.0721f));
+    vec2 sampleLuminance = vec2(tileExposure, 0.0f);
+    sampleLuminance = subgroupInclusiveAdd(sampleLuminance);
+    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+        prefixSumCache[gl_SubgroupID] = sampleLuminance;
+    barrier();
+
+    uint loopLength = uint(findMSB(gl_NumSubgroups));
+    loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
+    for (uint i = 0u; i < loopLength; ++i) {
+        if ((gl_SubgroupID & (1u << i)) > 0u) {
+            sampleLuminance += prefixSumCache[(gl_SubgroupID >> i << i) - 1u];
+            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+                prefixSumCache[gl_SubgroupID] = sampleLuminance;
+        }
+        barrier();
+    }
+    if (gl_LocalInvocationIndex == 511u)
+        prefixSumCache[0] = sampleLuminance / 512.0f;
+    barrier();
+
+    float avg = prefixSumCache[0].x;
+    vec2 tileDistance = texCoord * 2.0f - 1.0f;
+    tileDistance.y /= aspectRatio;
+    float centerDistance = length(tileDistance);
+    float tileWeight = remapSaturate(centerDistance, 0.6f, 0.4f);
+    tileExposure = max(7.0E-7f, tileExposure);
+    float lumaWeight = avg / tileExposure;
+    lumaWeight = pow(lumaWeight, remapSaturate(avg, 0.02f, 0.001f) * 0.4f + 0.2f);
+    tileWeight *= lumaWeight;
+
+    vec2 sampleExposure = vec2(tileExposure * tileWeight, tileWeight);
+    sampleExposure = subgroupInclusiveAdd(sampleExposure);
+    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+        prefixSumCache[gl_SubgroupID] = sampleExposure;
+    barrier();
+    for (uint i = 0u; i < loopLength; ++i) {
+        if ((gl_SubgroupID & (1u << i)) > 0u) {
+            sampleExposure += prefixSumCache[(gl_SubgroupID >> i << i) - 1u];
+            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+                prefixSumCache[gl_SubgroupID] = sampleExposure;
+        }
+        barrier();
+    }
+
+    if (gl_LocalInvocationIndex == 511u) {
+        float avgExposure = max(sampleExposure.x / sampleExposure.y * 29.3f, 1.0E-10f);
+        avgExposure = log2(avgExposure);
+        float prevAvgExposure = log2(texelFetch(pixelData2D, ivec2(0, 0), 0).x);
+        float frameTimeFixed = frameTime + step(frameCounter, 20) * 100.0f;
+        float exposureTime = clamp(frameTimeFixed * 2.0f, 0.0f, 1.0f);
+        avgExposure = mix(prevAvgExposure, avgExposure, exposureTime);
+        avgExposure = max(exp2(avgExposure), 1.0E-5f);
+        float exposure = GetExposureValue(avgExposure);
+        imageStore(img_pixelData2D, ivec2(0, 0), vec4(avgExposure, exposure, 0.0f, 0.0f));
+    }
+}
+)glsl";
+
+        GLuint CompileCompute(const char* mainSource, std::string* error) {
+            const std::array<const GLchar*, 2> sources = {kCommonSource, mainSource};
+            const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+            glShaderSource(shader, static_cast<GLsizei>(sources.size()), sources.data(), nullptr);
+            glCompileShader(shader);
+            GLint compiled = GL_FALSE;
+            glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+            if (compiled != GL_TRUE) {
+                std::array<char, 8192> log{};
+                glGetShaderInfoLog(shader, static_cast<GLsizei>(log.size() - 1), nullptr, log.data());
+                *error = log.data();
+                glDeleteShader(shader);
+                return 0;
+            }
+            const GLuint program = glCreateProgram();
+            glAttachShader(program, shader);
+            glLinkProgram(program);
+            glDeleteShader(shader);
+            GLint linked = GL_FALSE;
+            glGetProgramiv(program, GL_LINK_STATUS, &linked);
+            if (linked != GL_TRUE) {
+                std::array<char, 8192> log{};
+                glGetProgramInfoLog(program, static_cast<GLsizei>(log.size() - 1), nullptr, log.data());
+                *error = log.data();
+                glDeleteProgram(program);
+                return 0;
+            }
+            return program;
+        }
+
+        std::vector<Rgba32f> MakeSceneInput() {
+            std::vector<Rgba32f> texels(kSceneTexelCount);
+            for (int y = 0; y < kSceneHeight; ++y) {
+                for (int x = 0; x < kSceneWidth; ++x) {
+                    std::uint32_t h = static_cast<std::uint32_t>(x) * 0x9e3779b9u;
+                    h ^= static_cast<std::uint32_t>(y) * 0x85ebca6bu;
+                    h ^= h >> 16u;
+                    h *= 0x7feb352du;
+                    h ^= h >> 15u;
+                    const float noise = static_cast<float>(h & 0xffffu) / 65535.0f;
+                    float base = 0.0002f + noise * 0.075f;
+                    const float dx = static_cast<float>(x - 420);
+                    const float dy = static_cast<float>(y - 4);
+                    base += 0.65f * std::exp(-(dx * dx + dy * dy) / 18.0f);
+                    if (((x + y * 17) % 113) == 0) base += 1.75f;
+                    texels[static_cast<std::size_t>(y) * kSceneWidth + x] =
+                        {base * 0.83f, base * 1.07f, base * 1.31f, 1.0f};
+                }
+            }
+            return texels;
+        }
+
+        std::uint16_t FloatToHalf(float value) {
+            const std::uint32_t bits = std::bit_cast<std::uint32_t>(value);
+            const std::uint32_t sign = (bits >> 16u) & 0x8000u;
+            const std::uint32_t exponent = (bits >> 23u) & 0xffu;
+            std::uint32_t mantissa = bits & 0x7fffffu;
+
+            if (exponent == 0xffu) {
+                return static_cast<std::uint16_t>(sign | (mantissa == 0 ? 0x7c00u : 0x7e00u));
+            }
+            int halfExponent = static_cast<int>(exponent) - 127 + 15;
+            if (halfExponent >= 31) return static_cast<std::uint16_t>(sign | 0x7c00u);
+            if (halfExponent <= 0) {
+                if (halfExponent < -10) return static_cast<std::uint16_t>(sign);
+                mantissa |= 0x800000u;
+                const unsigned shift = static_cast<unsigned>(14 - halfExponent);
+                const std::uint32_t rounded = mantissa + ((1u << (shift - 1u)) - 1u) +
+                                              ((mantissa >> shift) & 1u);
+                return static_cast<std::uint16_t>(sign | (rounded >> shift));
+            }
+            mantissa += 0xfffu + ((mantissa >> 13u) & 1u);
+            if ((mantissa & 0x800000u) != 0) {
+                mantissa = 0;
+                if (++halfExponent >= 31) return static_cast<std::uint16_t>(sign | 0x7c00u);
+            }
+            return static_cast<std::uint16_t>(sign | (static_cast<std::uint32_t>(halfExponent) << 10u) |
+                                              (mantissa >> 13u));
+        }
+
+        std::vector<Rg16> MakePixelDataInput() {
+            std::vector<Rg16> texels(kPixelDataTexelCount);
+            for (std::size_t i = 0; i < texels.size(); ++i) {
+                texels[i] = {FloatToHalf(0.35f + static_cast<float>(i % 97u) * 0.0025f),
+                             FloatToHalf(-0.45f + static_cast<float>(i % 89u) * 0.01f)};
+            }
+            texels[0] = {FloatToHalf(0.73f), FloatToHalf(1.25f)};
+            return texels;
+        }
+
+        std::vector<Rg16> MakeGoldenOutput() {
+            std::vector<Rg16> golden = MakePixelDataInput();
+            golden[0] = kGoldenExposure;
+            return golden;
+        }
+
+        GLuint MakeTexture(GLenum internalFormat, GLenum format, GLenum type, int width, int height,
+                           const void* data) {
+            GLuint texture = 0;
+            glGenTextures(1, &texture);
+            glBindTexture(GL_TEXTURE_2D, texture);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+            glTexImage2D(GL_TEXTURE_2D, 0, static_cast<GLint>(internalFormat), width, height, 0, format,
+                         type, data);
+            return texture;
+        }
+
+        void BindAndDispatch(GLuint program, GLuint scene, GLuint pixelData) {
+            glUseProgram(program);
+            glActiveTexture(GL_TEXTURE3);
+            glBindTexture(GL_TEXTURE_2D, scene);
+            glUniform1i(glGetUniformLocation(program, "colortex2"), 3);
+            glActiveTexture(GL_TEXTURE4);
+            glBindTexture(GL_TEXTURE_2D, pixelData);
+            glUniform1i(glGetUniformLocation(program, "pixelData2D"), 4);
+            glBindImageTexture(0, pixelData, 0, GL_FALSE, 0, GL_READ_WRITE, GL_RG16F);
+            glUniform1i(glGetUniformLocation(program, "img_pixelData2D"), 0);
+            glUniform1i(glGetUniformLocation(program, "frameCounter"), 100);
+            glUniform1f(glGetUniformLocation(program, "frameTime"), 1.0f / 60.0f);
+            glUniform1f(glGetUniformLocation(program, "aspectRatio"),
+                        static_cast<float>(kSceneWidth) / kSceneHeight);
+            glUniform2f(glGetUniformLocation(program, "pixelSize"), 1.0f / kSceneWidth, 1.0f / kSceneHeight);
+            glUniform1f(glGetUniformLocation(program, "nightVision"), 0.23f);
+            glUniform1f(glGetUniformLocation(program, "darknessLightFactor"), 0.08f);
+            glDispatchCompute(1, 1, 1);
+            glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
+        }
+
+        std::vector<Rg16> ReadWholeRgTexture(GLuint texture) {
+            std::vector<Rg16> texels(kPixelDataTexelCount);
+            glBindTexture(GL_TEXTURE_2D, texture);
+            glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_HALF_FLOAT, texels.data());
+            return texels;
+        }
+
+        class IterationRPProgram203Scenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                GLint stages = 0;
+                GLint features = 0;
+                GLint invocations = 0;
+                glGetIntegerv(GL_SUBGROUP_SUPPORTED_STAGES_KHR, &stages);
+                glGetIntegerv(GL_SUBGROUP_SUPPORTED_FEATURES_KHR, &features);
+                glGetIntegerv(GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, &invocations);
+                const GLbitfield required =
+                    GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR;
+                if ((static_cast<GLbitfield>(stages) & GL_COMPUTE_SHADER_BIT) == 0 ||
+                    (static_cast<GLbitfield>(features) & required) != required || invocations < 512) {
+                    GTEST_SKIP() << "requires 512-invocation basic+arithmetic compute subgroups";
+                }
+
+                std::string error;
+                m_original = CompileCompute(kOriginalMain, &error);
+                ASSERT_NE(m_original, 0u) << "original Program 203: " << error;
+
+                const std::vector<Rgba32f> scene = MakeSceneInput();
+                const std::vector<Rg16> pixelData = MakePixelDataInput();
+                m_scene = MakeTexture(GL_RGBA16F, GL_RGBA, GL_FLOAT, kSceneWidth, kSceneHeight, scene.data());
+                m_originalOutput =
+                    MakeTexture(GL_RG16F, GL_RG, GL_HALF_FLOAT, kPixelDataWidth, kPixelDataHeight,
+                                pixelData.data());
+                ASSERT_EQ(FirstGLError(), static_cast<GLenum>(GL_NO_ERROR));
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                const std::array<GLuint, 2> textures = {m_scene, m_originalOutput};
+                glDeleteTextures(static_cast<GLsizei>(textures.size()), textures.data());
+                if (m_original != 0) glDeleteProgram(m_original);
+            }
+
+            GLuint m_original = 0;
+            GLuint m_scene = 0;
+            GLuint m_originalOutput = 0;
+        };
+    } // namespace
+
+    TEST_F(IterationRPProgram203Scenario, FixedCompleteInputProducesFixedCompleteGoldenOutput) {
+        if (!Ready()) return;
+
+        BindAndDispatch(m_original, m_scene, m_originalOutput);
+        glFinish();
+        const std::vector<Rg16> actual = ReadWholeRgTexture(m_originalOutput);
+        const std::vector<Rg16> expected = MakeGoldenOutput();
+        ASSERT_EQ(FirstGLError(), static_cast<GLenum>(GL_NO_ERROR));
+
+        std::size_t mismatchTexels = 0;
+        std::size_t firstMismatch = actual.size();
+        for (std::size_t i = 0; i < actual.size(); ++i) {
+            if (actual[i].r != expected[i].r || actual[i].g != expected[i].g) {
+                if (firstMismatch == actual.size()) firstMismatch = i;
+                ++mismatchTexels;
+            }
+        }
+
+        RecordProperty("program203_output_width", kPixelDataWidth);
+        RecordProperty("program203_output_height", kPixelDataHeight);
+        RecordProperty("program203_compared_texels", static_cast<long long>(actual.size()));
+        RecordProperty("program203_mismatch_texels", static_cast<long long>(mismatchTexels));
+        std::cout << "IterationRPProgram203Scenario complete-output actualExposureBits=(0x" << std::hex
+                  << actual[0].r << ", 0x" << actual[0].g << ") goldenExposureBits=(0x" << expected[0].r
+                  << ", 0x" << expected[0].g << std::dec << ") mismatches=" << mismatchTexels << '/'
+                  << actual.size() << '\n';
+
+        if (firstMismatch != actual.size()) {
+            const std::size_t x = firstMismatch % kPixelDataWidth;
+            const std::size_t y = firstMismatch / kPixelDataWidth;
+            ADD_FAILURE() << "complete Program 203 output differs at " << x << ',' << y
+                          << ": actual half bits=(0x" << std::hex << actual[firstMismatch].r << ", 0x"
+                          << actual[firstMismatch].g << ") golden half bits=(0x" << expected[firstMismatch].r
+                          << ", 0x" << expected[firstMismatch].g << std::dec << "); mismatched "
+                          << mismatchTexels << " of " << actual.size() << " texels";
+        }
+        EXPECT_EQ(mismatchTexels, 0u);
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/IterationRPScratchFixScenario.cpp

@@ -0,0 +1,302 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/IterationRPScratchFixScenario.cpp
+// Copyright (c) 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 - THE FIXTURE-SHAPED SUBGROUP REDUCTION, ON WHATEVER WIDTH THE DEVICE HAS.
+//
+// iterationRP hard-sizes the scratch its subgroup prefix scans write through
+// prefixSumCache[gl_SubgroupID], and ships that idiom twice: the auto-exposure pass
+// declares `shared vec2 prefixSumCache[32]` for a 512-invocation workgroup, and the
+// RTW importance warp declares `shared float prefixSumCache[64]` for a 1024-invocation
+// one. Both algorithms are width-agnostic; only the static lengths bake in "at most 32
+// (respectively 64) subgroups", which every desktop capture satisfies and an 8-lane
+// device (lavapipe: 64 and 128 subgroups) does not. DirectVulkan patches exactly that with
+// FixIterationRPSubgroupScratchPass, growing the array to ceil(invocations / native
+// width) on the modules that match the pack's reduction fingerprint.
+//
+// This scenario replays the fixture's reduction shape verbatim - the same 32-entry
+// declaration, the same last-lane handoff, the same findMSB combine loop, and NO
+// domain guard - and asserts only the width-independent result: the workgroup total.
+// The inputs are small integers, so the fp32 sum is exact under any lane order and any
+// association; a correct run produces the exact constant on a 4-lane device and a
+// 128-lane device alike. Without the patch, a sub-16-lane device indexes the
+// 32-entry array out of bounds - on lavapipe that is literal heap corruption - and
+// this scenario is the regression test that keeps the patch working, and it runs on every device that
+// has basic+arithmetic compute subgroups (unlike IterationRPFirstReductionScenario,
+// which probes the UNREPAIRED source contract and must skip outside [16, 256]).
+
+#include <cstdint>
+#include <cstring>
+#include <string>
+
+#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 std::uint32_t kInvocationCount = 512u;
+        // sum of 0..511, exactly representable and associativity-proof in fp32.
+        constexpr float kExpectedTotal = 130816.0f;
+        // The RTW warp's shape: 1024 invocations into a 64-entry float scratch.
+        constexpr std::uint32_t kWideInvocationCount = 1024u;
+        // sum of 0..1023, likewise exact in fp32.
+        constexpr float kWideExpectedTotal = 523776.0f;
+
+        constexpr const char* kComputeSource = R"(#version 430 core
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x = 32, local_size_y = 16, local_size_z = 1) in;
+
+layout(std430, binding = 0) buffer Output {
+    float total;
+    uint numSubgroups;
+    uint maxSubgroupId;
+} outputData;
+
+shared vec2 prefixSumCache[32];
+
+void main() {
+    vec2 sampleLuminance = vec2(float(gl_LocalInvocationIndex), 0.0);
+    sampleLuminance = subgroupInclusiveAdd(sampleLuminance);
+    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+        prefixSumCache[gl_SubgroupID] = sampleLuminance;
+    barrier();
+
+    uint loopLength = uint(findMSB(gl_NumSubgroups));
+    loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
+
+    for (uint scanStage = 0u; scanStage < loopLength; ++scanStage) {
+        if ((gl_SubgroupID & (1u << scanStage)) > 0u) {
+            sampleLuminance += prefixSumCache[(gl_SubgroupID >> scanStage << scanStage) - 1u];
+            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+                prefixSumCache[gl_SubgroupID] = sampleLuminance;
+        }
+        barrier();
+    }
+
+    if (gl_LocalInvocationIndex == 511u) {
+        outputData.total = sampleLuminance.x;
+        outputData.numSubgroups = gl_NumSubgroups;
+    }
+    atomicMax(outputData.maxSubgroupId, gl_SubgroupID);
+}
+)";
+
+        // The RTW importance warp's shape: a plain float scan over 1024 invocations
+        // into a 64-entry scratch. Same idiom, different dimensions - which is exactly
+        // what a fingerprint pinned to the exposure pass's shape walks past.
+        constexpr const char* kWideComputeSource = R"(#version 430 core
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x = 1024) in;
+
+layout(std430, binding = 0) buffer Output {
+    float total;
+    uint numSubgroups;
+    uint maxSubgroupId;
+} outputData;
+
+shared float prefixSumCache[64];
+
+void main() {
+    float importance = float(gl_LocalInvocationID.x);
+    float prefixSum = subgroupInclusiveAdd(importance);
+    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+        prefixSumCache[gl_SubgroupID] = prefixSum;
+    barrier();
+
+    uint loopLength = uint(findMSB(gl_NumSubgroups));
+    loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
+
+    for (uint scanStage = 0u; scanStage < loopLength; ++scanStage) {
+        if ((gl_SubgroupID & (1u << scanStage)) > 0u) {
+            prefixSum += prefixSumCache[(gl_SubgroupID >> scanStage << scanStage) - 1u];
+            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+                prefixSumCache[gl_SubgroupID] = prefixSum;
+        }
+        barrier();
+    }
+
+    if (gl_LocalInvocationID.x == 1023u) {
+        outputData.total = prefixSum;
+        outputData.numSubgroups = gl_NumSubgroups;
+    }
+    atomicMax(outputData.maxSubgroupId, gl_SubgroupID);
+}
+)";
+
+        struct OutputBlock {
+            float total = -1.0f;
+            std::uint32_t numSubgroups = 0;
+            std::uint32_t maxSubgroupId = 0;
+        };
+
+        bool HasExtension(const char* wanted) {
+            GLint extensionCount = 0;
+            glGetIntegerv(GL_NUM_EXTENSIONS, &extensionCount);
+            for (GLint i = 0; i < extensionCount; ++i) {
+                const auto* extension =
+                    reinterpret_cast<const char*>(glGetStringi(GL_EXTENSIONS, static_cast<GLuint>(i)));
+                if (extension != nullptr && std::string(extension) == wanted) return true;
+            }
+            return false;
+        }
+
+        class IterationRPScratchFixScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                GLint stages = 0;
+                GLint features = 0;
+                GLint invocations = 0;
+                const bool subgroupExtension = HasExtension("GL_KHR_shader_subgroup");
+                if (subgroupExtension) {
+                    glGetIntegerv(GL_SUBGROUP_SUPPORTED_STAGES_KHR, &stages);
+                    glGetIntegerv(GL_SUBGROUP_SUPPORTED_FEATURES_KHR, &features);
+                }
+                glGetIntegerv(GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, &invocations);
+                const GLbitfield requiredFeatures =
+                    GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR;
+                if (!subgroupExtension || (static_cast<GLbitfield>(stages) & GL_COMPUTE_SHADER_BIT) == 0 ||
+                    (static_cast<GLbitfield>(features) & requiredFeatures) != requiredFeatures ||
+                    invocations < static_cast<GLint>(kInvocationCount)) {
+                    GTEST_SKIP() << "needs GL_KHR_shader_subgroup basic+arithmetic in compute and a "
+                                    "512-invocation workgroup";
+                }
+
+                m_maxInvocations = static_cast<std::uint32_t>(invocations);
+
+                glGenBuffers(1, &m_output);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_output);
+                // maxSubgroupId starts at zero HOST-side: the word is touched only by
+                // atomicMax during the dispatch, since a plain shader-side zeroing store
+                // would race the other invocations' atomics (barrier() orders shared
+                // memory, not SSBO stores).
+                const OutputBlock poison{-1.0f, 0xa5a5a5a5u, 0u};
+                glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(OutputBlock), &poison, GL_DYNAMIC_READ);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_output);
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, 0);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                if (m_output != 0) glDeleteBuffers(1, &m_output);
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            unsigned int CompileComputeProgram(const char* source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute shader did not compile: ") + log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute program did not link: ") + log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            // Re-poisons the block, compiles the shape under test and runs it once.
+            OutputBlock Dispatch(const char* source) {
+                const OutputBlock poison{-1.0f, 0xa5a5a5a5u, 0u};
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_output);
+                glBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(OutputBlock), &poison);
+                m_program = CompileComputeProgram(source);
+                EXPECT_NE(m_program, 0u) << m_buildLog;
+                if (m_program == 0u) return OutputBlock{};
+                glUseProgram(m_program);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+                OutputBlock block{};
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_output);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(OutputBlock), &block);
+                return block;
+            }
+
+            GLuint m_program = 0;
+            GLuint m_output = 0;
+            std::uint32_t m_maxInvocations = 0;
+            std::string m_buildLog;
+        };
+    } // namespace
+
+    TEST_F(IterationRPScratchFixScenario, FixtureShapedReductionSumsEveryInvocation) {
+        const OutputBlock block = Dispatch(kComputeSource);
+        EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+
+        // The topology diagnostics catch the failure modes by name before the sum does:
+        // an out-of-bounds handoff corrupts the total, a wrong gl_NumSubgroups breaks
+        // the combine loop's length.
+        ASSERT_NE(block.numSubgroups, 0xa5a5a5a5u) << "invocation 511 never reached its store";
+        EXPECT_GE(block.numSubgroups, 1u);
+        EXPECT_LE(block.numSubgroups, kInvocationCount);
+        EXPECT_LT(block.maxSubgroupId, block.numSubgroups)
+            << "gl_SubgroupID exceeds gl_NumSubgroups - the inconsistency "
+               "DeriveNumSubgroupsPass exists to repair";
+
+        // Integer-valued fp32 inputs: the workgroup total is exact under any subgroup
+        // width, lane order, and association. This is the value iterationRP's exposure
+        // average is built from; without FixIterationRPSubgroupScratchPass an 8-lane
+        // device writes prefixSumCache[32..63] out of bounds and this comparison fails.
+        EXPECT_EQ(block.total, kExpectedTotal)
+            << "workgroup reduction produced " << block.total << " with gl_NumSubgroups="
+            << block.numSubgroups;
+    }
+
+    // The pack's second instance of the same bug, and the one that kept the CI
+    // retrace red after the exposure pass alone was patched.
+    TEST_F(IterationRPScratchFixScenario, WideFixtureShapedReductionSumsEveryInvocation) {
+        if (m_maxInvocations < kWideInvocationCount) {
+            GTEST_SKIP() << "needs a " << kWideInvocationCount << "-invocation workgroup";
+        }
+        const OutputBlock block = Dispatch(kWideComputeSource);
+        EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+
+        ASSERT_NE(block.numSubgroups, 0xa5a5a5a5u) << "invocation 1023 never reached its store";
+        EXPECT_GE(block.numSubgroups, 1u);
+        EXPECT_LE(block.numSubgroups, kWideInvocationCount);
+        EXPECT_LT(block.maxSubgroupId, block.numSubgroups)
+            << "gl_SubgroupID exceeds gl_NumSubgroups - the inconsistency "
+               "DeriveNumSubgroupsPass exists to repair";
+
+        // Without the patch an 8-lane device writes prefixSumCache[64..127] out of
+        // bounds and this comparison fails.
+        EXPECT_EQ(block.total, kWideExpectedTotal)
+            << "workgroup reduction produced " << block.total << " with gl_NumSubgroups="
+            << block.numSubgroups;
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/LayeredAttachmentBarrierScenario.cpp

@@ -0,0 +1,378 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/LayeredAttachmentBarrierScenario.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 TRANSFER OFF A NON-ZERO ATTACHMENT LAYER READS THE LAYER THE BARRIER MOVED.
+//
+// Every transfer DirectVulkan performs against a framebuffer attachment is three commands: a
+// barrier that puts the image in TRANSFER_SRC/DST, the copy or blit itself, and a barrier that
+// puts it back. The copy names the attachment's layer - glFramebufferTextureLayer(.., layer) ends
+// up in `srcSubresource.baseArrayLayer` - but TransitionImageLayout used to emit `layerCount = 1`
+// from `baseArrayLayer 0`, so for every attachment on a layer above zero the barrier moved layer 0
+// and the copy read layer N. The layer the transfer touched was never transitioned: it sat in
+// COLOR_ATTACHMENT_OPTIMAL (or DEPTH_STENCIL_ATTACHMENT_OPTIMAL) while being read as TRANSFER_SRC.
+//
+// That is undefined behaviour, not a guaranteed wrong pixel: a layout is a compression/tiling
+// promise, so a driver that stores both layouts identically returns the right bytes anyway. The
+// software lanes (lavapipe) are exactly such a driver, which is why this scenario is paired with a
+// validation-layer run - the layer names the mismatch outright
+// (VUID-vkCmdCopyImageToBuffer-srcImageLayout-00189, "srcImageLayout ... doesn't match the actual
+// current layout") where the pixels here cannot. On a tiler that really does re-tile per layout,
+// these are the reads that come back as garbage.
+//
+// The four cases below are the four transfer paths that take an attachment layer from GL:
+//
+//   glReadPixels (colour)                    -> VulkanRenderer::ReadPixels
+//   glBlitFramebuffer (colour)               -> VulkanRenderer::BlitNamedFramebuffer
+//   glReadPixels (GL_DEPTH_COMPONENT)        -> VulkanRenderer::ReadDepthStencilImageToClient
+//   glBlitFramebuffer (GL_DEPTH_BUFFER_BIT)  -> VulkanRenderer::BlitNamedFramebuffer, depth leg
+//
+// Each one renders or clears INTO the non-zero layer first, so the image is genuinely sitting in
+// its attachment layout when the transfer starts - a scenario that only uploaded texels would
+// leave it in a transfer layout already and the mismatched barrier would be a no-op.
+//
+// Every case also asserts the layers it did not name still hold their own fill, so a backend that
+// "fixed" the miss by transferring the whole image passes neither half.
+//
+// DirectGLES is the control: it hands the same calls to the driver, so a failure on both backends
+// means the scenario is wrong and a failure on DirectVulkan alone means Magma is.
+
+#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 int kWidth = 8;
+        constexpr int kHeight = 8;
+        // Four layers with the subject at index 2: layers on both sides of it stay untouched, so
+        // "moved the whole image" and "moved layer 0" are both distinguishable from correct.
+        constexpr int kLayers = 4;
+        constexpr int kSubjectLayer = 2;
+
+        // A value no correct read can produce, so "the backend wrote nothing" fails loudly.
+        constexpr float kDepthPoison = 0.2f;
+
+        std::string Describe(const Rgba8& color) {
+            return "(" + std::to_string(color.r) + ", " + std::to_string(color.g) + ", " + std::to_string(color.b) +
+                   ", " + std::to_string(color.a) + ")";
+        }
+
+        // Per-layer fill, uniform within a layer: the defect is about WHICH layer is addressed, and
+        // a value that also varied inside the layer would make the assertions depend on row order.
+        Rgba8 LayerFill(int layer) {
+            return {static_cast<GLubyte>(17 + layer * 30), static_cast<GLubyte>(200 - layer * 25),
+                    static_cast<GLubyte>(60 + layer * 40), 255};
+        }
+
+        // What the draw paints - matches kFS below, and is deliberately none of the LayerFill
+        // values so "the draw never landed" cannot read as a pass.
+        constexpr Rgba8 kPaintedColor{26, 51, 204, 255};
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+void main() { gl_Position = vec4(aPos, 0.0, 1.0); }
+)";
+
+        constexpr const char* kFS = R"(#version 330 core
+out vec4 o_color;
+void main() { o_color = vec4(0.1, 0.2, 0.8, 1.0); }
+)";
+
+        void DrawFullViewportQuad(unsigned int program) {
+            static const float kQuad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+            GLuint vao = 0, vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, sizeof(kQuad), kQuad, GL_STATIC_DRAW);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            glBindVertexArray(0);
+            glDeleteBuffers(1, &vbo);
+            glDeleteVertexArrays(1, &vao);
+        }
+
+        class LayeredAttachmentBarrierScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string error;
+                m_program = CompileProgram(kVS, kFS, &error);
+                ASSERT_NE(m_program, 0u) << error;
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                for (const GLuint fbo : m_fbos) {
+                    glDeleteFramebuffers(1, &fbo);
+                }
+                m_fbos.clear();
+                for (const GLuint texture : m_textures) {
+                    glDeleteTextures(1, &texture);
+                }
+                m_textures.clear();
+                if (m_program != 0) {
+                    glUseProgram(0);
+                    glDeleteProgram(m_program);
+                    m_program = 0;
+                }
+            }
+
+            // An RGBA8 2D array with a different uniform colour per layer.
+            GLuint MakeColorArray() {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_RGBA8, kWidth, kHeight, kLayers);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                for (int layer = 0; layer < kLayers; ++layer) {
+                    const std::vector<Rgba8> texels(static_cast<std::size_t>(kWidth) * kHeight, LayerFill(layer));
+                    glTexSubImage3D(GL_TEXTURE_2D_ARRAY, 0, 0, 0, layer, kWidth, kHeight, 1, GL_RGBA,
+                                    GL_UNSIGNED_BYTE, texels.data());
+                }
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                return texture;
+            }
+
+            // A depth 2D array. No initial upload: depth arrays are filled by clearing through an
+            // attachment, which is also the state the transfer paths have to cope with.
+            GLuint MakeDepthArray() {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_DEPTH_COMPONENT24, kWidth, kHeight, kLayers);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                return texture;
+            }
+
+            // One FBO naming `layer` of the given arrays. Depth is optional (0 = colour only).
+            GLuint MakeLayerFbo(GLuint colorArray, GLuint depthArray, int layer) {
+                GLuint fbo = 0;
+                glGenFramebuffers(1, &fbo);
+                m_fbos.push_back(fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+                glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, colorArray, 0, layer);
+                if (depthArray != 0) {
+                    glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depthArray, 0, layer);
+                }
+                EXPECT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                    << "layer " << layer << " is not attachable";
+                return fbo;
+            }
+
+            // glReadPixels of one whole layer, through an FBO that names it.
+            Rgba8 ReadLayer(GLuint colorArray, int layer) {
+                const GLuint fbo = MakeLayerFbo(colorArray, 0, layer);
+                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+                glReadBuffer(GL_COLOR_ATTACHMENT0);
+                glPixelStorei(GL_PACK_ALIGNMENT, 1);
+                std::vector<Rgba8> pixels(static_cast<std::size_t>(kWidth) * kHeight, Rgba8{});
+                glReadPixels(0, 0, kWidth, kHeight, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                // The fill is uniform within a layer, so any disagreement between texels is itself
+                // a failure - reported here rather than silently reduced to pixels[0].
+                for (std::size_t i = 1; i < pixels.size(); ++i) {
+                    EXPECT_TRUE(pixels[i] == pixels[0])
+                        << "layer " << layer << " is not uniform: texel 0 is " << Describe(pixels[0]) << ", texel "
+                        << i << " is " << Describe(pixels[i]);
+                }
+                return pixels[0];
+            }
+
+            // Every layer but `changed` still holds its own fill.
+            void ExpectOtherLayersUntouched(GLuint colorArray, int changed, const char* what) {
+                for (int layer = 0; layer < kLayers; ++layer) {
+                    if (layer == changed) continue;
+                    const Rgba8 actual = ReadLayer(colorArray, layer);
+                    EXPECT_TRUE(actual == LayerFill(layer))
+                        << what << ": layer " << layer << " should still hold its fill but is " << Describe(actual)
+                        << ", expected " << Describe(LayerFill(layer));
+                }
+            }
+
+            float ReadDepthAt(int x, int y) const {
+                float depth = kDepthPoison;
+                glReadPixels(x, y, 1, 1, GL_DEPTH_COMPONENT, GL_FLOAT, &depth);
+                return depth;
+            }
+
+            std::vector<GLuint> m_textures;
+            std::vector<GLuint> m_fbos;
+            unsigned int m_program = 0;
+        };
+
+        // glReadPixels straight off a layer that was just rendered to. The image is in
+        // COLOR_ATTACHMENT_OPTIMAL when the readback barrier runs, so the barrier and the copy
+        // disagreeing about the layer is a live layout mismatch, not a bookkeeping detail.
+        TEST_F(LayeredAttachmentBarrierScenario, ReadPixelsOffRenderedNonZeroLayer) {
+            if (!Ready()) return;
+
+            const GLuint colorArray = MakeColorArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            const GLuint fbo = MakeLayerFbo(colorArray, 0, kSubjectLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDisable(GL_DEPTH_TEST);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            DrawFullViewportQuad(m_program);
+
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glPixelStorei(GL_PACK_ALIGNMENT, 1);
+            std::vector<Rgba8> pixels(static_cast<std::size_t>(kWidth) * kHeight, Rgba8{});
+            glReadPixels(0, 0, kWidth, kHeight, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            for (std::size_t i = 0; i < pixels.size(); ++i) {
+                ASSERT_NEAR(pixels[i].r, kPaintedColor.r, 2)
+                    << "texel " << i << " of the rendered layer is " << Describe(pixels[i]);
+                ASSERT_NEAR(pixels[i].g, kPaintedColor.g, 2) << "texel " << i;
+                ASSERT_NEAR(pixels[i].b, kPaintedColor.b, 2) << "texel " << i;
+            }
+
+            ExpectOtherLayersUntouched(colorArray, kSubjectLayer, "readback off a rendered layer");
+        }
+
+        // glBlitFramebuffer between two non-zero layers of two different arrays. Both endpoints are
+        // above layer 0, so the source and destination barriers are each wrong on their own side.
+        TEST_F(LayeredAttachmentBarrierScenario, BlitBetweenNonZeroColorLayers) {
+            if (!Ready()) return;
+
+            const GLuint sourceArray = MakeColorArray();
+            const GLuint destinationArray = MakeColorArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            constexpr int kSourceLayer = 3;
+            constexpr int kDestinationLayer = 1;
+
+            const GLuint sourceFbo = MakeLayerFbo(sourceArray, 0, kSourceLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, sourceFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDisable(GL_DEPTH_TEST);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            DrawFullViewportQuad(m_program);
+
+            const GLuint destinationFbo = MakeLayerFbo(destinationArray, 0, kDestinationLayer);
+            glBindFramebuffer(GL_READ_FRAMEBUFFER, sourceFbo);
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, destinationFbo);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            const Rgba8 blitted = ReadLayer(destinationArray, kDestinationLayer);
+            EXPECT_NEAR(blitted.r, kPaintedColor.r, 2) << "blit destination layer is " << Describe(blitted);
+            EXPECT_NEAR(blitted.g, kPaintedColor.g, 2);
+            EXPECT_NEAR(blitted.b, kPaintedColor.b, 2);
+
+            ExpectOtherLayersUntouched(destinationArray, kDestinationLayer, "colour blit destination");
+            // The source layer was rendered, not blitted into, so it is checked separately.
+            const Rgba8 source = ReadLayer(sourceArray, kSourceLayer);
+            EXPECT_NEAR(source.r, kPaintedColor.r, 2) << "blit source layer is " << Describe(source);
+            ExpectOtherLayersUntouched(sourceArray, kSourceLayer, "colour blit source");
+        }
+
+        // The depth aspect of the same readback path: the depth image sits in
+        // DEPTH_STENCIL_ATTACHMENT_OPTIMAL after the clear, and the copy names the attached layer.
+        TEST_F(LayeredAttachmentBarrierScenario, ReadDepthOffClearedNonZeroLayer) {
+            if (!Ready()) return;
+
+            const GLuint colorArray = MakeColorArray();
+            const GLuint depthArray = MakeDepthArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            const GLuint fbo = MakeLayerFbo(colorArray, depthArray, kSubjectLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.375);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            const float centre = ReadDepthAt(kWidth / 2, kHeight / 2);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+            EXPECT_NEAR(centre, 0.375f, 1.0f / 4096.0f)
+                << "glReadPixels(GL_DEPTH_COMPONENT) off layer " << kSubjectLayer << " returned " << centre
+                << (std::fabs(centre - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+        }
+
+        // The depth leg of the blit path, both endpoints above layer 0. Verified by reading the
+        // destination's depth back, which is the same readback the case above pins - so a failure
+        // here with that one passing is the blit, not the readback.
+        TEST_F(LayeredAttachmentBarrierScenario, BlitDepthBetweenNonZeroLayers) {
+            if (!Ready()) return;
+
+            const GLuint sourceColor = MakeColorArray();
+            const GLuint sourceDepth = MakeDepthArray();
+            const GLuint destinationColor = MakeColorArray();
+            const GLuint destinationDepth = MakeDepthArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            constexpr int kSourceLayer = 3;
+            constexpr int kDestinationLayer = 1;
+
+            const GLuint sourceFbo = MakeLayerFbo(sourceColor, sourceDepth, kSourceLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, sourceFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.625);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            // A destination pre-cleared to something the blit must overwrite, so "the blit did
+            // nothing" and "the blit landed" are different answers.
+            const GLuint destinationFbo = MakeLayerFbo(destinationColor, destinationDepth, kDestinationLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, destinationFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.125);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            glBindFramebuffer(GL_READ_FRAMEBUFFER, sourceFbo);
+            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, destinationFbo);
+            glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            glBindFramebuffer(GL_FRAMEBUFFER, destinationFbo);
+            const float blitted = ReadDepthAt(kWidth / 2, kHeight / 2);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+            EXPECT_NEAR(blitted, 0.625f, 1.0f / 4096.0f)
+                << "depth blitted onto layer " << kDestinationLayer << " reads back as " << blitted
+                << (std::fabs(blitted - 0.125f) < 1e-3f ? " - the destination kept its own clear" : "");
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/VertexArrayEnableDisableScenario.cpp

@@ -0,0 +1,264 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/VertexArrayEnableDisableScenario.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
+//
+// KHR-GL45.direct_state_access.vertex_arrays_enable_disable_attributes, rebuilt.
+//
+// The case is small and does one unusual thing twice: it turns half of
+// GL_MAX_VERTEX_ATTRIBS attribute arrays on and the other half off with
+// glEnableVertexArrayAttrib / glDisableVertexArrayAttrib on a vertex array object
+// that is NOT bound (it binds the default one first, on purpose), draws one point
+// through a program that reads exactly the enabled half, and checks the sum those
+// arrays produced. Then it swaps which half is enabled, draws again through a
+// SECOND program, and checks the other sum.
+//
+// Both draws capture into ONE four-byte transform feedback buffer, allocated once
+// with immutable storage and read back with glMapBuffer - so anything that only
+// works on the first capture span through a buffer fails the second check while
+// leaving the first one green.
+//
+// It is reassembled here rather than shortened because every one of those details
+// is a candidate: the unbound-VAO enables, the two-program swap, the integer
+// attributes fetched with glVertexAttribIPointer at a stride wider than one
+// element, the second capture span, and the fact that the sums differ ONLY in
+// which arrays contributed (a fetch that ignored the enable state, or one that
+// read the wrong element, lands on a different number, not on garbage).
+
+#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 {
+
+        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;
+        }
+
+        // Declares and sums the even (parity 0) or odd (parity 1) attributes only, with the
+        // locations assigned by glBindAttribLocation rather than a layout qualifier - which is
+        // what the CTS case does, and which makes the attribute set the program reads a link
+        // property rather than a source one.
+        GLuint BuildSumProgram(int parity, int attributeCount, std::string* log) {
+            std::string declarations;
+            std::string copies = "  sum = 0;\n";
+            for (int i = parity; i < attributeCount; i += 2) {
+                declarations += "in int a_" + std::to_string(i) + ";\n";
+                copies += "  sum += a_" + std::to_string(i) + ";\n";
+            }
+            // `flat` where the CTS case has none: an integral shader output cannot be
+            // interpolated, so a driver is within its rights to reject the unqualified form
+            // even with no matching fragment input. The capture reads the same value either
+            // way, and the qualifier keeps this scenario portable off llvmpipe.
+            const std::string vertexSource = "#version 450\n\n" + declarations +
+                                             "flat out int sum;\n\nvoid main()\n{\n" + copies + "}\n";
+            const std::string fragmentSource = R"(#version 450
+
+out vec4 color;
+
+void main()
+{
+    color = vec4(1.0);
+}
+)";
+            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);
+            const char* varying = "sum";
+            glTransformFeedbackVaryings(program, 1, &varying, GL_INTERLEAVED_ATTRIBS);
+            for (int i = parity; i < attributeCount; i += 2) {
+                const std::string name = "a_" + std::to_string(i);
+                glBindAttribLocation(program, static_cast<GLuint>(i), name.c_str());
+            }
+            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;
+        }
+
+        class VertexArrayEnableDisableScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                glGetIntegerv(GL_MAX_VERTEX_ATTRIBS, &m_attributeCount);
+                ASSERT_GE(m_attributeCount, 16);
+
+                std::string log;
+                m_even = BuildSumProgram(0, m_attributeCount, &log);
+                ASSERT_NE(m_even, 0u) << "even program failed to build: " << log;
+                m_odd = BuildSumProgram(1, m_attributeCount, &log);
+                ASSERT_NE(m_odd, 0u) << "odd program failed to build: " << log;
+
+                // One element per attribute, read as one vertex whose stride spans them all.
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                std::vector<GLint> reference(static_cast<std::size_t>(m_attributeCount));
+                for (int i = 0; i < m_attributeCount; ++i) reference[static_cast<std::size_t>(i)] = i;
+                glBufferData(GL_ARRAY_BUFFER, static_cast<GLsizeiptr>(reference.size() * sizeof(GLint)),
+                             reference.data(), GL_STATIC_DRAW);
+                for (int i = 0; i < m_attributeCount; ++i) {
+                    glVertexAttribIPointer(static_cast<GLuint>(i), 1, GL_INT,
+                                           static_cast<GLsizei>(sizeof(GLint) * m_attributeCount),
+                                           reinterpret_cast<const void*>(static_cast<std::size_t>(i) * sizeof(GLint)));
+                }
+                glBindBuffer(GL_ARRAY_BUFFER, 0);
+
+                // Immutable storage, allocated once, read back with glMapBuffer - the capture
+                // buffer is never respecified between the two spans.
+                glGenBuffers(1, &m_xfb);
+                glBindBuffer(GL_TRANSFORM_FEEDBACK_BUFFER, m_xfb);
+                glBufferStorage(GL_TRANSFORM_FEEDBACK_BUFFER, sizeof(GLint), nullptr, GL_MAP_READ_BIT);
+                glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, m_xfb);
+                ASSERT_EQ(glGetError(), GL_NO_ERROR) << "capture buffer setup";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glUseProgram(0);
+                glBindVertexArray(0);
+                if (m_xfb != 0) glDeleteBuffers(1, &m_xfb);
+                if (m_vbo != 0) glDeleteBuffers(1, &m_vbo);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_even != 0) glDeleteProgram(m_even);
+                if (m_odd != 0) glDeleteProgram(m_odd);
+                ScenarioTest::TearDown();
+            }
+
+            // Enables one parity's arrays and disables the other's, THROUGH THE OBJECT NAME
+            // while a different vertex array object is bound.
+            void TurnOnAttributes(int enabledParity) {
+                glBindVertexArray(0);
+                for (int i = 0; i < m_attributeCount; ++i) {
+                    if (i % 2 == enabledParity % 2) {
+                        glEnableVertexArrayAttrib(m_vao, static_cast<GLuint>(i));
+                    } else {
+                        glDisableVertexArrayAttrib(m_vao, static_cast<GLuint>(i));
+                    }
+                    ASSERT_EQ(glGetError(), GL_NO_ERROR) << "attribute " << i << ", parity " << enabledParity;
+                }
+                glBindVertexArray(m_vao);
+            }
+
+            int ExpectedSum(int parity) const {
+                int sum = 0;
+                for (int i = parity; i < m_attributeCount; i += 2) sum += i;
+                return sum;
+            }
+
+            // One capture span, read back the way the CTS case does.
+            int DrawAndRead(int parity) {
+                glUseProgram(parity == 0 ? m_even : m_odd);
+                glBindVertexArray(m_vao);
+                glBeginTransformFeedback(GL_POINTS);
+                glDrawArrays(GL_POINTS, 0, 1);
+                glEndTransformFeedback();
+
+                const void* mapped = glMapBuffer(GL_TRANSFORM_FEEDBACK_BUFFER, GL_READ_ONLY);
+                if (mapped == nullptr) {
+                    ADD_FAILURE() << "glMapBuffer returned null for parity " << parity;
+                    return -1;
+                }
+                GLint result = -1;
+                std::memcpy(&result, mapped, sizeof(result));
+                glUnmapBuffer(GL_TRANSFORM_FEEDBACK_BUFFER);
+                return result;
+            }
+
+            GLint m_attributeCount = 16;
+            GLuint m_even = 0;
+            GLuint m_odd = 0;
+            GLuint m_vao = 0;
+            GLuint m_vbo = 0;
+            GLuint m_xfb = 0;
+        };
+
+        // The case verbatim: even half on, draw, check; odd half on, draw, check.
+        TEST_F(VertexArrayEnableDisableScenario, EitherHalfOfTheAttributesInTurn) {
+            if (!Ready()) GTEST_SKIP();
+
+            TurnOnAttributes(0);
+            EXPECT_EQ(DrawAndRead(0), ExpectedSum(0)) << "even attributes";
+
+            TurnOnAttributes(1);
+            EXPECT_EQ(DrawAndRead(1), ExpectedSum(1)) << "odd attributes";
+
+            EXPECT_EQ(glGetError(), GL_NO_ERROR);
+        }
+
+        // The first span on its own, so a failure of the case above can be read as "the second
+        // span" rather than "the enables".
+        TEST_F(VertexArrayEnableDisableScenario, TheEvenHalfAlone) {
+            if (!Ready()) GTEST_SKIP();
+
+            TurnOnAttributes(0);
+            EXPECT_EQ(DrawAndRead(0), ExpectedSum(0));
+            EXPECT_EQ(glGetError(), GL_NO_ERROR);
+        }
+
+        // And the odd half as the FIRST span, which separates "the odd program/arrays are
+        // wrong" from "the second span is wrong".
+        TEST_F(VertexArrayEnableDisableScenario, TheOddHalfAlone) {
+            if (!Ready()) GTEST_SKIP();
+
+            TurnOnAttributes(1);
+            EXPECT_EQ(DrawAndRead(1), ExpectedSum(1));
+            EXPECT_EQ(glGetError(), GL_NO_ERROR);
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/VertexAttribBindingScenario.cpp

@@ -139,16 +139,18 @@ void main() {
         // As CapturePoints, but through the baseInstance entry point, and on a capture buffer
         // of its own.
         //
-        // Kept separate from CapturePoints rather than defaulting a parameter, for two
+        // Kept separate from CapturePoints rather than defaulting a parameter, so that every
-        // reasons. Every existing caller stays on the draw command that carries no
+        // existing caller stays on the draw command that carries no baseInstance at all: the
-        // baseInstance at all, so the negative control is a DIFFERENT command rather than
+        // negative control is then a DIFFERENT command rather than the same one passed a zero.
-        // the same one passed a zero. And baseInstance is the first thing here that needs
+        //
-        // several captures in ONE test, which the shared helper cannot currently do: a
+        // The buffer per capture is a leftover. baseInstance was the first thing here that
-        // second capture into the same buffer object comes back empty on DirectVulkan
+        // needed several captures in ONE test, and at the time a second capture into the same
-        // (respecifying a buffer that is bound to a transform-feedback binding point does
+        // buffer object came back empty on DirectVulkan - respecifying a buffer whose bytes the
-        // not reach that binding - reproduced with two plain CapturePoints calls, so it is
+        // backend had handed the frontend a pointer into replaced the storage under that
-        // neither about baseInstance nor about this helper). A fresh buffer per capture
+        // pointer, so the capture wrote one store and the readback read another. That is fixed
-        // sidesteps it; without that, this scenario would be pinning that bug instead.
+        // and pinned by XfbCaptureBufferReuseScenario, which owns the shape now; a buffer per
+        // capture is simply the cheapest thing that still isolates these three draws from each
+        // other.
         std::vector<float> CaptureOwnBufferBaseInstance(GLuint program, int vertexCount, int instanceCount,
                                                         GLuint baseInstance, bool useBaseInstanceCommand) {
             const std::size_t floats = static_cast<std::size_t>(vertexCount) * instanceCount * 16;

MobileGL/MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp

@@ -0,0 +1,524 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ViewportArrayScenario.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 - gl_ViewportIndex ACTUALLY ROUTES, AND THE PER-INDEX STATE IT SELECTS IS REAL.
+//
+// The state half of ARB_viewport_array is asserted in MG_Test/State/RenderStateTest.cpp, which
+// is a pure set/get exercise and would pass just as green against a backend that stores all 16
+// rectangles and rasterizes only the first. This file is the other half: every case here routes
+// primitives to a viewport OTHER than 0 and then looks at where the pixels landed.
+//
+// Three claims, one per case:
+//   1. gl_ViewportIndex selects the viewport RECTANGLE - a 4x4 grid of 32x32 viewports, one
+//      geometry-shader invocation per cell, and every cell must hold its own index.
+//   2. gl_ViewportIndex selects the DEPTH RANGE - 16 one-pixel-wide viewports whose ranges are
+//      (i/16, 1 - i/16), a quad at each end of clip space, and gl_FragCoord.z read back.
+//      This is the claim that fails loudest against a single-viewport backend, because the
+//      geometry is still in the right place while every depth comes back as viewport 0's.
+//   3. The per-index SCISSOR TEST ENABLE is honoured. Vulkan has no per-viewport scissor-test
+//      toggle, so a disabled index has to be given the whole framebuffer as its rectangle; the
+//      case draws the same primitive into the same index twice, once with the test off and once
+//      with it on, and requires the two results to differ in the documented direction.
+//
+// Case 1 runs a second time against the DEFAULT framebuffer. MobileGL Y-flips (and pre-transform
+// rotates) the default framebuffer's rectangles and does not touch an FBO's, so a port that
+// applies the flip to viewport 0 and forgets the other fifteen renders a correct-looking FBO and
+// an upside-down window - the classic multi-viewport bug, and invisible to every FBO-only case.
+//
+// HONEST LIMIT OF THIS FILE. DirectGLES SKIPS every case: GLES has one viewport, one scissor
+// rectangle and no gl_ViewportIndex, so routing to index > 0 is an emulation feature that has
+// not been built (the Espryt half of KHR-GL43.viewport_array's rendering group is deliberately
+// still red). The skip is explicit rather than silent so a future emulation lands here as a
+// failing test and not as a test that was quietly never running. DirectVulkan additionally
+// skips when the device lacks the multiViewport feature - Vulkan then forbids a pipeline from
+// declaring more than one viewport at all, which is a device limit and not a MobileGL bug;
+// lavapipe (every CI lane) and both Mali/Adreno devices support it, so the cases do run where
+// it matters.
+
+#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 int kViewportCount = 16;
+        constexpr int kGridSide = 4;   // 4x4 grid of viewports
+        constexpr int kCellSize = 32;  // ... each 32x32
+        constexpr int kSurfaceSide = kGridSide * kCellSize;
+        constexpr GLint kUnwritten = -1;
+
+        // A geometry shader is the only stage GL 4.1 lets write gl_ViewportIndex, and
+        // `invocations` runs it once per viewport off a single input point - the same shape
+        // KHR-GL43.viewport_array.draw_to_single_layer_with_multiple_viewports uses.
+        const char* const kVertexSource = R"(#version 410 core
+void main() { gl_Position = vec4(0.0, 0.0, 0.0, 1.0); }
+)";
+
+        const char* const kGridGeometrySource = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 4) out;
+flat out int gsIndex;
+void main() {
+    gsIndex = gl_InvocationID;
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        // One invocation, viewport chosen by a uniform: lets a case draw the SAME primitive into
+        // the SAME index twice under two different scissor-enable states.
+        const char* const kSingleGeometrySource = R"(#version 410 core
+layout(points, invocations = 1) in;
+layout(triangle_strip, max_vertices = 4) out;
+uniform int uViewport;
+flat out int gsIndex;
+void main() {
+    gsIndex = uViewport;
+    gl_ViewportIndex = uViewport;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        const char* const kIntFragmentSource = R"(#version 410 core
+flat in int gsIndex;
+layout(location = 0) out int fragColor;
+void main() { fragColor = gsIndex; }
+)";
+
+        // Two quads, one at each end of clip space, so the fragment stage can report the depth
+        // the viewport's range mapped them to. gl_FragCoord.z IS the post-range window depth, so
+        // it reads back the per-viewport minDepth/maxDepth directly.
+        const char* const kDepthGeometrySource = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 8) out;
+void main() {
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  0.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  0.0, -1.0, 1.0); EmitVertex();
+    EndPrimitive();
+    gl_Position = vec4(-1.0,  0.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  0.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 1.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        const char* const kDepthFragmentSource = R"(#version 410 core
+layout(location = 0) out float fragColor;
+void main() { fragColor = gl_FragCoord.z; }
+)";
+
+        class ViewportArrayScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                if (Gl().BackendName() == "DirectGLES") {
+                    GTEST_SKIP() << "gl_ViewportIndex routing is not emulated on DirectGLES: GLES has one viewport "
+                                    "and one scissor rectangle, so every index rasterizes as index 0. The indexed "
+                                    "STATE is still asserted (MG_Test RenderStateTest); this is the deferred "
+                                    "rendering half of KHR-GL43.viewport_array.";
+                }
+
+                GLint maxViewports = 0;
+                glGetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+                ASSERT_GE(maxViewports, kViewportCount) << "GL 4.3 core requires GL_MAX_VIEWPORTS >= 16";
+
+                m_program = BuildProgram(kGridGeometrySource, kIntFragmentSource);
+                ASSERT_NE(m_program, 0u) << "grid program failed to build: " << m_buildLog;
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                ResetViewportArrayState();
+                ASSERT_EQ(glGetError(), GL_NO_ERROR) << "setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready() || IsSkipped()) return;
+                ResetViewportArrayState();
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_program != 0) glDeleteProgram(m_program);
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+            }
+
+            // Every case starts from the same slate: this fixture shares its context with every
+            // other scenario in the process, and a leftover per-index scissor enable is exactly
+            // the kind of state that would make a later case pass or fail for the wrong reason.
+            static void ResetViewportArrayState() {
+                for (int i = 0; i < kViewportCount; ++i) {
+                    glDisablei(GL_SCISSOR_TEST, static_cast<GLuint>(i));
+                }
+                glDisable(GL_SCISSOR_TEST);
+                glViewport(0, 0, kSurfaceSide, kSurfaceSide);
+                glScissor(0, 0, kSurfaceSide, kSurfaceSide);
+                glDepthRange(0.0, 1.0);
+                glDisable(GL_DEPTH_TEST);
+            }
+
+            // The 4x4 grid: viewport y*4+x covers the cell whose lower-left corner is
+            // (x*cellW, y*cellH), in GL's bottom-left-origin window coordinates. Parameterized on
+            // the cell size because the default framebuffer this scenario also renders into is
+            // deliberately non-square (HeadlessGL is 128x96, so a transposing bug cannot hide).
+            static void SetupGridViewports(int cellW, int cellH) {
+                std::vector<GLfloat> data(static_cast<size_t>(kViewportCount) * 4);
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        const size_t base = static_cast<size_t>(y * kGridSide + x) * 4;
+                        data[base + 0] = static_cast<GLfloat>(x * cellW);
+                        data[base + 1] = static_cast<GLfloat>(y * cellH);
+                        data[base + 2] = static_cast<GLfloat>(cellW);
+                        data[base + 3] = static_cast<GLfloat>(cellH);
+                    }
+                }
+                glViewportArrayv(0, kViewportCount, data.data());
+            }
+
+            GLuint BuildProgram(const char* geometrySource, const char* fragmentSource) {
+                const GLuint vs = CompileStage(GL_VERTEX_SHADER, kVertexSource);
+                if (vs == 0) return 0;
+                const GLuint gs = CompileStage(GL_GEOMETRY_SHADER, geometrySource);
+                if (gs == 0) {
+                    glDeleteShader(vs);
+                    return 0;
+                }
+                const GLuint fs = CompileStage(GL_FRAGMENT_SHADER, fragmentSource);
+                if (fs == 0) {
+                    glDeleteShader(vs);
+                    glDeleteShader(gs);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, vs);
+                glAttachShader(program, gs);
+                glAttachShader(program, fs);
+                glLinkProgram(program);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                glDeleteShader(vs);
+                glDeleteShader(gs);
+                glDeleteShader(fs);
+                if (!linked) {
+                    GLint length = 0;
+                    glGetProgramiv(program, GL_INFO_LOG_LENGTH, &length);
+                    std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                    glGetProgramInfoLog(program, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                    m_buildLog = log.data();
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            GLuint CompileStage(GLenum stage, const char* source) {
+                const GLuint shader = glCreateShader(stage);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled) return shader;
+                GLint length = 0;
+                glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+                std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                glGetShaderInfoLog(shader, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                m_buildLog = log.data();
+                glDeleteShader(shader);
+                return 0;
+            }
+
+            // An R32I colour target, pre-filled with kUnwritten so "nothing was drawn here" is
+            // distinguishable from "index 0 was drawn here".
+            struct IntTarget {
+                GLuint fbo = 0;
+                GLuint texture = 0;
+            };
+
+            // The "nothing drawn here" value is UPLOADED, not cleared: the CTS fills its R32I
+            // targets the same way (fillTexture), and an upload cannot be confused with a clear
+            // that a backend defers, reorders or drops - which is exactly the ambiguity a case
+            // asserting "this cell must be untouched" cannot afford.
+            static void FillIntTarget(const IntTarget& target, int width, int height) {
+                const std::vector<GLint> unwritten(static_cast<size_t>(width) * height, kUnwritten);
+                glBindTexture(GL_TEXTURE_2D, target.texture);
+                glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED_INTEGER, GL_INT, unwritten.data());
+            }
+
+            static IntTarget MakeIntTarget(int width, int height) {
+                IntTarget target;
+                glGenTextures(1, &target.texture);
+                glBindTexture(GL_TEXTURE_2D, target.texture);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glTexImage2D(GL_TEXTURE_2D, 0, GL_R32I, width, height, 0, GL_RED_INTEGER, GL_INT, nullptr);
+                glGenFramebuffers(1, &target.fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+                glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, target.texture, 0);
+                FillIntTarget(target, width, height);
+                return target;
+            }
+
+            static void DestroyIntTarget(IntTarget& target) {
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                if (target.fbo != 0) glDeleteFramebuffers(1, &target.fbo);
+                if (target.texture != 0) glDeleteTextures(1, &target.texture);
+            }
+
+            static std::vector<GLint> ReadInts(int width, int height) {
+                std::vector<GLint> pixels(static_cast<size_t>(width) * height, 0);
+                glReadPixels(0, 0, width, height, GL_RED_INTEGER, GL_INT, pixels.data());
+                return pixels;
+            }
+
+            // The centre of grid cell (x, y), in the bottom-left-origin coordinates glReadPixels
+            // returns. Sampling the centre rather than a corner keeps the assertion about WHICH
+            // viewport was selected rather than about edge rounding.
+            static GLint CellCentre(const std::vector<GLint>& pixels, int stride, int x, int y) {
+                const int px = x * kCellSize + kCellSize / 2;
+                const int py = y * kCellSize + kCellSize / 2;
+                return pixels[static_cast<size_t>(py) * stride + px];
+            }
+
+            std::string m_buildLog;
+            GLuint m_program = 0;
+            GLuint m_vao = 0;
+        };
+
+        // --- 1. the viewport rectangle -------------------------------------------------------
+
+        TEST_F(ViewportArrayScenario, EachViewportIndexRasterizesIntoItsOwnRectangle) {
+            IntTarget target = MakeIntTarget(kSurfaceSide, kSurfaceSide);
+            SetupGridViewports(kCellSize, kCellSize);
+            glUseProgram(m_program);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const GLint expected = y * kGridSide + x;
+                    EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), expected)
+                        << "cell (" << x << ", " << y << ") should hold viewport index " << expected
+                        << "; a single-viewport backend paints the whole image with 15 (the last invocation)";
+                }
+            }
+            DestroyIntTarget(target);
+        }
+
+        // The same claim against the DEFAULT framebuffer, where MobileGL applies its Y-flip and
+        // pre-transform rotation. Index 0 alone getting the mapping is the classic bug.
+        TEST_F(ViewportArrayScenario, TheDefaultFramebufferAppliesTheSameFlipToEveryViewport) {
+            const int surfaceW = Gl().Width();
+            const int surfaceH = Gl().Height();
+            ASSERT_GE(surfaceW, kGridSide);
+            ASSERT_GE(surfaceH, kGridSide);
+            const int cellW = surfaceW / kGridSide;
+            const int cellH = surfaceH / kGridSide;
+
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            // Paint a value no viewport index can produce, so an unwritten cell is obvious.
+            glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
+            glClear(GL_COLOR_BUFFER_BIT);
+
+            // The default framebuffer is 8-bit RGBA, so the index travels as a colour: cell i is
+            // painted with red = i * 16, which is exact in 8 bits for i in [0, 16).
+            const char* const kColorFragmentSource = R"(#version 410 core
+flat in int gsIndex;
+layout(location = 0) out vec4 fragColor;
+void main() { fragColor = vec4(float(gsIndex) * 16.0 / 255.0, 0.0, 0.0, 1.0); }
+)";
+            const GLuint colorProgram = BuildProgram(kGridGeometrySource, kColorFragmentSource);
+            ASSERT_NE(colorProgram, 0u) << "colour program failed to build: " << m_buildLog;
+
+            SetupGridViewports(cellW, cellH);
+            glUseProgram(colorProgram);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            std::vector<unsigned char> pixels(static_cast<size_t>(surfaceW) * surfaceH * 4, 0);
+            glReadPixels(0, 0, surfaceW, surfaceH, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const int px = x * cellW + cellW / 2;
+                    const int py = y * cellH + cellH / 2;
+                    const int red = pixels[(static_cast<size_t>(py) * surfaceW + px) * 4];
+                    const int expected = (y * kGridSide + x) * 16;
+                    // One LSB of slack for an 8-bit round trip; the values are 16 apart, so this
+                    // cannot confuse two neighbouring indices.
+                    EXPECT_LE(std::abs(red - expected), 1)
+                        << "default-framebuffer cell (" << x << ", " << y << ") holds red=" << red << ", expected "
+                        << expected << ". A vertically mirrored grid means the Y-flip was applied to viewport 0 "
+                        << "only";
+                }
+            }
+            glDeleteProgram(colorProgram);
+        }
+
+        // --- 2. the depth range --------------------------------------------------------------
+
+        TEST_F(ViewportArrayScenario, EachViewportIndexUsesItsOwnDepthRange) {
+            // 16 columns one pixel wide and two rows tall: row 0 gets the near-plane quad, row 1
+            // the far-plane one, so both ends of viewport i's range land in the same column.
+            constexpr int kWidth = kViewportCount;
+            constexpr int kHeight = 2;
+
+            GLuint texture = 0;
+            GLuint fbo = 0;
+            glGenTextures(1, &texture);
+            glBindTexture(GL_TEXTURE_2D, texture);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+            glTexImage2D(GL_TEXTURE_2D, 0, GL_R32F, kWidth, kHeight, 0, GL_RED, GL_FLOAT, nullptr);
+            glGenFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, 0);
+            const GLfloat clearValue[4] = {-1.0f, 0.0f, 0.0f, 0.0f};
+            glClearBufferfv(GL_COLOR, 0, clearValue);
+
+            std::vector<GLfloat> viewports(static_cast<size_t>(kViewportCount) * 4);
+            std::vector<GLdouble> ranges(static_cast<size_t>(kViewportCount) * 2);
+            for (int i = 0; i < kViewportCount; ++i) {
+                viewports[static_cast<size_t>(i) * 4 + 0] = static_cast<GLfloat>(i);
+                viewports[static_cast<size_t>(i) * 4 + 1] = 0.0f;
+                viewports[static_cast<size_t>(i) * 4 + 2] = 1.0f;
+                viewports[static_cast<size_t>(i) * 4 + 3] = 2.0f;
+                ranges[static_cast<size_t>(i) * 2 + 0] = static_cast<GLdouble>(i) / 16.0;
+                ranges[static_cast<size_t>(i) * 2 + 1] = 1.0 - static_cast<GLdouble>(i) / 16.0;
+            }
+            glViewportArrayv(0, kViewportCount, viewports.data());
+            glDepthRangeArrayv(0, kViewportCount, ranges.data());
+
+            const GLuint depthProgram = BuildProgram(kDepthGeometrySource, kDepthFragmentSource);
+            ASSERT_NE(depthProgram, 0u) << "depth program failed to build: " << m_buildLog;
+            glUseProgram(depthProgram);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            std::vector<GLfloat> pixels(static_cast<size_t>(kWidth) * kHeight, 0.0f);
+            glReadPixels(0, 0, kWidth, kHeight, GL_RED, GL_FLOAT, pixels.data());
+            for (int i = 0; i < kViewportCount; ++i) {
+                const float nearDepth = static_cast<float>(i) / 16.0f;
+                const float farDepth = 1.0f - static_cast<float>(i) / 16.0f;
+                // The tolerance covers depth-buffer-free rasterization of gl_FragCoord.z on a
+                // software rasterizer; the per-index values are 1/16 apart, so it cannot let a
+                // neighbouring viewport's range through, and viewport 0's range (0, 1) differs
+                // from every other index by at least 1/16.
+                EXPECT_NEAR(pixels[i], nearDepth, 1.0e-3f)
+                    << "viewport " << i << " near-plane depth; got viewport 0's range if this is 0";
+                EXPECT_NEAR(pixels[static_cast<size_t>(kWidth) + i], farDepth, 1.0e-3f)
+                    << "viewport " << i << " far-plane depth; got viewport 0's range if this is 1";
+            }
+
+            glDeleteProgram(depthProgram);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            glDeleteFramebuffers(1, &fbo);
+            glDeleteTextures(1, &texture);
+        }
+
+        // --- 3. the per-index scissor-test enable --------------------------------------------
+
+        TEST_F(ViewportArrayScenario, AnIndexedScissorEnableClipsOnlyThatIndex) {
+            IntTarget target = MakeIntTarget(kSurfaceSide, kSurfaceSide);
+
+            // One full-size viewport per index so the scissor rectangle is the ONLY thing that
+            // can shrink the quad - the same separation KHR-GL43.viewport_array.scissor uses.
+            glViewport(0, 0, kSurfaceSide, kSurfaceSide);
+            std::vector<GLint> boxes(static_cast<size_t>(kViewportCount) * 4);
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const size_t base = static_cast<size_t>(y * kGridSide + x) * 4;
+                    boxes[base + 0] = x * kCellSize;
+                    boxes[base + 1] = y * kCellSize;
+                    boxes[base + 2] = kCellSize;
+                    boxes[base + 3] = kCellSize;
+                }
+            }
+            glScissorArrayv(0, kViewportCount, boxes.data());
+
+            const GLuint singleProgram = BuildProgram(kSingleGeometrySource, kIntFragmentSource);
+            ASSERT_NE(singleProgram, 0u) << "single-viewport program failed to build: " << m_buildLog;
+            glUseProgram(singleProgram);
+            glBindVertexArray(m_vao);
+            const GLint uViewport = glGetUniformLocation(singleProgram, "uViewport");
+            ASSERT_NE(uViewport, -1);
+
+            constexpr GLint kProbeIndex = 6;  // grid cell (2, 1)
+            constexpr int kProbeX = kProbeIndex % kGridSide;
+            constexpr int kProbeY = kProbeIndex / kGridSide;
+
+            // (a) scissor test ENABLED for this index: the quad is clipped to its 32x32 box.
+            glUniform1i(uViewport, kProbeIndex);
+            glEnablei(GL_SCISSOR_TEST, kProbeIndex);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+            {
+                const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+                EXPECT_EQ(CellCentre(pixels, kSurfaceSide, kProbeX, kProbeY), kProbeIndex)
+                    << "the scissored index must still paint inside its own box";
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        if (x == kProbeX && y == kProbeY) continue;
+                        EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), kUnwritten)
+                            << "cell (" << x << ", " << y << ") is outside scissor rectangle " << kProbeIndex
+                            << " and must be untouched";
+                    }
+                }
+            }
+
+            // (b) scissor test DISABLED for the same index, everything else identical: with no
+            // per-viewport toggle in Vulkan this is the case that needs the disabled index to be
+            // given the full framebuffer rectangle, and it is exactly where "leave the last
+            // rectangle bound" would show up as a still-clipped quad.
+            FillIntTarget(target, kSurfaceSide, kSurfaceSide);
+            glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+            glDisablei(GL_SCISSOR_TEST, kProbeIndex);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+            {
+                const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), kProbeIndex)
+                            << "with the scissor test off for index " << kProbeIndex
+                            << ", its full-viewport quad must cover cell (" << x << ", " << y << ")";
+                    }
+                }
+            }
+
+            glDeleteProgram(singleProgram);
+            DestroyIntTarget(target);
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/XfbCaptureBufferReuseScenario.cpp

@@ -0,0 +1,317 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/XfbCaptureBufferReuseScenario.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
+//
+// ONE capture buffer, SEVERAL capture spans - the shape most KHR-GL4x cases that
+// use transform feedback as a readback channel are built on. They allocate the
+// capture buffer once in a setup step and then run span after span through it,
+// so a defect that only shows from the second span onwards fails the whole case
+// while the first span (and every single-span scenario in this suite) stays
+// green. The first thing checked here is therefore not the capture itself but
+// that the bytes the capture wrote are the bytes the readback reads.
+//
+// Two ways of reusing the buffer, because they exercise different machinery:
+//
+//   * respecified between spans (glBufferData while the buffer is still bound to
+//     the transform-feedback binding point), which is what a test helper that
+//     poisons its capture buffer before every span does;
+//   * allocated ONCE with immutable storage and never touched again, which is
+//     what KHR-GL45.direct_state_access.vertex_arrays_enable_disable_attributes
+//     does - glBufferStorage(4 bytes) in its setup, then two draws.
+//
+// The negative control (a fresh buffer object per span) is a separate case
+// rather than a parameter: it is the configuration that already worked, so it
+// has to keep working for the others to mean anything.
+
+#include <cmath>
+#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 float kPoison = -1234.0f;
+        // One vec4 per point, one point per draw.
+        constexpr std::size_t kCaptureFloats = 4;
+        constexpr std::size_t kCaptureBytes = kCaptureFloats * sizeof(float);
+
+        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;
+        }
+
+        // Vertex-only capture program: whatever the draw fetched at location 0 comes
+        // straight back out through the capture. Runs under GL_RASTERIZER_DISCARD, so
+        // there is no fragment stage.
+        GLuint BuildCaptureProgram(std::string* log) {
+            const std::string vertexSource = R"(#version 430 core
+layout(location = 0) in vec4 vs_in_value;
+out vec4 vs_out_value;
+void main() {
+  vs_out_value = vs_in_value;
+}
+)";
+            const GLuint vertexShader = CompileShader(GL_VERTEX_SHADER, vertexSource, log);
+            if (vertexShader == 0) return 0;
+            const GLuint program = glCreateProgram();
+            glAttachShader(program, vertexShader);
+            const char* varying = "vs_out_value";
+            glTransformFeedbackVaryings(program, 1, &varying, GL_INTERLEAVED_ATTRIBS);
+            glLinkProgram(program);
+            glDeleteShader(vertexShader);
+            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;
+        }
+
+        class XfbCaptureBufferReuseScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                std::string log;
+                m_program = BuildCaptureProgram(&log);
+                ASSERT_NE(m_program, 0u) << "capture program failed to build: " << log;
+
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferData(GL_ARRAY_BUFFER, kCaptureBytes, nullptr, GL_DYNAMIC_DRAW);
+                glVertexAttribPointer(0, 4, GL_FLOAT, GL_FALSE, 0, nullptr);
+                glEnableVertexAttribArray(0);
+                glBindBuffer(GL_ARRAY_BUFFER, 0);
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glBindVertexArray(0);
+                if (m_vbo != 0) glDeleteBuffers(1, &m_vbo);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_program != 0) glDeleteProgram(m_program);
+                glUseProgram(0);
+                ScenarioTest::TearDown();
+            }
+
+            // The vertex the next span will fetch and capture.
+            void SetVertex(float value) {
+                const float data[kCaptureFloats] = {value, value + 1.0f, value + 2.0f, value + 3.0f};
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferSubData(GL_ARRAY_BUFFER, 0, kCaptureBytes, data);
+                glBindBuffer(GL_ARRAY_BUFFER, 0);
+            }
+
+            // One capture span over the buffer currently bound to capture point 0.
+            void RunSpan() {
+                glEnable(GL_RASTERIZER_DISCARD);
+                glUseProgram(m_program);
+                glBindVertexArray(m_vao);
+                glBeginTransformFeedback(GL_POINTS);
+                glDrawArrays(GL_POINTS, 0, 1);
+                glEndTransformFeedback();
+                glDisable(GL_RASTERIZER_DISCARD);
+                glUseProgram(0);
+            }
+
+            static ::testing::AssertionResult CapturedIs(const float* data, float value) {
+                for (std::size_t i = 0; i < kCaptureFloats; ++i) {
+                    const float expected = value + static_cast<float>(i);
+                    const float got = data[i];
+                    // isfinite first: every ordered comparison against a NaN is false, so a
+                    // pair of one-sided range tests REPORTS SUCCESS for uninitialised
+                    // storage that happens to read as NaN - which is exactly the failure
+                    // these scenarios exist to catch.
+                    if (!std::isfinite(got) || std::fabs(got - expected) > 0.01f) {
+                        return ::testing::AssertionFailure()
+                               << "component " << i << " is " << got << ", expected " << expected
+                               << (got == kPoison ? " (the capture never reached these bytes)" : "");
+                    }
+                }
+                return ::testing::AssertionSuccess();
+            }
+
+            GLuint m_program = 0;
+            GLuint m_vao = 0;
+            GLuint m_vbo = 0;
+        };
+
+        // The negative control: one buffer object per span. This is the configuration
+        // every multi-span scenario in this suite works around the others with, so it
+        // has to hold or nothing below is interpretable.
+        TEST_F(XfbCaptureBufferReuseScenario, EverySpanIntoABufferObjectOfItsOwn) {
+            if (!Ready()) GTEST_SKIP();
+
+            for (int span = 0; span < 3; ++span) {
+                const float value = 10.0f * static_cast<float>(span + 1);
+                const std::vector<float> poison(kCaptureFloats, kPoison);
+
+                GLuint xfbBuffer = 0;
+                glGenBuffers(1, &xfbBuffer);
+                glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, xfbBuffer);
+                glBufferData(GL_TRANSFORM_FEEDBACK_BUFFER, kCaptureBytes, poison.data(), GL_DYNAMIC_DRAW);
+
+                SetVertex(value);
+                RunSpan();
+
+                float readback[kCaptureFloats] = {kPoison, kPoison, kPoison, kPoison};
+                glGetBufferSubData(GL_TRANSFORM_FEEDBACK_BUFFER, 0, kCaptureBytes, readback);
+                EXPECT_TRUE(CapturedIs(readback, value)) << "span " << span;
+
+                glDeleteBuffers(1, &xfbBuffer);
+            }
+            EXPECT_EQ(glGetError(), GL_NO_ERROR);
+        }
+
+        // The same three spans through ONE buffer object, respecified before each of
+        // them WHILE it is bound to capture point 0 - a helper poisoning its capture
+        // buffer, which is what makes "captured nothing" legible in the first place.
+        //
+        // A respecification is free to replace the storage underneath (that is what
+        // orphaning is), and on a buffer whose bytes the backend has already handed
+        // the frontend a pointer into, the replacement has to reach that pointer too.
+        // It did not: the capture wrote the new storage and the readback kept reading
+        // the old one, so every span after the first came back poison.
+        TEST_F(XfbCaptureBufferReuseScenario, EverySpanIntoOneRespecifiedBufferObject) {
+            if (!Ready()) GTEST_SKIP();
+
+            GLuint xfbBuffer = 0;
+            glGenBuffers(1, &xfbBuffer);
+            glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, xfbBuffer);
+
+            for (int span = 0; span < 3; ++span) {
+                const float value = 10.0f * static_cast<float>(span + 1);
+                const std::vector<float> poison(kCaptureFloats, kPoison);
+                glBufferData(GL_TRANSFORM_FEEDBACK_BUFFER, kCaptureBytes, poison.data(), GL_DYNAMIC_DRAW);
+
+                SetVertex(value);
+                RunSpan();
+
+                float readback[kCaptureFloats] = {kPoison, kPoison, kPoison, kPoison};
+                glGetBufferSubData(GL_TRANSFORM_FEEDBACK_BUFFER, 0, kCaptureBytes, readback);
+                EXPECT_TRUE(CapturedIs(readback, value)) << "span " << span;
+            }
+
+            glDeleteBuffers(1, &xfbBuffer);
+            EXPECT_EQ(glGetError(), GL_NO_ERROR);
+        }
+
+        // A respecification that CHANGES the size, which is the case a re-pointing
+        // that only handled same-size storage would still get wrong - and, before the
+        // fix, the case that wrote the new (larger) contents through a mapping sized
+        // for the old ones.
+        TEST_F(XfbCaptureBufferReuseScenario, ARespecificationMayChangeTheCaptureBufferSize) {
+            if (!Ready()) GTEST_SKIP();
+
+            GLuint xfbBuffer = 0;
+            glGenBuffers(1, &xfbBuffer);
+            glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, xfbBuffer);
+
+            // Sized for one point, then for four, then back down to one.
+            const std::size_t pointCapacity[] = {1, 4, 1};
+            for (int span = 0; span < 3; ++span) {
+                const float value = 10.0f * static_cast<float>(span + 1);
+                const std::size_t floats = kCaptureFloats * pointCapacity[span];
+                const std::vector<float> poison(floats, kPoison);
+                glBufferData(GL_TRANSFORM_FEEDBACK_BUFFER, static_cast<GLsizeiptr>(floats * sizeof(float)),
+                             poison.data(), GL_DYNAMIC_DRAW);
+
+                SetVertex(value);
+                RunSpan();
+
+                std::vector<float> readback(floats, kPoison);
+                glGetBufferSubData(GL_TRANSFORM_FEEDBACK_BUFFER, 0,
+                                   static_cast<GLsizeiptr>(floats * sizeof(float)), readback.data());
+                EXPECT_TRUE(CapturedIs(readback.data(), value)) << "span " << span;
+                // The bytes past the one point the draw produced must still be the
+                // poison the respecification put there, not whatever the previous
+                // (differently sized) storage held.
+                for (std::size_t i = kCaptureFloats; i < floats; ++i) {
+                    EXPECT_FLOAT_EQ(readback[i], kPoison) << "span " << span << " float " << i;
+                }
+            }
+
+            glDeleteBuffers(1, &xfbBuffer);
+            EXPECT_EQ(glGetError(), GL_NO_ERROR);
+        }
+
+        // The KHR-GL45.direct_state_access.vertex_arrays_enable_disable_attributes
+        // shape: the capture buffer gets IMMUTABLE storage once, in a setup step, and
+        // is never respecified - two spans simply run through it, each read back with
+        // glMapBuffer. Nothing here may depend on a respecification to reset the
+        // capture: glBeginTransformFeedback does that on its own.
+        TEST_F(XfbCaptureBufferReuseScenario, EverySpanIntoOneImmutableStorageBuffer) {
+            if (!Ready()) GTEST_SKIP();
+
+            GLuint xfbBuffer = 0;
+            glGenBuffers(1, &xfbBuffer);
+            glBindBuffer(GL_TRANSFORM_FEEDBACK_BUFFER, xfbBuffer);
+            // Poisoned at creation - the storage is immutable, so this is the only chance to
+            // put a recognisable value there, and without it a span that captured nothing
+            // would be indistinguishable from one that captured the right thing whenever the
+            // untouched bytes happened to read back as the expected number.
+            const std::vector<float> poison(kCaptureFloats, kPoison);
+            glBufferStorage(GL_TRANSFORM_FEEDBACK_BUFFER, kCaptureBytes, poison.data(), GL_MAP_READ_BIT);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR) << "glBufferStorage on the capture buffer";
+            glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, xfbBuffer);
+
+            for (int span = 0; span < 3; ++span) {
+                const float value = 10.0f * static_cast<float>(span + 1);
+                SetVertex(value);
+                RunSpan();
+
+                const void* mapped = glMapBuffer(GL_TRANSFORM_FEEDBACK_BUFFER, GL_READ_ONLY);
+                ASSERT_NE(mapped, nullptr) << "span " << span << ": glMapBuffer returned null";
+                float readback[kCaptureFloats] = {kPoison, kPoison, kPoison, kPoison};
+                std::memcpy(readback, mapped, kCaptureBytes);
+                glUnmapBuffer(GL_TRANSFORM_FEEDBACK_BUFFER);
+                EXPECT_TRUE(CapturedIs(readback, value)) << "span " << span;
+            }
+
+            glBindBuffer(GL_TRANSFORM_FEEDBACK_BUFFER, 0);
+            glDeleteBuffers(1, &xfbBuffer);
+            EXPECT_EQ(glGetError(), GL_NO_ERROR);
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_State/GLState/BufferState/BufferObject.cpp

@@ -75,10 +75,42 @@ namespace MobileGL::MG_State::GLState {
         NotifySubData(offset, size);
     }
 
-    void BufferObject::Respecify(SizeT size, const void* data) {
+    // A (re)definition of the store is about to write `size` bytes through Bytes().
-        ReleaseMemory();
+    // Sizing the shadow is all that takes for a shadow-backed buffer. A buffer whose
+    // bytes were adopted into backend GPU memory has to give the adoption back first,
+    // because the mapping it holds describes exactly the OLD store: writing the new
+    // contents through it runs past its end the moment the store grows, and a backend
+    // that replaces the storage for the new store - which is what an orphaning
+    // respecification asks for - would leave that mapping, and therefore every later
+    // read of this buffer, addressing storage nothing writes to any more. That was the
+    // transform feedback capture that wrote one buffer while the readback read another.
+    //
+    // Given back rather than renewed here, deliberately. Renewing in place would mean
+    // memcpying the new contents into storage that submitted-but-unretired draws may
+    // still be reading, which is precisely what the orphaning idiom exists to avoid;
+    // avoiding THAT would mean either stalling on a fence in the middle of a frame or
+    // teaching the persistent-map op to orphan, and the op must never orphan for the
+    // other kind of caller (an application-held GL_MAP_PERSISTENT_BIT mapping, whose
+    // pointer has to stay valid for the buffer's whole life). Handing the store back to
+    // the CPU shadow needs none of that: the backend's ordinary respecification path
+    // then does the busy-tracking and the conditional orphan it has always done, and the
+    // next binding that wants GPU residency takes a fresh mapping of the new store.
+    void BufferObject::RedefineStorage(SizeT size) {
+        if (m_resource.IsGpuResident()) {
+            m_resource.ReleasePersistentMap();
+            // Whatever a shader or a capture wrote is in the store being replaced, so
+            // there is nothing left to reconcile - and leaving the flag set would make
+            // the next read of this buffer wait for GPU work on behalf of bytes the
+            // application has just thrown away.
+            m_gpuWritePending = false;
+        }
         m_size = size;
         m_resource.ResizeShadow(size);
+    }
+
+    void BufferObject::Respecify(SizeT size, const void* data) {
+        ReleaseMemory();
+        RedefineStorage(size);
         if (data && size > 0) {
             Memcpy(m_resource.Bytes(), data, size);
         }
@@ -96,8 +128,7 @@ namespace MobileGL::MG_State::GLState {
 
     void BufferObject::AllocateImmutableStorage(SizeT size, const void* data, GLbitfield storageFlags) {
         ReleaseMemory();
-        m_size = size;
+        RedefineStorage(size);
-        m_resource.ResizeShadow(size);
         if (data) {
             Memcpy(m_resource.Bytes(), data, size);
         } else if (size > 0) {
@@ -219,6 +250,34 @@ namespace MobileGL::MG_State::GLState {
         NotifyContentWrite(atOffset, data.size);
     }
 
+    void BufferObject::FillSubData(DataPtr pattern, SizeT atOffset, SizeT size) {
+        MOBILEGL_ASSERT(pattern.data != nullptr && pattern.size > 0,
+                        "FillSubData requires a non-empty pattern.");
+        MOBILEGL_ASSERT(size % pattern.size == 0,
+                        "FillSubData size (%zu) must be a multiple of pattern size (%zu).", size, pattern.size);
+        MOBILEGL_ASSERT(atOffset <= m_size && size <= m_size - atOffset,
+                        "FillSubData out of bounds: atOffset (%zu) + size (%zu) > m_size (%zu)", atOffset, size,
+                        m_size);
+        MOBILEGL_ASSERT(!m_isMapped || (m_mappingAccess & BufferMappingAccessBit::Persistent),
+                        "Cannot fill data while buffer is non-persistently mapped.");
+        if (size == 0) return;
+
+        // A clear is ordered after all earlier GPU writes. Partial clears additionally need the
+        // retained shadow bytes; whole-store clears need the same synchronization before writing
+        // an adopted persistent mapping that the GPU may still be accessing.
+        SyncGpuWrites();
+
+        Uint8* dst = m_resource.Bytes() + atOffset;
+        if (pattern.size == 1) {
+            Memset(dst, *static_cast<const Uint8*>(pattern.data), size);
+        } else {
+            for (SizeT at = 0; at < size; at += pattern.size) {
+                Memcpy(dst + at, pattern.data, pattern.size);
+            }
+        }
+        NotifyContentWrite(atOffset, size);
+    }
+
     void BufferObject::DownloadSubData(void* dst, SizeT atOffset, SizeT size) const {
         MOBILEGL_ASSERT(atOffset + size <= m_size,
                         "DownloadSubData out of bounds: atOffset (%zu) + size (%zu) > m_size (%zu)", atOffset, size,

MobileGL/MG_State/GLState/BufferState/BufferObject.h

@@ -132,6 +132,9 @@ namespace MobileGL {
 
             void UploadData(DataPtr data, SizeT atOffset);
             void UploadSubData(DataPtr data, SizeT atOffset);
+            // Repeats one already-converted element through [atOffset, atOffset + size) and
+            // publishes the range as one content mutation.
+            void FillSubData(DataPtr pattern, SizeT atOffset, SizeT size);
             // Reads `size` bytes from the CPU shadow at `atOffset` into `dst` (glGetBufferSubData).
             // The shadow reflects CPU writes (BufferData/SubData/maps) and backend write-backs, but not
             // arbitrary GPU-side writes.
@@ -205,6 +208,9 @@ namespace MobileGL {
             void SetBackendResource(SharedPtr<BackendBufferResource> resource);
 
         private:
+            // Sizes the store for a (re)definition, renewing an adopted GPU-resident
+            // mapping across it. See the definition for why the renewal is not optional.
+            void RedefineStorage(SizeT size);
             void NotifyRespecify();
             void NotifySubData(SizeT offset, SizeT size);
             void NotifyFlushMappedRange(Range1D range, Flags<BufferMappingAccessBit> appAccess);

MobileGL/MG_State/GLState/BufferState/PipeResource.h

@@ -70,6 +70,15 @@ namespace MobileGL::MG_State::GLState {
             m_shadow->shrink_to_fit();
         }
 
+        // Give the adoption back: the bytes resolve against the shadow again (which
+        // the caller must (re)size, it was released on adoption). Used when the store
+        // itself is redefined - the mapping describes exactly the store that is going
+        // away, so it may neither be written through nor kept. It is NOT a general
+        // "unmap": a persistent map the application holds outlives every unmap by
+        // definition, and the calls that could redefine such a buffer's store are
+        // errors the frontend refuses before reaching here.
+        void ReleasePersistentMap() { m_gpuMapped = nullptr; }
+
         // Backend GPU resource, owned here in both modes.
         const SharedPtr<BackendBufferResource>& Backend() const { return m_backend; }
         void SetBackend(SharedPtr<BackendBufferResource> backend) { m_backend = std::move(backend); }

MobileGL/MG_State/GLState/Core.cpp

@@ -39,6 +39,11 @@ namespace MobileGL::MG_State {
             return m_compileEnv;
         }
 
+        void GLContext::InvalidateCompileEnv() {
+            m_compileEnv.reset();
+            m_compileEnvBackend = nullptr;
+        }
+
         // Error
         void GLContext::RecordError(ErrorCode code, UniquePtr<ErrorInfo> info) {
             // Invariant I1, mechanically enforced: the GL error state is GL-thread-owned.
@@ -196,7 +201,7 @@ namespace MobileGL::MG_State {
         // (which expands to nothing outside debug builds).
         void GLContext::SetCurrentVertexAttributeFloat(Uint index, const Array<Float, 4>& value) {
             if (index >= m_currentVertexAttributes.size()) {
-                MGLOG_E("SetCurrentVertexAttributeFloat: index %u is out of range", index);
+                MGLOG_E_ONCE("SetCurrentVertexAttributeFloat: index %u is out of range", index);
                 return;
             }
 
@@ -210,7 +215,7 @@ namespace MobileGL::MG_State {
 
         void GLContext::SetCurrentVertexAttributeInt(Uint index, const Array<Int32, 4>& value) {
             if (index >= m_currentVertexAttributes.size()) {
-                MGLOG_E("SetCurrentVertexAttributeInt: index %u is out of range", index);
+                MGLOG_E_ONCE("SetCurrentVertexAttributeInt: index %u is out of range", index);
                 return;
             }
 
@@ -224,7 +229,7 @@ namespace MobileGL::MG_State {
 
         void GLContext::SetCurrentVertexAttributeUint(Uint index, const Array<Uint32, 4>& value) {
             if (index >= m_currentVertexAttributes.size()) {
-                MGLOG_E("SetCurrentVertexAttributeUint: index %u is out of range", index);
+                MGLOG_E_ONCE("SetCurrentVertexAttributeUint: index %u is out of range", index);
                 return;
             }
 
@@ -239,7 +244,7 @@ namespace MobileGL::MG_State {
         const CurrentVertexAttributeValue& GLContext::GetCurrentVertexAttribute(Uint index) const {
             static const CurrentVertexAttributeValue defaultValue{};
             if (index >= m_currentVertexAttributes.size()) {
-                MGLOG_E("GetCurrentVertexAttribute: index %u is out of range", index);
+                MGLOG_E_ONCE("GetCurrentVertexAttribute: index %u is out of range", index);
                 return defaultValue;
             }
             return m_currentVertexAttributes[index];
@@ -641,9 +646,17 @@ namespace MobileGL::MG_State {
             for (SizeT stage = 0; stage < ProgramPipelineObject::kGraphicsStageCount; ++stage) {
                 const auto& stageProgram = pipeline->GetStageProgram(static_cast<ShaderStage>(stage));
                 if (!stageProgram) continue;
-                for (const auto& shader : stageProgram->GetAttachedShaders()) {
+                // The stage program contributes the shaders its LAST LINK consumed, never
-                    if (!shader || static_cast<SizeT>(shader->GetShaderStage()) != stage) continue;
+                // its live attach list: per GL 4.6 7.3/7.4 a pipeline stage executes the
-                    composite->AttachShader(shader);
+                // stage program as last linked - glAttachShader and glCompileShader take
+                // effect only at the program's next link - and neither of those moves the
+                // link version this cache keys on, so reading live state here would let a
+                // post-link attach or recompile leak into the composite while the signature
+                // still hits. The pinned (source, node) makes the composite's Link()
+                // consume the very inputs that link consumed.
+                for (const auto& ref : stageProgram->GetLinkedShaderSnapshot()) {
+                    if (!ref.shader || static_cast<SizeT>(ref.shader->GetShaderStage()) != stage) continue;
+                    composite->AttachShaderWithPinnedLinkInput(ref);
                     anyStage = true;
                 }
             }
@@ -712,10 +725,18 @@ namespace MobileGL::MG_State {
             m_renderState.SetViewport(viewport);
         }
 
-        const IntVec4& GLContext::GetViewport() const {
+        IntVec4 GLContext::GetViewport() const {
             return m_renderState.GetViewport();
         }
 
+        void GLContext::SetViewportIndexed(Uint index, FloatVec4 viewport) {
+            m_renderState.SetViewportIndexed(index, viewport);
+        }
+
+        const FloatVec4& GLContext::GetViewportIndexed(Uint index) const {
+            return m_renderState.GetViewportIndexed(index);
+        }
+
         void GLContext::SetLineWidth(Float width) {
             m_renderState.SetLineWidth(width);
         }
@@ -953,6 +974,14 @@ namespace MobileGL::MG_State {
             return m_renderState.GetDepthRange();
         }
 
+        void GLContext::SetDepthRangeIndexed(Uint index, FloatVec2 range) {
+            m_renderState.SetDepthRangeIndexed(index, range);
+        }
+
+        const FloatVec2& GLContext::GetDepthRangeIndexed(Uint index) const {
+            return m_renderState.GetDepthRangeIndexed(index);
+        }
+
         void GLContext::SetSampleCoverage(Float value, Bool invert) {
             m_renderState.SetSampleCoverage(value, invert);
         }
@@ -1017,6 +1046,14 @@ namespace MobileGL::MG_State {
             return m_renderState.GetScissorBox();
         }
 
+        void GLContext::SetScissorBoxIndexed(Uint index, IntVec4 box) {
+            m_renderState.SetScissorBoxIndexed(index, box);
+        }
+
+        const IntVec4& GLContext::GetScissorBoxIndexed(Uint index) const {
+            return m_renderState.GetScissorBoxIndexed(index);
+        }
+
         // Framebuffer
         void GLContext::GenFramebufferNames(Uint number, Vector<Uint>& framebuffers) {
             m_framebufferState.GenerateNames(number, framebuffers);

MobileGL/MG_State/GLState/Core.h

@@ -198,8 +198,10 @@ namespace MobileGL {
                 // Only the pipeline-relevant subset - see RenderState::m_pipelineStateVersion.
                 Uint GetPipelineStateVersion() const;
                 const RenderStateParameters& GetRenderStateParameters() const;
-                void SetViewport(IntVec4 viewport); // x, y, width, height
+                void SetViewport(IntVec4 viewport); // x, y, width, height; writes ALL viewports
-                const IntVec4& GetViewport() const; // x, y, width, height
+                IntVec4 GetViewport() const;        // x, y, width, height; viewport 0, rounded
+                void SetViewportIndexed(Uint index, FloatVec4 viewport);
+                const FloatVec4& GetViewportIndexed(Uint index) const;
                 void SetLineWidth(Float width);
                 Float GetLineWidth() const;
                 void SetPointSize(Float size);
@@ -260,8 +262,10 @@ namespace MobileGL {
                 Uint32 GetClearStencil() const;
                 void SetBlendColor(FloatVec4 color);
                 const FloatVec4& GetBlendColor() const;
-                void SetDepthRange(FloatVec2 range);
+                void SetDepthRange(FloatVec2 range); // writes ALL viewports' depth ranges
                 const FloatVec2& GetDepthRange() const;
+                void SetDepthRangeIndexed(Uint index, FloatVec2 range);
+                const FloatVec2& GetDepthRangeIndexed(Uint index) const;
                 void SetSampleCoverage(Float value, Bool invert);
                 Float GetSampleCoverageValue() const;
                 Bool GetSampleCoverageInvert() const;
@@ -276,8 +280,10 @@ namespace MobileGL {
                 FrontFaceMode GetFrontFaceMode() const;
                 void SetProvokingVertexMode(ProvokingVertexMode mode);
                 ProvokingVertexMode GetProvokingVertexMode() const;
-                void SetScissorBox(IntVec4 box);      // x, y, width, height
+                void SetScissorBox(IntVec4 box);      // x, y, width, height; writes ALL rectangles
-                const IntVec4& GetScissorBox() const; // x, y, width, height
+                const IntVec4& GetScissorBox() const; // x, y, width, height; rectangle 0
+                void SetScissorBoxIndexed(Uint index, IntVec4 box);
+                const IntVec4& GetScissorBoxIndexed(Uint index) const;
 
                 // Transform feedback. The fields below are the state of the transform
                 // feedback object currently bound to GL_TRANSFORM_FEEDBACK; see the object
@@ -407,9 +413,12 @@ namespace MobileGL {
                 // 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.
+                // every P0b preprocess memo keyed against the old one. A backend whose dynamic
+                // capabilities become available without changing object identity must call
+                // InvalidateCompileEnv() after publishing them.
                 // GL thread only.
                 const SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv>& GetCompileEnv();
+                void InvalidateCompileEnv();
 
             private:
                 // State Components

MobileGL/MG_State/GLState/ErrorState/Error.cpp

@@ -14,10 +14,10 @@
 namespace MobileGL::MG_State::GLState {
     void ErrorState::RecordError(ErrorCode code, UniquePtr<ErrorInfo> info) {
         if (code == ErrorCode::NoError) {
-            MGLOG_E("Recording Non-OpenGL error:\n%s", info->toString().c_str());
+            MGLOG_D("Recording Non-OpenGL error:\n%s", info->toString().c_str());
             m_nonGLErrors.push_back(MakeUnique<Error>(code, Move(info)));
         } else {
-            MGLOG_E("Recording OpenGL error (%s):\n%s",
+            MGLOG_D("Recording OpenGL error (%s):\n%s",
                     MG_Util::ConvertGLEnumToString(MG_Util::ConvertErrorCodeToGLEnum(code)).c_str(),
                     info->toString().c_str());
             // GL error semantics are sticky flags, not a queue (GL 3.3 core §2.5): with multiple

MobileGL/MG_State/GLState/FramebufferState/FramebufferObject.cpp

@@ -9,7 +9,18 @@
 #include "FramebufferObject.h"
 #include "MG_Util/Types.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 for the same reason as the VAO counter: it costs nothing, and a duplicate
+    // id would resurrect exactly the ABA this id exists to kill.
+    static std::atomic<Uint64> s_nextFramebufferLifetimeId{1};
+
+    Uint64 FramebufferObject::AllocateLifetimeId() {
+        return s_nextFramebufferLifetimeId.fetch_add(1, std::memory_order_relaxed);
+    }
+
     // FramebufferAttachmentObject
     FramebufferAttachmentObject::FramebufferAttachmentObject(
         const SharedPtr<MG_State::GLState::ITextureObject>& texture, TextureUploadTarget textureUploadTarget, Int level,

MobileGL/MG_State/GLState/FramebufferState/FramebufferObject.h

@@ -148,13 +148,25 @@ namespace MobileGL {
 
             Uint16 GetObjectVersion() const { return m_objectVersion; }
 
+            // Globally-unique, never-reused id for THIS object's lifetime - the same
+            // contract as VertexArrayObject::GetLifetimeId(), and needed for the same
+            // reason: neither the GL name nor the heap address can tell a
+            // deleted-and-recreated framebuffer from the original, and m_objectVersion
+            // starts at 0 for every new object, so a backend memo keyed on
+            // (pointer, version) alone would silently inherit the dead object's entry
+            // (see VkRenderPassManager's per-draw fast-path memo).
+            Uint64 GetLifetimeId() const { return m_lifetimeId; }
+
             Uint GetExternalIndex() const;
             Bool IsDefaultFramebuffer() const { return m_externalIndex == 0; }
 
         private:
+            static Uint64 AllocateLifetimeId();
+
             void BumpAttachmentVersion(FramebufferAttachmentType type);
 
             const Uint m_externalIndex = 0;
+            const Uint64 m_lifetimeId = AllocateLifetimeId();
             FramebufferAttachmentObjectArray m_attachmentObjects;
             FramebufferAttachmentVersionArray m_attachmentVersions;
 

MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.h

@@ -60,6 +60,8 @@ namespace MobileGL::MG_State::GLState {
             Uint externalIndex = 0; // logs only
             Vector<LinkShaderInput> shaders; // already stage-sorted
             SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> env;
+            // Startup configuration copied with the task, never read from worker code.
+            Bool enableSpirvValidation = false;
             // 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

MobileGL/MG_State/GLState/ProgramState/ProgramObject.cpp

@@ -255,7 +255,7 @@ namespace MobileGL::MG_State::GLState {
                 static_cast<SizeT>(offset) + write.byteOffsetInUniform + write.byteSize > uboSize) {
                 // Same verdict the live write path reaches for a uniform without backing
                 // storage: log and drop, rather than fault.
-                MGLOG_E("ProgramObject %u: buffered uniform write at location %u has no backing storage "
+                MGLOG_E_ONCE("ProgramObject %u: buffered uniform write at location %u has no backing storage "
                         "(offset=%u size=%u uboSize=%zu); dropping write",
                         m_externalIndex, write.location, offset, write.byteSize, uboSize);
                 continue;
@@ -393,6 +393,14 @@ namespace MobileGL::MG_State::GLState {
         return true;
     }
 
+    bool ProgramObject::AttachShaderWithPinnedLinkInput(const LinkedShaderRef& ref) {
+        if (!AttachShader(ref.shader)) {
+            return false;
+        }
+        m_pinnedLinkInputs[ref.shader.get()] = ref;
+        return true;
+    }
+
     SizeT ProgramObject::DetachShader(const SharedPtr<ShaderObject>& shader) {
         MGLOG_D("DetachShader called for shader %p from ProgramObject %u", shader.get(), m_externalIndex);
         if (!ShaderIsAttached(shader)) {
@@ -436,7 +444,7 @@ namespace MobileGL::MG_State::GLState {
         defaultFS->Compile(); // TODO: use a global default FS object.
         auto status = defaultFS->GetCompileStatus();
         if (!status) {
-            MGLOG_E("ProgramObject %u: Failed to compile default fragment shader. InfoLog:\n%s", m_externalIndex,
+            MGLOG_E_ONCE("ProgramObject %u: Failed to compile default fragment shader. InfoLog:\n%s", m_externalIndex,
                     defaultFS->GetInfoLog().c_str());
             return;
         }
@@ -475,6 +483,8 @@ namespace MobileGL::MG_State::GLState {
             AddDefaultFragmentShaderIfMissing();
         }
         if (m_shaders.empty()) {
+            // This IS the last link now, and it consumed nothing.
+            m_linkedShaderSnapshot.clear();
             m_artifacts.infoLog = "No shader objects are attached to program.";
             MGLOG_E("ProgramObject %u: Link failed - no shader objects attached.", m_externalIndex);
             return;
@@ -494,6 +504,7 @@ namespace MobileGL::MG_State::GLState {
         auto task = MakeShared<ProgramLinkTask>();
         task->in.externalIndex = m_externalIndex;
         task->in.env = MG_Util::ShaderTranspiler::GetCurrentCompileEnv();
+        task->in.enableSpirvValidation = MG_Config::Features.EnableSpirvValidation;
         task->in.explicitAttribLocations = m_explicitAttribLocations;
         task->in.explicitFragDataLocation = m_explicitFragDataLocation;
         task->in.explicitFragDataIndex = m_explicitFragDataIndex;
@@ -504,8 +515,19 @@ namespace MobileGL::MG_State::GLState {
         Vector<SharedPtr<ShaderCompileTask>> deps;
         deps.reserve(m_shaders.size());
         task->in.shaders.reserve(m_shaders.size());
+        m_linkedShaderSnapshot.clear();
+        m_linkedShaderSnapshot.reserve(m_shaders.size());
         for (const auto& shader : m_shaders) {
-            const SharedPtr<ShaderCompileTask>& node = shader->CompiledNodeForLink();
+            // A pipeline composite pins the (source, node) each stage program's LAST link
+            // consumed (AttachShaderWithPinnedLinkInput); an ordinary program takes the
+            // shader's current ones. Without the pin a post-link recompile would leak a
+            // shader the stage program never linked into the composite.
+            SharedPtr<const String> sourcePtr = shader->GetShaderSourcePtr();
+            SharedPtr<ShaderCompileTask> node = shader->CompiledNodeForLink();
+            if (const auto pinned = m_pinnedLinkInputs.find(shader.get()); pinned != m_pinnedLinkInputs.end()) {
+                sourcePtr = pinned->second.source;
+                node = pinned->second.node;
+            }
             if (node) {
                 // This link is now an observer of that node's result, and the ShaderObject is
                 // no longer the only route to it: without the marker, the ordinary
@@ -514,7 +536,10 @@ namespace MobileGL::MG_State::GLState {
                 node->MarkLinkReferenced();
                 if (!node->IsTerminal()) deps.push_back(node);
             }
-            task->in.shaders.push_back({shader->GetShaderStage(), shader->GetShaderSourcePtr(), node});
+            task->in.shaders.push_back({shader->GetShaderStage(), sourcePtr, node});
+            // What "as last linked" will mean for this program from now on - the pipeline
+            // composite cache rebuilds from exactly this set (GetProgramForDraw).
+            m_linkedShaderSnapshot.push_back({shader, sourcePtr, node});
         }
 
         // Phase B of the same link: SPIR-V generation, spirv-opt and the global-UBO routing

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

@@ -60,6 +60,26 @@ namespace MobileGL::MG_State::GLState {
 
         Vector<SharedPtr<ShaderObject>>& GetAttachedShaders();
         const Vector<SharedPtr<ShaderObject>>& GetAttachedShaders() const;
+
+        // One shader exactly as this program's last Link() consumed it: the object, the
+        // source snapshot, and the compile node taken at that link's enqueue. GL 4.6 7.3/7.4
+        // makes this triple - not the live attach list, not the shader's current compile -
+        // what a program pipeline stage executes ("as last linked"): glAttachShader and
+        // glCompileShader take effect only at the program's next link, yet neither moves
+        // m_linkVersion, so anything keyed on the link generation must consume this
+        // snapshot rather than re-read the live state.
+        struct LinkedShaderRef {
+            SharedPtr<ShaderObject> shader;
+            SharedPtr<const String> source;
+            SharedPtr<ShaderCompileTask> node;
+        };
+        // The last link's full input set; empty when this program has never linked (or its
+        // last link had no shaders attached). GL-thread-owned, rebuilt in Link()'s prologue.
+        const Vector<LinkedShaderRef>& GetLinkedShaderSnapshot() const { return m_linkedShaderSnapshot; }
+        // Pipeline-composite attach: AttachShader plus a pin that makes THIS program's
+        // Link() consume ref's (source, node) instead of the shader's current ones, so a
+        // post-link recompile of the stage program's shader cannot leak into the composite.
+        bool AttachShaderWithPinnedLinkInput(const LinkedShaderRef& ref);
         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.
@@ -786,6 +806,13 @@ namespace MobileGL::MG_State::GLState {
         // 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) {
+            // Equality bail-out like SetUniformBlockBinding's: the pipeline composite
+            // mirror replays every override each draw, and without this every replay
+            // would churn m_blockBindingVersion and rebuild whatever keys on it.
+            const auto it = Artifacts().shaderStorageBlockBinding.find(blockName);
+            if (it != Artifacts().shaderStorageBlockBinding.end() && it->second == static_cast<Int>(binding)) {
+                return;
+            }
             Artifacts().shaderStorageBlockBinding[blockName] = static_cast<Int>(binding);
             // Deliberately NOT m_backendStateVersion: Espryt's entry point never forces a
             // program build off this, and bumping that version would start doing so. The
@@ -819,6 +846,9 @@ namespace MobileGL::MG_State::GLState {
         // backend asks this exactly where it used to ask GetLinkStatus(), i.e. right before
         // it builds or draws with the program.
         Bool GetSpirvStatus() const { return Spirv().spirvStatus; }
+        // Copied from the link task that generated this program's SPIR-V. Backends use it for
+        // their final transforms, which must honor the same diagnostic setting as phase B.
+        Bool GetSpirvValidationEnabled() const { return Spirv().enableSpirvValidation; }
 
         // The linked glslang reflection itself, for the ONE consumer that needs resource
         // lists no typed getter above exposes: the GL program-interface query layer
@@ -985,6 +1015,7 @@ namespace MobileGL::MG_State::GLState {
         // cannot be lifted out of glslang's reflection instead.
         struct SpirvArtifacts {
             Vector<Vector<unsigned>> generatedSpirv;
+            Bool enableSpirvValidation = false;
             // Byte offset of each uniform location inside globalUboScratch, or
             // kInvalidUniformOffset. Sized maxUniformLocation + 1 by the routing pass.
             Vector<Uint> uniformOffsets;
@@ -1219,6 +1250,13 @@ namespace MobileGL::MG_State::GLState {
         // 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
+        // See GetLinkedShaderSnapshot. Holding the SharedPtrs here is deliberate: the
+        // "as last linked" set must survive detach-and-delete of its shaders (the
+        // glCreateShaderProgramv shape) until the next link replaces it.
+        Vector<LinkedShaderRef> m_linkedShaderSnapshot;
+        // See AttachShaderWithPinnedLinkInput. Populated only on pipeline composites,
+        // which never detach, so entries need no removal path. GL-thread-owned.
+        UnorderedMap<const ShaderObject*, LinkedShaderRef> m_pinnedLinkInputs;
 
         // Link INPUTS (all "take effect at the next link" per GL): glBindAttribLocation,
         // glBindFragDataLocation(Indexed), glTransformFeedbackVaryings, and the draw-buffer