[Feat] (DirectVulkan): relax fragment precision where the bound formats allow it

- WIP, parked: measures 80.9 -> 94.8 fps on Adreno 650 / MC 26.2 (same scene,
  device cooled to 38-40C), but is NOT validated. Desktop GLSL carries no
  precision qualifiers, so every fragment value reaches the driver as fp32 while
  Adreno runs fp16 at twice the rate.
- RelaxTextureDerivedPrecisionPass taints the values a fragment shader derives
  from built-in inputs and decorates everything else RelaxedPrecision. The
  taint direction matters: whitelisting outward from texture reads captures
  nothing, because MC multiplies every texel by an interpolated colour and a UBO
  value and one un-relaxed operand vetoes the expression - measured at 80.4 fps,
  i.e. no gain, both with and without varyings seeded. Precision-critical
  sources are few (gl_FragCoord cannot even hold a 3044-pixel x exactly), so
  tainting them and relaxing the rest is what actually pays.
- SPIR-V cannot see the bound formats - sampler2D yields vec4 whether the
  texture is RGBA8 or RGBA32F - so the decision is made per draw and passed in
  as a compile option, the same shape ExplicitLod0Sampling already uses.
  RelaxedFragmentPrecision is only requested when every sampled texture and
  every colour attachment is an 8-bit-or-less normalized format, where fp16's
  11-bit mantissa already carries the value exactly. Shaderpack HDR gbuffers,
  float data textures and 16-bit normalized targets therefore keep full
  precision, as do shaders that write gl_FragDepth or gl_SampleMask.
- LocalMultiStoreElim runs first: glslang emits function-local variables, and a
  load can never be relaxed, so without SSA promotion the analysis dies at the
  first temporary.
- WHY THIS IS PARKED: the retrace correctness gate never ran green. Every
  DirectVulkan retrace on Adreno 650 dies with DEVICE_LOST in
  UploadDirtyMipLevels on unmodified dev (pre-existing, device-gated), and on
  Adreno 830 - where the gate does pass on dev - minecraft-1.21.4-in-world times
  out at 900s with this change, which still needs explaining. Do not merge until
  that is understood and vanilla plus non-Photon shaderpack cases pass.
  (photon-v1.3b is broken on Adreno independently of this work.)
- The /sdcard/MG/exp_relaxed_precision_all and exp_no_relaxed_precision file
  toggles are development scaffolding for A/B measurement; they must go before
  this ships.

CMakeLists.txt

@@ -303,6 +303,13 @@ if (ANDROID)
     )
 endif()
 
+if (WIN32)
+    list(APPEND SOURCE_FILES
+        MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp
+        MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.cpp
+    )
+endif()
+
 set(MOBILEGL_LINK_LIBRARIES
         glslang::glslang
         spirv-cross-c
@@ -335,6 +342,14 @@ add_library(${CMAKE_PROJECT_NAME} SHARED
     ${SOURCE_FILES}
 )
 
+if (WIN32)
+    # The wgl* entry points are exported via .def (see the comment in wgl.def);
+    # only the shared library links it.
+    target_sources(${CMAKE_PROJECT_NAME} PRIVATE
+        MobileGL/MG_Impl/WGLImpl/Exporting/wgl.def
+    )
+endif()
+
 if (CMAKE_BUILD_TYPE STREQUAL "Debug")
     set_target_properties(${CMAKE_PROJECT_NAME} PROPERTIES
         C_VISIBILITY_PRESET default
@@ -377,6 +392,18 @@ if(UNIX AND NOT APPLE AND NOT ANDROID)
     endforeach()
 endif()
 
+if(WIN32)
+    # Drop-in for the classic GL loader path: a copy named opengl32.dll placed
+    # next to a host executable is what LoadLibrary("opengl32.dll") and gdi32's
+    # pixel-format forwarding will resolve.
+    add_custom_command(TARGET ${CMAKE_PROJECT_NAME} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                "$<TARGET_FILE:${CMAKE_PROJECT_NAME}>"
+                "$<TARGET_FILE_DIR:${CMAKE_PROJECT_NAME}>/opengl32.dll"
+        COMMENT "Creating opengl32.dll drop-in copy"
+    )
+endif()
+
 if(NOT ANDROID)
     add_library(${CMAKE_PROJECT_NAME}_s STATIC 
         ${SOURCE_FILES}
@@ -428,8 +455,21 @@ if (ANDROID)
 endif()
 
 if (APPLE AND NOT MOBILEGL_IOS)
+    # MobileGL statically embeds glslang, SPIRV-Tools, and SPIRV-Cross. When
+    # this dylib is injected with DYLD_INSERT_LIBRARIES, exporting those C++
+    # symbols interposes incompatible copies embedded by host libraries such
+    # as shaderc. Keep only the public GL/EGL/CGL loader surface globally
+    # visible; GetProcAddress can still return pointers to hidden internals.
+    set(MOBILEGL_MACOS_EXPORTED_SYMBOLS
+        "${CMAKE_CURRENT_SOURCE_DIR}/MobileGL/MG_Impl/DyldInterpose/ExportedSymbols.txt")
+    target_link_options(${CMAKE_PROJECT_NAME} PRIVATE
+        "LINKER:-exported_symbols_list,${MOBILEGL_MACOS_EXPORTED_SYMBOLS}")
+    set_property(TARGET ${CMAKE_PROJECT_NAME} APPEND PROPERTY
+        LINK_DEPENDS "${MOBILEGL_MACOS_EXPORTED_SYMBOLS}")
+
     target_link_libraries(${CMAKE_PROJECT_NAME} PUBLIC
         "-framework Cocoa"
+        "-framework CoreVideo"
         "-framework QuartzCore"
         "-framework Foundation"
         "-framework OpenGL"
@@ -437,6 +477,7 @@ if (APPLE AND NOT MOBILEGL_IOS)
     if(TARGET ${CMAKE_PROJECT_NAME}_s)
         target_link_libraries(${CMAKE_PROJECT_NAME}_s PUBLIC
             "-framework Cocoa"
+            "-framework CoreVideo"
             "-framework QuartzCore"
             "-framework Foundation"
             "-framework OpenGL"

MobileGL/Defines.h

@@ -34,6 +34,7 @@
 #define MOBILEGL_EGL_API MOBILEGL_API
 #define MOBILEGL_CGL_API MOBILEGL_API
 #define MOBILEGL_NSOPENGL_API MOBILEGL_API
+#define MOBILEGL_WGL_API MOBILEGL_API
 
 // ====================== MobileGL configurations ======================= //
 #ifndef MOBILEGL_LOG_ACTIVE_LEVEL

MobileGL/GlobalObjects.cpp

@@ -14,7 +14,13 @@ namespace MobileGL {
     } // namespace MG_Config
 
     namespace MG_Backend {
-        UniquePtr<BackendObject> pActiveBackendObject;
+        // Leak-at-exit storage: the UniquePtr itself lives on the heap and is
+        // never destroyed by the runtime, so process exit runs no backend
+        // destructors (static destruction order across TUs is undefined).
+        // Deterministic teardown happens inside the EGL lifecycle instead:
+        // the last eglTerminate calls MobileGL::Destroy(), which .reset()s
+        // these singletons while the process is still healthy.
+        UniquePtr<BackendObject>& pActiveBackendObject = *new UniquePtr<BackendObject>();
         GlobalBackendFunctionsTable gBackendFunctionsTable;
     } // namespace MG_Backend
 } // namespace MobileGL

MobileGL/Init.cpp

@@ -9,14 +9,25 @@
 #include "Init.h"
 #include "Config.h"
 #include <MG_Backend/BackendObjects.h>
+#include <MG_Backend/DirectVulkan/DirectVulkan.h>
 #include <MG_State/GLState/Core.h>
 #include <MG_State/EGLState/Core.h>
 #include <MG_Impl/GLImpl/Texture/ProxyTexture.h>
 #include <MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h>
+#include <MG_Impl/GLImpl/Sync/GL_Sync.h>
+
+#include <atomic>
+#include <mutex>
 
 namespace MobileGL {
     namespace {
-        Bool g_isInitialized = false;
+        std::atomic<Bool> g_isInitialized = false;
+        thread_local Bool tl_initializing = false;
+
+        std::mutex& InitMutex() {
+            static std::mutex mutex;
+            return mutex;
+        }
 
         void DestroyImpl(Bool logLifecycle) {
             if (!g_isInitialized) {
@@ -27,6 +38,12 @@ namespace MobileGL {
                 MGLOG_I("MobileGL closing...");
             }
             glslang::FinalizeProcess();
+            // GL syncs die with their contexts, and every context is gone by the
+            // time full teardown runs: drain the live-sync registry while the
+            // backend function table can still release the backend handles (and
+            // before a re-initialized library could pair them with the wrong
+            // backend's DeleteSync).
+            MG_Impl::GLImpl::DestroyAllSyncObjects();
             MG_Backend::pActiveBackendObject.reset();
             MG_State::pGLContext.reset();
             MG_State::pEGLContext.reset();
@@ -64,40 +81,37 @@ namespace MobileGL {
         MGLOG_I("MobileGL initialized");
     }
 
+    void EnsureInitialized() {
+        if (g_isInitialized.load(std::memory_order_acquire)) {
+            return;
+        }
+        // Re-entrant call while this thread is already inside Initialize()
+        // (e.g. an init step routing back through a public entry point).
+        if (tl_initializing) {
+            return;
+        }
+        const std::lock_guard<std::mutex> lock(InitMutex());
+        if (g_isInitialized.load(std::memory_order_acquire)) {
+            return;
+        }
+        tl_initializing = true;
+        Initialize();
+        tl_initializing = false;
+    }
+
     void Destroy() {
         DestroyImpl(true);
     }
 
-#if defined(__linux__) || defined(__APPLE__)
-    __attribute__((constructor)) static void AutoInit() {
-        Initialize();
-    }
-
-    __attribute__((destructor)) static void AutoDestroy() {
-        if (MG_Config::Features.TraceSkipAutodestroy) {
-            return;
-        }
-#if defined(__APPLE__)
-        // macOS injected dylibs can run destructors after logging/backend static state is already torn down.
-        return;
-#else
-        DestroyImpl(false);
-#endif
-    }
-#endif
-
-#ifdef _WIN32
-    BOOL WINAPI DllMain(HMODULE hModule, DWORD ul_reason_for_call, LPVOID lpReserved) {
-        switch (ul_reason_for_call) {
-        case DLL_PROCESS_ATTACH:
-            Initialize();
-            break;
-
-        case DLL_PROCESS_DETACH:
-            Destroy();
-            break;
-        }
-        return TRUE;
-    }
-#endif
+    // MobileGL's lifecycle is owned entirely by the host-API layers
+    // (EGL/WGL/CGL): initialization happens lazily on the first entry point
+    // via EnsureInitialized(), and full teardown happens deterministically
+    // when the last EGL display is terminated with nothing current (EGLImpl
+    // calls Destroy()). There is intentionally no backend-initializing static
+    // constructor, no static destructor, and no DllMain: the global singletons
+    // use leak-at-exit storage (see GlobalObjects.cpp), so a process that exits
+    // without eglTerminate simply leaks them to the OS instead of running
+    // backend destructors during static teardown. macOS has a lightweight
+    // dyld constructor that installs NSOpenGL dispatch hooks only; full backend
+    // initialization still enters here from the first hooked CGL context.
 } // namespace MobileGL

MobileGL/Init.h

@@ -11,6 +11,13 @@
 
 namespace MobileGL {
     void Initialize();
+    // Thread-safe, idempotent, and re-entrant wrapper around Initialize().
+    // Host layers (EGL/WGL/CGL entry points) call this lazily on first use so
+    // full backend initialization never depends on ELF/DLL static constructors,
+    // and so a fresh init can follow a full Destroy() (e.g. after the last
+    // eglTerminate). The macOS dyld bootstrap installs only lightweight
+    // NSOpenGL method hooks.
+    void EnsureInitialized();
     void Destroy();
 
     namespace MG_Util::Debug {

MobileGL/MG_Backend/BackendObject.h

@@ -313,7 +313,8 @@ namespace MobileGL {
             Android,
             X11,
             MetalLayer,
-            // TODO: Wayland, Windows, etc.
+            Win32, // Handle is an HWND
+            // TODO: Wayland, etc.
             WindowBackendCount,
             Unknown = -1
         };

MobileGL/MG_Backend/BackendObjects.h

@@ -13,6 +13,6 @@
 #include "DirectVulkan/BackendObject_DirectVulkan.h"
 
 namespace MobileGL::MG_Backend {
-    extern UniquePtr<BackendObject> pActiveBackendObject;
+    extern UniquePtr<BackendObject>& pActiveBackendObject;
     extern GlobalBackendFunctionsTable gBackendFunctionsTable;
 } // namespace MobileGL::MG_Backend

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -701,9 +701,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
         if ((handle.Backend != WindowBackend::Android &&
              handle.Backend != WindowBackend::X11 &&
-             handle.Backend != WindowBackend::MetalLayer) ||
+             handle.Backend != WindowBackend::MetalLayer &&
+             handle.Backend != WindowBackend::Win32) ||
             !handle.Handle) {
-            MGLOG_E("DirectGLES backend only supports Android, X11, and CAMetalLayer native windows");
+            MGLOG_E("DirectGLES backend only supports Android, X11, CAMetalLayer, and Win32 native windows");
             return false;
         }
 

MobileGL/MG_Backend/DirectGLES/DirectGLES.cpp

@@ -4223,9 +4223,23 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
         if (!ChooseConfigForSurface(surfaceBit, g_Config, window)) return false;
 
-        const EGLint contextAttribs[] = {EGL_CONTEXT_CLIENT_VERSION, 3, EGL_NONE};
-
+        // Negotiate the highest ES 3.x context. Version-strict EGL implementations
+        // (ANGLE) return exactly the requested minor, and a bare CLIENT_VERSION 3
+        // request yields a 3.0 context that lacks the 3.1/3.2 texture targets the
+        // capability probes exercise; mobile drivers ignore the minor and hand out
+        // their maximum either way.
+        for (const EGLint minorVersion : {2, 1, 0}) {
+            const EGLint contextAttribs[] = {EGL_CONTEXT_MAJOR_VERSION, 3,
+                                             EGL_CONTEXT_MINOR_VERSION, minorVersion,
+                                             EGL_NONE};
             g_Context = g_EGLFuncs.eglCreateContext(g_Display, g_Config, EGL_NO_CONTEXT, contextAttribs);
+            if (g_Context != EGL_NO_CONTEXT) {
+                return true;
+            }
+        }
+
+        const EGLint legacyContextAttribs[] = {EGL_CONTEXT_CLIENT_VERSION, 3, EGL_NONE};
+        g_Context = g_EGLFuncs.eglCreateContext(g_Display, g_Config, EGL_NO_CONTEXT, legacyContextAttribs);
         return g_Context != EGL_NO_CONTEXT;
     }
 

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -411,8 +411,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         if (!handle.Handle || (handle.Backend != WindowBackend::Android &&
                                handle.Backend != WindowBackend::X11 &&
-                               handle.Backend != WindowBackend::MetalLayer)) {
-            MGLOG_E("DirectVulkan backend only supports Android, X11, and CAMetalLayer native windows");
+                               handle.Backend != WindowBackend::MetalLayer &&
+                               handle.Backend != WindowBackend::Win32)) {
+            MGLOG_E("DirectVulkan backend only supports Android, X11, CAMetalLayer, and Win32 native windows");
             return false;
         }
 
@@ -468,6 +469,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // treat them as signaled/available with zero results from here on.
         BumpRendererGeneration();
         pVulkanRenderer.reset();
+        // The reflection cache is file-scope, not renderer-owned; without this the
+        // deleted programs' reflection strings survive full context teardown.
+        ClearProgramResourceCaches();
         BackendObject::ReleaseEGLResources();
     }
 
@@ -477,6 +481,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // treat them as signaled/available with zero results from here on.
         BumpRendererGeneration();
         pVulkanRenderer.reset();
+        // The reflection cache is file-scope, not renderer-owned; without this the
+        // deleted programs' reflection strings survive full context teardown.
+        ClearProgramResourceCaches();
     }
 
     const RendererInfo& BackendObject_DirectVulkan::GetRendererInfo() const {

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp

@@ -20,7 +20,8 @@
 #include <spirv_reflect.h>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
-    UniquePtr<VulkanRenderer> pVulkanRenderer = nullptr;
+    // Leak-at-exit storage; see GlobalObjects.cpp.
+    UniquePtr<VulkanRenderer>& pVulkanRenderer = *new UniquePtr<VulkanRenderer>();
 
     namespace {
         // Generation of the live VulkanRenderer instance, mirroring
@@ -60,6 +61,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         };
 
         struct ProgramResourceCache {
+            // Lifetime id of the program the cached reflection belongs to. GL names are
+            // recycled (IndexGenerator hands freed indices straight back), and a
+            // recreated program's backendStateVersion restarts at the same small values,
+            // so the version alone can collide; the never-reused lifetime id makes the
+            // slot's ownership unambiguous.
+            Uint64 programLifetimeId = 0;
             Uint32 backendStateVersion = 0;
             Vector<StorageBlockResource> storageBlocks;
             Vector<BufferVariableResource> bufferVariables;
@@ -81,6 +88,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint32 baseInstance = 0;
         };
 
+        // Keyed by GL program name so the freed-name reuse in IndexGenerator bounds the
+        // map at the peak-simultaneous-program high-water mark; each slot's ownership is
+        // checked against the program's lifetime id before it is served (see
+        // GetProgramResourceCache). Cleared wholesale at EGL teardown via
+        // ClearProgramResourceCaches.
         UnorderedMap<GLuint, ProgramResourceCache> g_programResourceCaches;
 
         void ClearReadPixelsOutput(GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels) {
@@ -141,13 +153,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         ProgramResourceCache& GetProgramResourceCache(const MG_State::GLState::ProgramObject& program) {
             auto& cache = g_programResourceCaches[program.GetExternalIndex()];
+            const Uint64 programLifetimeId = program.GetLifetimeId();
             const Uint32 backendStateVersion = program.GetBackendStateVersion();
-            if (cache.backendStateVersion == backendStateVersion &&
+            // 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())) {
                 return cache;
             }
 
             cache = {};
+            cache.programLifetimeId = programLifetimeId;
             cache.backendStateVersion = backendStateVersion;
 
             Vector<SpvReflectShaderModule> modules;
@@ -365,6 +383,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     } // namespace
 
+    void ClearProgramResourceCaches() {
+        // Called from EGL teardown while the backend's m_eglStateMutex is held; GL
+        // calls are serialized in this codebase (contexts migrate threads but never
+        // run concurrently), so no other thread can be inside the unsynchronized map.
+        // Live programs in another context self-heal: their entry rebuilds from the
+        // retained generated SPIR-V on the next resource query.
+        g_programResourceCaches.clear();
+    }
+
     GLuint GetShaderStorageBlockIndex(const MG_State::GLState::ProgramObject& program, const String& name) {
         auto& cache = GetProgramResourceCache(program);
         const auto it = std::find_if(cache.storageBlocks.begin(), cache.storageBlocks.end(),

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.h

@@ -12,7 +12,7 @@
 #include "Renderer/VulkanRenderer.h"
 
 namespace MobileGL::MG_Backend::DirectVulkan {
-    extern UniquePtr<VulkanRenderer> pVulkanRenderer;
+    extern UniquePtr<VulkanRenderer>& pVulkanRenderer;
 
     // Generation of the live VulkanRenderer instance, mirroring DirectGLES's
     // g_syncContextGeneration. BackendObject_DirectVulkan bumps it wherever
@@ -23,6 +23,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     Uint64 GetRendererGeneration();
     void BumpRendererGeneration();
 
+    // Drops every cached program-resource reflection entry (CPU-side strings/vectors
+    // only, no Vulkan handles). Called at EGL teardown next to the renderer reset;
+    // safe because GL calls are serialized in this codebase, and any still-live
+    // program rebuilds its entry from the retained generated SPIR-V on demand.
+    void ClearProgramResourceCaches();
+
     void ClearBufferfi(GLenum buffer, GLint drawbuffer, GLfloat depth, GLint stencil);
     void ClearBufferfv(GLenum buffer, GLint drawbuffer, const GLfloat* value);
     void ClearBufferuiv(GLenum buffer, GLint drawbuffer, const GLuint* value);

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.cpp

@@ -150,12 +150,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool FrameContext::TransitionToPresent(VkImage image, VkImageLayout oldLayout, VkImageLayout presentLayout) {
         auto& frame = GetCurrent();
-        if (frame.hasCommandBufferRecorded || frame.isCommandRecording || oldLayout == presentLayout ||
-            oldLayout == VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR) {
+        if (oldLayout == presentLayout || oldLayout == VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR) {
             return false;
         }
 
-        auto& commandBuffer = BeginCommandRecording();
+        // The barrier belongs in the frame's own recording. Bailing out because
+        // something was already recorded (the previous behaviour) dropped the
+        // transition entirely for every frame that never ran a default-framebuffer
+        // render pass - the only other thing that carries the image to
+        // PRESENT_SRC_KHR, via that pass's finalLayout - so the swapchain image was
+        // handed to the WSI still in the layout it was acquired in.
+        // A closed-but-unsubmitted buffer can only come from a submit that already
+        // failed (SubmitPendingCommandBuffer leaves the flag set on error), and
+        // appending to it is illegal while reopening would reset the frame's own
+        // 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");
+            return false;
+        }
+
+        // Reopening a recording here would vkResetCommandBuffer this frame's own
+        // commands away, so append to the open one and let the caller close it.
+        const Bool openedRecording = !frame.isCommandRecording;
+        VkCommandBuffer commandBuffer = openedRecording ? BeginCommandRecording() : frame.commandBuffer;
 
         VkImageMemoryBarrier presentBarrier{};
         presentBarrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
@@ -174,7 +192,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, 0,
                              nullptr, 0, nullptr, 1, &presentBarrier);
 
+        if (openedRecording) {
             EndCommandRecording();
+        }
         return true;
     }
 
@@ -227,12 +247,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         result = vkAcquireNextImageKHR(device, swapchain, timeout, frame.imageAvailableSemaphore, acquireFence,
                                        &outImageIndex);
-        if (result != VK_SUCCESS) {
+        // VK_SUBOPTIMAL_KHR is a success code: an image *was* acquired and
+        // imageAvailableSemaphore *will* be signaled. Bailing out on it skipped both
+        // the consumed-flag reset (leaving a stale "already consumed", so the next
+        // submit never waited on the pending signal) and the fence reset (leaving
+        // the slot's fence signaled for the next submit to reuse). Only a genuine
+        // failure - VK_ERROR_OUT_OF_DATE_KHR and friends, where nothing is acquired
+        // and nothing is signaled - skips the bookkeeping.
+        if (result != VK_SUCCESS && result != VK_SUBOPTIMAL_KHR) {
             return result;
         }
 
         frame.imageAvailableSemaphoreConsumed = false;
-        return vkResetFences(device, 1, &frame.imageInFlightFence);
+        const VkResult resetResult = vkResetFences(device, 1, &frame.imageInFlightFence);
+        // Hand the acquire's own code back so the caller can schedule a rebuild.
+        return resetResult == VK_SUCCESS ? result : resetResult;
     }
 
     Uint32 FrameContext::GetCurrentFrameIndex() const {
@@ -264,7 +293,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (result != VK_SUCCESS) {
             return result;
         }
-        frame.retiredCommandBuffers.push_back(frame.commandBuffer);
+        // lastSubmitIndex was just written by the renderer for the submission
+        // that carried this command buffer.
+        frame.retiredCommandBuffers.push_back({frame.commandBuffer, frame.lastSubmitIndex});
         frame.commandBuffer = replacement;
         return VK_SUCCESS;
     }
@@ -274,12 +305,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return;
         }
         if (m_device != VK_NULL_HANDLE && m_commandPool != VK_NULL_HANDLE) {
-            vkFreeCommandBuffers(m_device, m_commandPool, static_cast<Uint32>(frame.retiredCommandBuffers.size()),
-                                 frame.retiredCommandBuffers.data());
+            for (const auto& retired : frame.retiredCommandBuffers) {
+                vkFreeCommandBuffers(m_device, m_commandPool, 1, &retired.commandBuffer);
+            }
         }
         frame.retiredCommandBuffers.clear();
     }
 
+    void FrameContext::FreeRetiredCommandBuffersCompletedUpTo(Uint64 completedSubmitIndex) {
+        if (m_device == VK_NULL_HANDLE || m_commandPool == VK_NULL_HANDLE) {
+            return;
+        }
+        for (auto& frame : m_frames) {
+            // Retired buffers are appended in submit order, so the completed
+            // ones form a prefix.
+            SizeT completedCount = 0;
+            while (completedCount < frame.retiredCommandBuffers.size() &&
+                   frame.retiredCommandBuffers[completedCount].submitIndex <= completedSubmitIndex) {
+                vkFreeCommandBuffers(m_device, m_commandPool, 1,
+                                     &frame.retiredCommandBuffers[completedCount].commandBuffer);
+                ++completedCount;
+            }
+            if (completedCount > 0) {
+                frame.retiredCommandBuffers.erase(frame.retiredCommandBuffers.begin(),
+                                                  frame.retiredCommandBuffers.begin() + completedCount);
+            }
+        }
+    }
+
+    void FrameContext::FreeAllRetiredCommandBuffers() {
+        for (auto& frame : m_frames) {
+            FreeRetiredCommandBuffers(frame);
+        }
+    }
+
     void FrameContext::AssertValidFrameIndex(Uint32 frameIndex) const {
         MOBILEGL_ASSERT(frameIndex < m_frames.size(), "FrameContext index out of range");
     }

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.h

@@ -40,6 +40,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkPresentInfoKHR presentInfo{VK_STRUCTURE_TYPE_PRESENT_INFO_KHR};
         };
 
+        // A command buffer submitted mid-frame (FlushPendingCommands), tagged
+        // with the submit-tracker index it was submitted under so it can be
+        // freed as soon as that submission is observed complete - without
+        // waiting for the slot's fence to be waited again (present-less flush
+        // loops never wait it).
+        struct RetiredCommandBuffer {
+            VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
+            Uint64 submitIndex = 0;
+        };
+
         struct FrameData {
             VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
             VkSemaphore imageAvailableSemaphore = VK_NULL_HANDLE;
@@ -47,10 +57,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool isCommandRecording = false;
             Bool hasCommandBufferRecorded = false;
             Bool imageAvailableSemaphoreConsumed = false;
-            // Command buffers submitted mid-frame (FlushPendingCommands) whose
-            // execution is only known complete once this slot's fence has been
-            // waited again; freed at that point.
-            Vector<VkCommandBuffer> retiredCommandBuffers;
+            // Command buffers submitted mid-frame (FlushPendingCommands),
+            // appended in submit order; freed once their submission is known
+            // complete (fence wait or completion poll).
+            Vector<RetiredCommandBuffer> retiredCommandBuffers;
             // Submit-tracker index of this slot's most recent queue submission
             // (written by the renderer at submit time).
             Uint64 lastSubmitIndex = 0;
@@ -79,9 +89,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Parks the current (already ended and submitted) command buffer on the
         // slot's retired list and installs a freshly allocated one, so recording
         // can restart while the submitted buffer is still executing. Retired
-        // buffers are freed after the slot's fence is next waited.
+        // buffers are freed after the slot's fence is next waited, or as soon
+        // as their submission is observed complete.
         VkResult RetireCurrentCommandBuffer();
 
+        // Frees every retired command buffer whose tagged submission index is
+        // known complete. Driven by the renderer's submit tracker on completion
+        // events (fence waits and non-blocking polls), so present-less flush
+        // loops reclaim their buffers without any extra wait.
+        void FreeRetiredCommandBuffersCompletedUpTo(Uint64 completedSubmitIndex);
+        // Frees every slot's retired command buffers. Only valid when the
+        // caller has proven every queue submission complete.
+        void FreeAllRetiredCommandBuffers();
+
         Uint32 GetCurrentFrameIndex() const;
         Uint32 GetFrameCount() const;
 

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -8,6 +8,7 @@
 
 #include "PipelineFactory.h"
 
+#include <algorithm>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
     static const char* PrimitiveTopologyToString(VkPrimitiveTopology topology) {
@@ -243,23 +244,108 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const HashType hash = ComputeHash(payload);
         auto it = m_cache.find(hash);
         if (it != m_cache.end()) {
-            return it->second;
+            it->second.lastUsedFrame = m_frameCounter;
+            return it->second.pipeline;
         }
 
         VkPipeline pipeline = CreatePipeline(payload);
-        m_cache.emplace(hash, pipeline);
+        m_cache.emplace(hash, PipelineCacheEntry{pipeline, payload.programHash, payload.renderPass,
+                                                 m_frameCounter});
         return pipeline;
     }
 
     void PipelineFactory::DestroyAll() {
         for (auto& pair : m_cache) {
-            if (pair.second != VK_NULL_HANDLE) {
-                vkDestroyPipeline(m_device, pair.second, nullptr);
+            if (pair.second.pipeline != VK_NULL_HANDLE) {
+                vkDestroyPipeline(m_device, pair.second.pipeline, nullptr);
             }
         }
         m_cache.clear();
     }
 
+    Uint32 PipelineFactory::OnFrameBoundary() {
+        ++m_frameCounter;
+
+        // Sweep cadence and retire age mirror VkRenderPassManager::OnPresent: an entry
+        // idle for more than kRetireAgeFrames frame boundaries cannot be referenced by
+        // any in-flight command buffer (frames-in-flight <= MOBILEGL_MAGMA_FRAMESINFLIGHT),
+        // so immediate vkDestroyPipeline is safe. The caller must drop its "last
+        // pipeline" memo when this returns non-zero: the memo can return a cached
+        // handle without touching this cache, so an evicted pipeline may still be
+        // memoized (present-less flush loops never reset the memo per frame).
+        constexpr Uint64 kSweepInterval = 256;
+        constexpr Uint64 kRetireAgeFrames = 1024;
+        if ((m_frameCounter % kSweepInterval) != 0) {
+            return 0;
+        }
+
+        Uint32 evicted = 0;
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (m_frameCounter - it->second.lastUsedFrame > kRetireAgeFrames) {
+                if (it->second.pipeline != VK_NULL_HANDLE) {
+                    vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
+                }
+                it = m_cache.erase(it);
+                ++evicted;
+            } else {
+                ++it;
+            }
+        }
+        if (evicted > 0) {
+            MGLOG_D("PipelineFactory::OnFrameBoundary: evicted %u idle pipelines (%zu remain)", evicted,
+                    m_cache.size());
+        }
+        return evicted;
+    }
+
+    Uint32 PipelineFactory::EvictByRenderPasses(const Vector<VkRenderPass>& renderPasses) {
+        if (renderPasses.empty() || m_cache.empty()) {
+            return 0;
+        }
+        // Sorted-batch membership test keeps a mass eviction (shader-pack switch,
+        // dimension exit) at one O(cache * log batch) scan instead of one full scan
+        // per dying pass.
+        Vector<VkRenderPass> sortedPasses = renderPasses;
+        std::sort(sortedPasses.begin(), sortedPasses.end());
+        Uint32 evicted = 0;
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (std::binary_search(sortedPasses.begin(), sortedPasses.end(), it->second.renderPass)) {
+                if (it->second.pipeline != VK_NULL_HANDLE) {
+                    vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
+                }
+                it = m_cache.erase(it);
+                ++evicted;
+            } else {
+                ++it;
+            }
+        }
+        if (evicted > 0) {
+            MGLOG_D("PipelineFactory::EvictByRenderPasses: evicted %u pipelines for %zu destroyed render passes",
+                    evicted, sortedPasses.size());
+        }
+        return evicted;
+    }
+
+    Uint32 PipelineFactory::EvictByProgramHash(HashType programHash) {
+        Uint32 evicted = 0;
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (it->second.programHash == programHash) {
+                if (it->second.pipeline != VK_NULL_HANDLE) {
+                    vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
+                }
+                it = m_cache.erase(it);
+                ++evicted;
+            } else {
+                ++it;
+            }
+        }
+        if (evicted > 0) {
+            MGLOG_D("PipelineFactory::EvictByProgramHash: evicted %u pipelines for program hash 0x%llx",
+                    evicted, static_cast<unsigned long long>(programHash));
+        }
+        return evicted;
+    }
+
     VkPipeline PipelineFactory::CreatePipeline(const PipelineCreatePayload& payload) const {
         MOBILEGL_ASSERT(payload.stages != nullptr && !payload.stages->empty(), "PipelineFactory: stages are empty");
         MOBILEGL_ASSERT(payload.vertexInputState != nullptr, "PipelineFactory: vertexInputState is null");

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -65,6 +65,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkPipeline GetOrCreatePipeline(const PipelineCreatePayload& payload);
         void DestroyAll();
 
+        // Frame boundary hook: ages the pipeline cache and destroys long-unused entries
+        // (their command buffers retired many frames ago), mirroring
+        // VkRenderPassManager::OnPresent's sweep. Returns the number of pipelines
+        // destroyed so the caller can drop any memoized VkPipeline handle.
+        Uint32 OnFrameBoundary();
+        // Destroys every cached pipeline hashed on one of `renderPasses`. Only safe
+        // when the caller guarantees GPU idleness for them - the render-pass manager
+        // calls this (via the renderer) for passes its own >1024-boundary-idle sweep
+        // just evicted, and a pipeline hashed on those handles is only ever bound by
+        // draws that also hit the render-pass entries. Also closes the handle-recycling
+        // hazard: a recycled VkRenderPass value must never serve a stale pipeline.
+        // Batched: one cache scan regardless of how many passes died in the sweep.
+        // Returns the number destroyed (callers invalidate memos when non-zero).
+        Uint32 EvictByRenderPasses(const Vector<VkRenderPass>& renderPasses);
+        // Destroys every cached pipeline built from the program with content hash
+        // `programHash`. Called from the ProgramFactory eviction path, which proves the
+        // same >1024-boundary idleness (the program's pipelines are only bound by draws
+        // that stamp its factory entry). Returns the number destroyed.
+        Uint32 EvictByProgramHash(HashType programHash);
+
         // Driver quirk: suppress depth writes on accumulation-blended pipelines. Multi-pass
         // depth-equality rendering (a blended prepass writes depth that later passes re-test
         // with an equality-inclusive compare on the re-rasterized geometry) requires
@@ -87,12 +107,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         static Bool ShouldSuppressDepthWrite(const PipelineCreatePayload& payload);
 
     private:
+        struct PipelineCacheEntry {
+            VkPipeline pipeline = VK_NULL_HANDLE;
+            // The hashed inputs the eviction paths key on: programHash ties the entry to
+            // its ProgramFactory entry, renderPass records the exact handle the hash
+            // folded in (the hash is one-way, so targeted eviction needs them verbatim).
+            HashType programHash = 0;
+            VkRenderPass renderPass = VK_NULL_HANDLE;
+            // Frame-boundary counter value of the last GetOrCreatePipeline hit; drives
+            // cache eviction (see OnFrameBoundary).
+            Uint64 lastUsedFrame = 0;
+        };
+
         VkPipeline CreatePipeline(const PipelineCreatePayload& payload) const;
 
         VkDevice m_device = VK_NULL_HANDLE;
         const VulkanRendererConfig& m_config;
         VkPipelineCache m_pipelineCache = VK_NULL_HANDLE;
-        UnorderedMap<HashType, VkPipeline> m_cache;
+        UnorderedMap<HashType, PipelineCacheEntry> m_cache;
+        // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+        Uint64 m_frameCounter = 0;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
         static inline Bool s_suppressBlendedDepthWrite = false;
     };

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -12,7 +12,10 @@
 #include "MG_Util/ShaderTranspiler/ShaderCompiler.h"
 #include "MG_Util/ShaderTranspiler/SpvcSession.h"
 #include "MG_Util/ShaderTranspiler/Types.h"
+#include <cmath>
+#include <cstdio>
 #include <cstring>
+#include <unordered_set>
 #include <spirv-tools/libspirv.h>
 #include <spirv-tools/optimizer.hpp>
 #include <source/opt/build_module.h>
@@ -923,11 +926,610 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             ProgramFactory::CompileOptionFlags m_transformFlags;
         };
 
+        // Adreno 650 (driver 512.502) faults the GPU on an implicit-LOD sample of a full-screen
+        // colour render target: the texture unit's derivative path reads outside the image's
+        // allocation even though the sampler clamps LOD to 0 and the mapping is 1:1. MobileGL's
+        // own default-framebuffer blit shader works around it with textureLod, but an
+        // application's shader (Minecraft's blit.fsh is `texture(InSampler, texCoord)`) cannot be
+        // edited - so rewrite the sample at the SPIR-V level instead.
+        //
+        // The rewrite is only requested for draws whose every sampler binding is clamped to one
+        // mip level, where explicit LOD 0 is exactly what the implicit form must already produce:
+        // lambda' = clamp(lambda + bias, minLod, maxLod) with minLod = maxLod = 0. Bias and MinLod
+        // operands are therefore dropped rather than translated.
+        class ForceExplicitLod0SamplePass final : public spvtools::opt::Pass {
+        public:
+            const char* name() const override { return "force-explicit-lod0-sample"; }
+
+            Status Process() override {
+                Bool isFragment = false;
+                for (auto& entryPoint : get_module()->entry_points()) {
+                    if (entryPoint.opcode() != spv::Op::OpEntryPoint) continue;
+                    if (static_cast<spv::ExecutionModel>(entryPoint.GetSingleWordInOperand(0)) ==
+                        spv::ExecutionModel::Fragment) {
+                        isFragment = true;
+                        break;
+                    }
+                }
+                if (!isFragment) return Status::SuccessWithoutChange;
+
+                // Plan first, mutate second. Materializing the LOD constant is itself a module
+                // change, so it must not happen unless at least one rewrite is going to follow -
+                // otherwise the pass would grow the binary while reporting SuccessWithoutChange.
+                Vector<RewritePlan> plans;
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            RewritePlan plan{};
+                            if (PlanRewrite(&inst, plan)) plans.push_back(Move(plan));
+                        }
+                    }
+                }
+                if (plans.empty()) return Status::SuccessWithoutChange;
+
+                const Uint32 zeroId = GetFloatZeroId();
+                if (zeroId == 0) return Status::SuccessWithoutChange;
+
+                for (auto& plan : plans) {
+                    plan.operands.push_back({SPV_OPERAND_TYPE_ID, {zeroId}});
+                    for (auto& operand : plan.trailingOperands) {
+                        plan.operands.push_back(operand);
+                    }
+                    plan.instruction->SetOpcode(plan.opcode);
+                    plan.instruction->SetInOperands(Move(plan.operands));
+                }
+                // Opcodes and operand lists changed underneath every cached analysis.
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+        private:
+            struct RewritePlan {
+                spvtools::opt::Instruction* instruction = nullptr;
+                spv::Op opcode = spv::Op::OpNop;
+                // Everything up to and including the Image Operands mask; the Lod id and the
+                // trailing operand values are appended once the constant exists.
+                Vector<spvtools::opt::Operand> operands;
+                Vector<spvtools::opt::Operand> trailingOperands;
+            };
+
+            // Image Operands bits that may accompany an implicit-LOD sample, in the canonical
+            // ascending order SPIR-V requires the operand values to appear in.
+            static constexpr Uint32 kBias = 0x1;
+            static constexpr Uint32 kLod = 0x2;
+            static constexpr Uint32 kGrad = 0x4;
+            static constexpr Uint32 kConstOffset = 0x8;
+            static constexpr Uint32 kOffset = 0x10;
+            static constexpr Uint32 kConstOffsets = 0x20;
+            static constexpr Uint32 kSample = 0x40;
+            static constexpr Uint32 kMinLod = 0x80;
+            static constexpr Uint32 kKnownMask = 0xFF;
+
+            Uint32 GetFloatZeroId() {
+                // Reuse a 32-bit float type already in the module; a shader that samples always has
+                // one, and looking it up avoids depending on type-creation API details.
+                Uint32 floatTypeId = 0;
+                for (auto& inst : get_module()->types_values()) {
+                    if (inst.opcode() == spv::Op::OpTypeFloat && inst.NumInOperands() >= 1 &&
+                        inst.GetSingleWordInOperand(0) == 32) {
+                        floatTypeId = inst.result_id();
+                        break;
+                    }
+                }
+                if (floatTypeId == 0) return 0;
+
+                const auto* floatType = context()->get_type_mgr()->GetType(floatTypeId);
+                if (floatType == nullptr) return 0;
+                const auto zeroBits = std::bit_cast<Uint32>(0.0f);
+                const auto* zeroConst = context()->get_constant_mgr()->GetConstant(floatType, {zeroBits});
+                if (zeroConst == nullptr) return 0;
+                auto* zeroInst = context()->get_constant_mgr()->GetDefiningInstruction(zeroConst);
+                return zeroInst != nullptr ? zeroInst->result_id() : 0;
+            }
+
+            static Bool MapOpcode(spv::Op op, spv::Op& outOpcode, Uint32& outFixedOperandCount) {
+                switch (op) {
+                case spv::Op::OpImageSampleImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleExplicitLod;
+                    outFixedOperandCount = 2; // sampled image, coordinate
+                    return true;
+                case spv::Op::OpImageSampleProjImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleProjExplicitLod;
+                    outFixedOperandCount = 2;
+                    return true;
+                case spv::Op::OpImageSampleDrefImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleDrefExplicitLod;
+                    outFixedOperandCount = 3; // sampled image, coordinate, Dref
+                    return true;
+                case spv::Op::OpImageSampleProjDrefImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleProjDrefExplicitLod;
+                    outFixedOperandCount = 3;
+                    return true;
+                default:
+                    return false;
+                }
+            }
+
+            static Bool PlanRewrite(spvtools::opt::Instruction* inst, RewritePlan& outPlan) {
+                spv::Op newOpcode = spv::Op::OpNop;
+                Uint32 fixedCount = 0;
+                if (!MapOpcode(inst->opcode(), newOpcode, fixedCount)) return false;
+                if (inst->NumInOperands() < fixedCount) return false;
+
+                Uint32 mask = 0;
+                Uint32 next = fixedCount;
+                if (inst->NumInOperands() > fixedCount) {
+                    mask = inst->GetSingleWordInOperand(fixedCount);
+                    next = fixedCount + 1;
+                }
+                // An operand this pass does not model would be silently reordered or dropped, and
+                // Grad cannot legally accompany an implicit-LOD sample: leave such an instruction be.
+                if ((mask & ~kKnownMask) != 0 || (mask & kGrad) != 0) return false;
+
+                Vector<spvtools::opt::Operand> fixedOperands;
+                fixedOperands.reserve(fixedCount + 1);
+                for (Uint32 i = 0; i < fixedCount; ++i) {
+                    fixedOperands.push_back(inst->GetInOperand(i));
+                }
+
+                // Collect the surviving operand values in the same ascending-bit order they were
+                // encoded in, so the rebuilt list stays canonical.
+                Uint32 keptMask = kLod;
+                Vector<spvtools::opt::Operand> keptOperands;
+                static constexpr Uint32 kOrderedBits[] = {kBias,   kLod,          kGrad,   kConstOffset,
+                                                          kOffset, kConstOffsets, kSample, kMinLod};
+                for (const Uint32 bit : kOrderedBits) {
+                    if ((mask & bit) == 0) continue;
+                    if (next >= inst->NumInOperands()) return false;
+                    const spvtools::opt::Operand value = inst->GetInOperand(next++);
+                    // Bias and MinLod only shift a lambda that is already clamped to 0, and any
+                    // original Lod is replaced by the constant the caller appends.
+                    if (bit == kBias || bit == kMinLod || bit == kLod) continue;
+                    keptMask |= bit;
+                    keptOperands.push_back(value);
+                }
+
+                fixedOperands.push_back({SPV_OPERAND_TYPE_IMAGE, {keptMask}});
+                outPlan.instruction = inst;
+                outPlan.opcode = newOpcode;
+                outPlan.operands = Move(fixedOperands);
+                outPlan.trailingOperands = Move(keptOperands);
+                return true;
+            }
+        };
+
+        spvtools::Optimizer::PassToken CreateForceExplicitLod0SamplePass() {
+            return spvtools::Optimizer::PassToken(MakeUnique<ForceExplicitLod0SamplePass>());
+        }
+
+        // TEMP-PERFDIAG: measure what fragment-stage fp32 costs on this GPU. Desktop GLSL carries
+        // no precision qualifiers, so everything reaches the driver as full fp32 while Adreno runs
+        // fp16 at twice the rate. Decorating every float-typed result in a fragment entry point
+        // with RelaxedPrecision is the blunt "all mediump" upper bound - it changes results, so it
+        // is a probe, not a shipping transform. Toggled by /sdcard/MG/exp_relaxed_precision.
+        class RelaxedPrecisionProbePass final : public spvtools::opt::Pass {
+        public:
+            const char* name() const override { return "relaxed-precision-probe"; }
+
+            Status Process() override {
+                Bool isFragment = false;
+                for (auto& entryPoint : get_module()->entry_points()) {
+                    if (entryPoint.opcode() != spv::Op::OpEntryPoint) continue;
+                    if (static_cast<spv::ExecutionModel>(entryPoint.GetSingleWordInOperand(0)) ==
+                        spv::ExecutionModel::Fragment) {
+                        isFragment = true;
+                        break;
+                    }
+                }
+                if (!isFragment) return Status::SuccessWithoutChange;
+
+                // Every 32-bit-float scalar/vector/matrix type in the module. Anything wider (f64)
+                // or narrower is left alone: RelaxedPrecision only has meaning for 32-bit floats.
+                std::unordered_set<Uint32> relaxableTypes;
+                for (auto& type : get_module()->types_values()) {
+                    const Uint32 typeId = type.result_id();
+                    if (typeId == 0) continue;
+                    switch (type.opcode()) {
+                    case spv::Op::OpTypeFloat:
+                        if (type.GetSingleWordInOperand(0) == 32) relaxableTypes.insert(typeId);
+                        break;
+                    case spv::Op::OpTypeVector:
+                    case spv::Op::OpTypeMatrix:
+                        if (relaxableTypes.count(type.GetSingleWordInOperand(0)) != 0) {
+                            relaxableTypes.insert(typeId);
+                        }
+                        break;
+                    default:
+                        break;
+                    }
+                }
+                if (relaxableTypes.empty()) return Status::SuccessWithoutChange;
+
+                Vector<Uint32> targets;
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            const Uint32 resultId = inst.result_id();
+                            if (resultId == 0) continue;
+                            if (relaxableTypes.count(inst.type_id()) == 0) continue;
+                            targets.push_back(resultId);
+                        }
+                    }
+                }
+                if (targets.empty()) return Status::SuccessWithoutChange;
+
+                for (const Uint32 id : targets) {
+                    context()->get_decoration_mgr()->AddDecoration(
+                        id, static_cast<Uint32>(spv::Decoration::RelaxedPrecision));
+                }
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+        };
+
+        // Relax fragment-stage arithmetic that provably came out of a texture read. Desktop GLSL
+        // has no precision qualifiers, so every fragment value reaches the driver as fp32 while
+        // Adreno runs fp16 at twice the rate - and a texel is at most 8 bits per channel, which
+        // fp16's 11-bit mantissa carries exactly. Seeding at image reads and propagating only
+        // through operations whose every input is already relaxed keeps everything the shader
+        // computes from other sources (screen coordinates, depth, wide-range uniforms) at full
+        // precision, which is where fp16 would actually go wrong: fp16 cannot even represent a
+        // 3044-pixel gl_FragCoord.x exactly.
+        class RelaxTextureDerivedPrecisionPass final : public spvtools::opt::Pass {
+        public:
+            const char* name() const override { return "relax-texture-derived-precision"; }
+
+            Status Process() override {
+                if (!IsFragmentEntryPoint()) return Status::SuccessWithoutChange;
+                // A shader that drives depth or coverage itself is out of scope: those values must
+                // stay exact, and proving which computations feed them is not worth it here.
+                if (WritesDepthOrSampleMask()) return Status::SuccessWithoutChange;
+
+                CollectRelaxableFloatTypes();
+                if (m_relaxableTypes.empty()) return Status::SuccessWithoutChange;
+
+                // Whitelisting from texture reads captures nothing in practice: MC's fragment
+                // shaders multiply every texel by an interpolated colour and a UBO value, so one
+                // un-relaxed operand vetoes the whole expression (measured: no fps change).
+                // Taint the few genuinely precision-critical sources instead and relax the rest.
+                std::unordered_set<Uint32> tainted;
+                CollectPrecisionCriticalSeeds(tainted);
+                Bool grew = true;
+                while (grew) {
+                    grew = false;
+                    for (auto& function : *get_module()) {
+                        for (auto& block : function) {
+                            for (auto& inst : block) {
+                                const Uint32 resultId = inst.result_id();
+                                if (resultId == 0 || tainted.count(resultId) != 0) continue;
+                                if (!AnyOperandTainted(inst, tainted)) continue;
+                                tainted.insert(resultId);
+                                grew = true;
+                            }
+                        }
+                    }
+                }
+
+                std::unordered_set<Uint32> relaxed;
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            const Uint32 resultId = inst.result_id();
+                            if (resultId == 0 || tainted.count(resultId) != 0) continue;
+                            if (m_relaxableTypes.count(inst.type_id()) == 0) continue;
+                            relaxed.insert(resultId);
+                        }
+                    }
+                }
+                if (relaxed.empty()) return Status::SuccessWithoutChange;
+
+                for (const Uint32 id : relaxed) {
+                    context()->get_decoration_mgr()->AddDecoration(
+                        id, static_cast<Uint32>(spv::Decoration::RelaxedPrecision));
+                }
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+        private:
+            std::unordered_set<Uint32> m_relaxableTypes;
+
+            Bool IsFragmentEntryPoint() const {
+                for (auto& entryPoint : get_module()->entry_points()) {
+                    if (entryPoint.opcode() != spv::Op::OpEntryPoint) continue;
+                    if (static_cast<spv::ExecutionModel>(entryPoint.GetSingleWordInOperand(0)) ==
+                        spv::ExecutionModel::Fragment) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
+            Bool WritesDepthOrSampleMask() const {
+                for (auto& annotation : get_module()->annotations()) {
+                    if (annotation.opcode() != spv::Op::OpDecorate) continue;
+                    if (static_cast<spv::Decoration>(annotation.GetSingleWordInOperand(1)) !=
+                        spv::Decoration::BuiltIn) {
+                        continue;
+                    }
+                    const auto builtIn = static_cast<spv::BuiltIn>(annotation.GetSingleWordInOperand(2));
+                    if (builtIn == spv::BuiltIn::FragDepth || builtIn == spv::BuiltIn::SampleMask) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
+            void CollectRelaxableFloatTypes() {
+                m_relaxableTypes.clear();
+                for (auto& type : get_module()->types_values()) {
+                    const Uint32 typeId = type.result_id();
+                    if (typeId == 0) continue;
+                    switch (type.opcode()) {
+                    case spv::Op::OpTypeFloat:
+                        if (type.GetSingleWordInOperand(0) == 32) m_relaxableTypes.insert(typeId);
+                        break;
+                    case spv::Op::OpTypeVector:
+                        if (m_relaxableTypes.count(type.GetSingleWordInOperand(0)) != 0) {
+                            m_relaxableTypes.insert(typeId);
+                        }
+                        break;
+                    default:
+                        break;
+                    }
+                }
+            }
+
+            void CollectImageReadSeeds(std::unordered_set<Uint32>& relaxed) const {
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            const Uint32 resultId = inst.result_id();
+                            if (resultId == 0 || m_relaxableTypes.count(inst.type_id()) == 0) continue;
+                            // Interpolated user varyings seed too, or propagation dies at the
+                            // first `texel * vertexColour`: the load of an Input can never be
+                            // relaxed by the rule below (its operand is a pointer), so a single
+                            // varying vetoes every downstream operation. This is what ESSL's
+                            // mediump varyings already mean. Built-ins are excluded - gl_FragCoord
+                            // carries pixel coordinates that fp16 cannot represent exactly.
+                            if (inst.opcode() == spv::Op::OpLoad && IsNonBuiltInFragmentInput(inst)) {
+                                relaxed.insert(resultId);
+                                continue;
+                            }
+                            switch (inst.opcode()) {
+                            case spv::Op::OpImageSampleImplicitLod:
+                            case spv::Op::OpImageSampleExplicitLod:
+                            case spv::Op::OpImageSampleProjImplicitLod:
+                            case spv::Op::OpImageSampleProjExplicitLod:
+                            case spv::Op::OpImageSampleDrefImplicitLod:
+                            case spv::Op::OpImageSampleDrefExplicitLod:
+                            case spv::Op::OpImageFetch:
+                            case spv::Op::OpImageRead:
+                            case spv::Op::OpImageGather:
+                                relaxed.insert(resultId);
+                                break;
+                            default:
+                                break;
+                            }
+                        }
+                    }
+                }
+            }
+
+            // OpLoad straight out of a fragment Input variable that carries no BuiltIn decoration.
+            // Only a direct load counts: a load through an access chain could be indexing a
+            // structure whose other members are not interpolated colour data.
+            Bool IsNonBuiltInFragmentInput(const spvtools::opt::Instruction& load) const {
+                const Uint32 pointerId = load.GetSingleWordInOperand(0);
+                const auto* pointer = context()->get_def_use_mgr()->GetDef(pointerId);
+                if (pointer == nullptr || pointer->opcode() != spv::Op::OpVariable) return false;
+                if (static_cast<spv::StorageClass>(pointer->GetSingleWordInOperand(0)) !=
+                    spv::StorageClass::Input) {
+                    return false;
+                }
+                Bool isBuiltIn = false;
+                context()->get_decoration_mgr()->ForEachDecoration(
+                    pointerId, static_cast<Uint32>(spv::Decoration::BuiltIn),
+                    [&isBuiltIn](const spvtools::opt::Instruction&) { isBuiltIn = true; });
+                return !isBuiltIn;
+            }
+
+            // A float constant small enough that fp16 represents it without surprise. Colour math
+            // constants (0, 1, 0.5, 255, gamma exponents) all live here; anything larger is
+            // treated as unknown so it stops propagation.
+            Bool IsBoundedFloatConstant(Uint32 id) const {
+                const auto* constant = context()->get_constant_mgr()->FindDeclaredConstant(id);
+                if (constant == nullptr) return false;
+                if (const auto* scalar = constant->AsFloatConstant()) {
+                    const float value = scalar->GetFloat();
+                    return std::isfinite(value) && std::fabs(value) <= 1024.0f;
+                }
+                if (const auto* composite = constant->AsVectorConstant()) {
+                    for (const auto* component : composite->GetComponents()) {
+                        const auto* scalar = component->AsFloatConstant();
+                        if (scalar == nullptr) return false;
+                        const float value = scalar->GetFloat();
+                        if (!std::isfinite(value) || std::fabs(value) > 1024.0f) return false;
+                    }
+                    return true;
+                }
+                return false;
+            }
+
+            // Precision-critical sources: a built-in fragment input. gl_FragCoord is the one that
+            // matters - fp16 cannot represent a 3044-pixel x coordinate exactly, and anything
+            // derived from it (screen-space effects, manual depth reconstruction) would visibly
+            // quantise. Everything else a fragment shader reads is colour-range data.
+            void CollectPrecisionCriticalSeeds(std::unordered_set<Uint32>& tainted) const {
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            if (inst.opcode() != spv::Op::OpLoad || inst.result_id() == 0) continue;
+                            if (IsBuiltInInputLoad(inst)) tainted.insert(inst.result_id());
+                        }
+                    }
+                }
+            }
+
+            Bool IsBuiltInInputLoad(const spvtools::opt::Instruction& load) const {
+                const Uint32 pointerId = load.GetSingleWordInOperand(0);
+                const auto* pointer = context()->get_def_use_mgr()->GetDef(pointerId);
+                if (pointer == nullptr || pointer->opcode() != spv::Op::OpVariable) return false;
+                if (static_cast<spv::StorageClass>(pointer->GetSingleWordInOperand(0)) !=
+                    spv::StorageClass::Input) {
+                    return false;
+                }
+                Bool isBuiltIn = false;
+                context()->get_decoration_mgr()->ForEachDecoration(
+                    pointerId, static_cast<Uint32>(spv::Decoration::BuiltIn),
+                    [&isBuiltIn](const spvtools::opt::Instruction&) { isBuiltIn = true; });
+                return isBuiltIn;
+            }
+
+            Bool AnyOperandTainted(const spvtools::opt::Instruction& inst,
+                                   const std::unordered_set<Uint32>& tainted) const {
+                const Uint32 operandCount = inst.NumInOperands();
+                for (Uint32 i = 0; i < operandCount; ++i) {
+                    const auto& operand = inst.GetInOperand(i);
+                    if (!spvIsIdType(operand.type)) continue;
+                    if (IsNonNumericOperand(inst, i)) continue;
+                    if (tainted.count(operand.words[0]) != 0) return true;
+                }
+                return false;
+            }
+
+            Bool AllValueOperandsRelaxed(const spvtools::opt::Instruction& inst,
+                                         const std::unordered_set<Uint32>& relaxed) const {
+                switch (inst.opcode()) {
+                // Pointer-typed plumbing: relaxing the loaded value would say nothing about the
+                // memory it came from, and the pointer operand can never be in the set.
+                case spv::Op::OpLoad:
+                case spv::Op::OpStore:
+                case spv::Op::OpAccessChain:
+                case spv::Op::OpInBoundsAccessChain:
+                case spv::Op::OpFunctionCall:
+                    return false;
+                default:
+                    break;
+                }
+
+                Bool sawValueOperand = false;
+                Bool allRelaxed = true;
+                const Uint32 operandCount = inst.NumInOperands();
+                for (Uint32 i = 0; i < operandCount; ++i) {
+                    const auto& operand = inst.GetInOperand(i);
+                    if (!spvIsIdType(operand.type)) continue; // literals: selectors, swizzle indices
+                    const Uint32 id = operand.words[0];
+                    // OpPhi's block labels, OpSelect's condition and OpExtInst's instruction-set id
+                    // are ids that carry no numeric precision; skip them rather than let them veto.
+                    if (IsNonNumericOperand(inst, i)) continue;
+                    sawValueOperand = true;
+                    if (relaxed.count(id) != 0) continue;
+                    if (IsBoundedFloatConstant(id)) continue;
+                    allRelaxed = false;
+                    break;
+                }
+                return sawValueOperand && allRelaxed;
+            }
+
+            static Bool IsNonNumericOperand(const spvtools::opt::Instruction& inst, Uint32 index) {
+                switch (inst.opcode()) {
+                case spv::Op::OpPhi:
+                    return (index % 2) == 1; // parent block labels
+                case spv::Op::OpSelect:
+                    return index == 0; // condition
+                case spv::Op::OpExtInst:
+                    return index == 0; // extended instruction set
+                default:
+                    return false;
+                }
+            }
+        };
+
+        // TEMP-PERFDIAG: A/B switch between the scoped transform and the all-float upper bound.
+        Bool PerfDiagRelaxAllPrecision() {
+            static const Bool enabled = [] {
+                std::FILE* probe = std::fopen("/sdcard/MG/exp_relaxed_precision_all", "rb");
+                if (probe == nullptr) return false;
+                std::fclose(probe);
+                MGLOG_I("[PERFDIAG] fragment RelaxedPrecision: ALL floats (upper-bound probe)");
+                return true;
+            }();
+            return enabled;
+        }
+
+        // TEMP-PERFDIAG: lets a run turn the transform off entirely for an A/B baseline.
+        Bool PerfDiagRelaxedPrecisionEnabled() {
+            static const Bool disabled = [] {
+                std::FILE* probe = std::fopen("/sdcard/MG/exp_no_relaxed_precision", "rb");
+                if (probe == nullptr) return false;
+                std::fclose(probe);
+                MGLOG_I("[PERFDIAG] fragment RelaxedPrecision DISABLED");
+                return true;
+            }();
+            return !disabled;
+        }
+
+        Bool TransformSpirvForExplicitLod0Sampling(const Vector<Uint>& input, Vector<Uint>& output) {
+            if (input.empty()) {
+                output.clear();
+                return true;
+            }
+            spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
+            spvtools::OptimizerOptions options;
+            // Matches the position-fix pass: this build of spirv-tools asserts rather than
+            // reporting, so validation stays off in the shipping path.
+            options.set_run_validator(false);
+            optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
+                                            const char* message) {
+                MGLOG_E("Vulkan: explicit-LOD0 pass: %s", message != nullptr ? message : "");
+            });
+            optimizer.RegisterPass(CreateForceExplicitLod0SamplePass());
+
+            const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
+            if (!success) {
+                MGLOG_E("Vulkan: explicit-LOD0 sampling pass failed; keeping the original module");
+                output = input;
+            }
+            return success;
+        }
+
         spvtools::Optimizer::PassToken CreateGlToVulkanPositionFixPass(
             ProgramFactory::CompileOptionFlags transformFlags) {
             return spvtools::Optimizer::PassToken(MakeUnique<GlToVulkanPositionFixPass>(transformFlags));
         }
 
+        // TEMP-PERFDIAG
+        Bool TransformSpirvForRelaxedPrecisionProbe(const Vector<Uint>& input, Vector<Uint>& output) {
+            if (input.empty()) {
+                output.clear();
+                return true;
+            }
+            spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
+            spvtools::OptimizerOptions options;
+            options.set_run_validator(false);
+            optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
+                                            const char* message) {
+                MGLOG_E("Vulkan: relaxed-precision probe: %s", message != nullptr ? message : "");
+            });
+            // SSA promotion first: glslang emits function-local variables with stores and loads,
+            // and a load can never be relaxed (its operand is a pointer), so without this the
+            // propagation below dies at the first temporary.
+            optimizer.RegisterPass(spvtools::CreateLocalMultiStoreElimPass());
+            if (PerfDiagRelaxAllPrecision()) {
+                optimizer.RegisterPass(spvtools::Optimizer::PassToken(MakeUnique<RelaxedPrecisionProbePass>()));
+            } else {
+                optimizer.RegisterPass(
+                    spvtools::Optimizer::PassToken(MakeUnique<RelaxTextureDerivedPrecisionPass>()));
+            }
+            const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
+            if (!success) {
+                MGLOG_E("Vulkan: relaxed-precision probe failed; keeping the original module");
+                output = input;
+            }
+            return success;
+        }
+
         Bool TransformSpirvForVulkanPositionFix(const Vector<Uint>& input, Vector<Uint>& output,
                                                 ProgramFactory::CompileOptionFlags transformFlags) {
             if (input.empty()) {
@@ -1950,11 +2552,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         auto it = m_cache.find(hash);
         if (it != m_cache.end()) {
+            // Every draw/dispatch funnels through this lookup (the renderer memos only
+            // skip re-hashing, never the factory lookup), so an actively-used entry is
+            // stamped at least once per frame boundary and can never be aged out while
+            // any in-flight command buffer still references it.
+            it->second.lastUsedFrame = m_frameCounter;
             return it->second;
         }
 
         auto& entry = m_cache[hash];
         entry.hash = hash;
+        entry.lastUsedFrame = m_frameCounter;
         auto& shaders = program.GetAttachedShaders();
         auto& spirv = program.GetGeneratedSpirv();
         Vector<Vector<Uint>> moduleSpirvs(spirv.size());
@@ -1972,6 +2580,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 moduleSpirvs[i] = spv;
             }
 
+            if ((flags & ProgramFactory::CompileOptionBit::ExplicitLod0Sampling) && shaders[i] &&
+                shaders[i]->GetShaderStage() == ShaderStage::Fragment) {
+                Vector<Uint> explicitLodSpirv;
+                if (TransformSpirvForExplicitLod0Sampling(moduleSpirvs[i], explicitLodSpirv)) {
+                    moduleSpirvs[i] = Move(explicitLodSpirv);
+                }
+            }
+
+            if ((flags & ProgramFactory::CompileOptionBit::RelaxedFragmentPrecision) &&
+                PerfDiagRelaxedPrecisionEnabled() && shaders[i] &&
+                shaders[i]->GetShaderStage() == ShaderStage::Fragment) {
+                Vector<Uint> relaxedSpirv;
+                if (TransformSpirvForRelaxedPrecisionProbe(moduleSpirvs[i], relaxedSpirv)) {
+                    moduleSpirvs[i] = Move(relaxedSpirv);
+                }
+            }
+
             // GL apps depend on cross-program position invariance for multi-pass equality
             // depth tests (MC 26.3's OIT re-draws the cloud geometry with GEQUAL against the
             // depth its own first pass wrote); decorate Position outputs Invariant so
@@ -2068,4 +2693,42 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         return entry;
     }
+
+    void ProgramFactory::OnFrameBoundary() {
+        ++m_frameCounter;
+
+        // Sweep cadence and retire age mirror VkRenderPassManager::OnPresent: an entry
+        // idle for more than kRetireAgeFrames frame boundaries cannot be referenced by
+        // any in-flight command buffer (frames-in-flight <= MOBILEGL_MAGMA_FRAMESINFLIGHT),
+        // so its shader modules and layouts are destroyed immediately - no deferred-
+        // destroy machinery needed. Eviction is content-based, never tied to
+        // glDeleteProgram: the cache is content-hash-shared across GL programs, so a
+        // delete-driven erase could free an entry another live program still resolves.
+        // An evicted entry self-heals - the frontend program keeps its generated
+        // SPIR-V, so the next GetOrCreateProgram rebuilds it (this also covers the
+        // renderer's internal blit/depth-mipmap programs).
+        constexpr Uint64 kSweepInterval = 256;
+        constexpr Uint64 kRetireAgeFrames = 1024;
+        if ((m_frameCounter % kSweepInterval) != 0) {
+            return;
+        }
+
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (m_frameCounter - it->second.lastUsedFrame > kRetireAgeFrames) {
+                const HashType hash = it->first;
+                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
+                // so an observer never observes a half-destroyed entry through a lookup.
+                // Observers only need the handle values to purge their keyed caches.
+                it = m_cache.erase(it);
+                if (m_evictionObserver != nullptr) {
+                    m_evictionObserver->OnProgramEvicted(hash, descriptorSetLayout);
+                }
+            } else {
+                ++it;
+            }
+        }
+    }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -42,6 +42,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             SurfaceRotate90 = 1 << 2,
             SurfaceRotate180 = 1 << 3,
             SurfaceRotate270 = 1 << 4,
+            // Rewrites the fragment stage's implicit-LOD image samples to explicit LOD 0.
+            // Only ever set for a draw whose every sampler binding is clamped to a single mip
+            // level, which makes the two forms produce identical texels (the implicit lambda is
+            // clamped into [minLod, maxLod] = [0, 0] regardless of derivatives or bias).
+            ExplicitLod0Sampling = 1 << 5,
+            // Fragment arithmetic may run at relaxed (fp16) precision. Only requested for draws
+            // where every sampled texture and every colour attachment is an 8-bit-or-less
+            // normalized format, so nothing the shader reads or writes carries more precision
+            // than fp16 already represents exactly.
+            RelaxedFragmentPrecision = 1 << 6,
         };
         using CompileOptionFlags = Flags<CompileOptionBit>;
         using HashType = Uint64;
@@ -88,6 +98,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // gl_FragDepth); shader-computed depth is immune to the cross-pipeline
             // position-invariance quirk (see PipelineFactory::ShouldSuppressDepthWrite).
             Bool fragmentReplacesDepth = false;
+            // Frame-boundary counter value of the last GetOrCreateProgram hit; drives
+            // cache eviction (see OnFrameBoundary).
+            Uint64 lastUsedFrame = 0;
 
             static inline VkDevice s_device = VK_NULL_HANDLE;
 
@@ -124,6 +137,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 producerOutputComponentCount = other.producerOutputComponentCount;
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
+                lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
@@ -135,6 +149,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.producerOutputComponentCount = 0;
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
+                other.lastUsedFrame = 0;
             }
             VkProgramObject& operator=(VkProgramObject&& other) noexcept {
                 if (this == &other) {
@@ -170,6 +185,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 producerOutputComponentCount = other.producerOutputComponentCount;
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
+                lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
@@ -181,6 +197,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.producerOutputComponentCount = 0;
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
+                other.lastUsedFrame = 0;
                 return *this;
             }
 
@@ -210,6 +227,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         };
 
+        // Notified when the OnFrameBoundary sweep destroys an aged-out cache entry,
+        // carrying the entry's content hash and the VkDescriptorSetLayout it owned.
+        // Dependent caches (compute pipelines, PipelineFactory entries, UniformManager's
+        // per-layout descriptor sets) must purge in the same step: after vkDestroy the
+        // layout handle value may be recycled for an unrelated layout, and the program
+        // hash may be re-inserted by a later rebuild of the same content.
+        class IEvictionObserver {
+        public:
+            virtual ~IEvictionObserver() = default;
+            virtual void OnProgramEvicted(HashType programHash, VkDescriptorSetLayout descriptorSetLayout) = 0;
+        };
+
         explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings = 16,
                                 Bool shaderDrawParametersEnabled = false,
                                 Bool unformattedFloatStorageImagesEnabled = false)
@@ -225,6 +254,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkProgramObject& GetOrCreateProgram(
             const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags);
 
+        // Observer may be null (no notifications). Not owned.
+        void SetEvictionObserver(IEvictionObserver* observer) { m_evictionObserver = observer; }
+        // Frame boundary hook: ages the program cache and evicts long-unused entries
+        // (their command buffers retired many frames ago), mirroring
+        // VkRenderPassManager::OnPresent's sweep.
+        void OnFrameBoundary();
+
         static VkShaderStageFlagBits ToVkStage(ShaderStage stage);
         static VkFormat ConvertSpirvImageFormatToVkFormat(SpvImageFormat format);
         static SamplerNumericDomain UniformTypeToSamplerNumericDomain(GLenum glType);
@@ -266,6 +302,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
         Bool m_unformattedFloatStorageImagesEnabled = false;
         mutable ProgramLookupCache m_lastLookup;
+        // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+        Uint64 m_frameCounter = 0;
+        IEvictionObserver* m_evictionObserver = nullptr;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.cpp

@@ -247,6 +247,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         m_surfaceFormat = {createInfo.imageFormat, createInfo.imageColorSpace};
         m_extent = createInfo.imageExtent;
+        // The surface-space extent this swapchain was built from, i.e. before the
+        // quarter-turn swap above. Out-of-date checks must compare in THIS space: comparing a
+        // freshly queried currentExtent against the swapped m_extent flips axes every rotation
+        // and makes the comparison alternate forever.
+        m_surfaceExtent = defaultFramebufferExtent;
         m_preTransform = createInfo.preTransform;
 
         VK_VERIFY(vkCreateSwapchainKHR(device, &createInfo, nullptr, &m_swapchain));

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.h

@@ -35,6 +35,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkSwapchainKHR GetHandle() const { return m_swapchain; }
         const VkSurfaceFormatKHR& GetSurfaceFormat() const { return m_surfaceFormat; }
         VkExtent2D GetExtent() const { return m_extent; }
+        // Surface-space extent (before the pre-rotation quarter-turn swap) this swapchain was
+        // created from - the value to compare a freshly queried currentExtent against.
+        VkExtent2D GetSurfaceExtent() const { return m_surfaceExtent; }
         VkSurfaceTransformFlagBitsKHR GetPreTransform() const { return m_preTransform; }
         const Vector<VkImage>& GetImages() const { return m_images; }
         const Vector<VkImageView>& GetImageViews() const { return m_imageViews; }
@@ -63,6 +66,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkSwapchainKHR m_swapchain = VK_NULL_HANDLE;
         VkSurfaceFormatKHR m_surfaceFormat{};
         VkExtent2D m_extent{};
+        VkExtent2D m_surfaceExtent{};
         VkSurfaceTransformFlagBitsKHR m_preTransform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
         Vector<VkImage> m_images;
         Vector<VkImageView> m_imageViews;

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -16,6 +16,7 @@
 #include "MG_Util/Converters/GLToMG/TextureEnumConverter.h"
 #include "MG_Util/Converters/MGToStr/FramebufferEnumConverter.h"
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
+#include <vulkan/utility/vk_format_utils.h>
 #include "MG_Util/Metrics/TextureMetrics.h"
 #include <Config.h>
 #include <cstdio>
@@ -211,6 +212,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    void UniformManager::OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout) {
+        SizeT purgedSets = 0;
+        for (auto& frame : m_frames) {
+            const auto it = frame.descriptorSetCacheByLayout.find(descriptorSetLayout);
+            if (it == frame.descriptorSetCacheByLayout.end()) {
+                continue;
+            }
+            // Free the sets back to their pools and credit the bucket accounting, so
+            // program churn recycles pool capacity instead of abandoning the slots.
+            // GPU-safe: the layout only dies after >1024 idle frame boundaries, so no
+            // in-flight command buffer references these sets.
+            for (const auto& cached : it->second.sets) {
+                if (cached.set == VK_NULL_HANDLE) {
+                    continue;
+                }
+                vkFreeDescriptorSets(m_device, cached.pool, 1, &cached.set);
+                const auto bucket = std::find_if(
+                    frame.descriptorPools.begin(), frame.descriptorPools.end(),
+                    [&cached](const DescriptorPoolBucket& candidate) { return candidate.handle == cached.pool; });
+                if (bucket != frame.descriptorPools.end() && bucket->allocatedSets > 0) {
+                    --bucket->allocatedSets;
+                }
+            }
+            purgedSets += it->second.sets.size();
+            frame.descriptorSetCacheByLayout.erase(it);
+        }
+        if (purgedSets > 0) {
+            // The per-draw reuse memo folds the layout handle into its signature; drop
+            // it so a recycled handle value cannot revive a purged set mid-frame.
+            m_hasLastDescriptor = false;
+            MGLOG_D("UniformDescriptorBinder: freed %zu descriptor sets for destroyed layout", purgedSets);
+        }
+    }
+
     Bool UniformManager::ResolveSamplerDescriptor(VkCommandBuffer commandBuffer,
                                                             const MG_State::GLState::ProgramObject& program,
                                                             const ProgramFactory::VkProgramObject& programObj,
@@ -350,24 +385,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const Uint16 samplerVersion = samplerToUse->GetVersion();
             const Uint64 textureLifetimeId = texture->GetLifetimeId();
             const Uint16 textureParamsVersion = texture->GetTextureParamsVersion();
+            // The sampler's LOD clamp depends on how many levels the sampled view exposes, and that
+            // follows uploads as well as GL parameters - so it belongs in the memo key too.
+            const Uint32 viewLevelCount = resource->sampledLevelCount;
             if (memo.valid && memo.samplerLifetimeId == samplerLifetimeId && memo.samplerVersion == samplerVersion &&
                 memo.textureLifetimeId == textureLifetimeId && memo.textureParamsVersion == textureParamsVersion &&
-                memo.forceNearestFiltering == forceNearestFiltering) {
+                memo.forceNearestFiltering == forceNearestFiltering && memo.viewLevelCount == viewLevelCount) {
                 resolvedSampler = memo.sampler;
             } else {
-                resolvedSampler =
-                    m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
+                resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture,
+                                                                       forceNearestFiltering, viewLevelCount);
                 memo.samplerLifetimeId = samplerLifetimeId;
                 memo.samplerVersion = samplerVersion;
                 memo.textureLifetimeId = textureLifetimeId;
                 memo.textureParamsVersion = textureParamsVersion;
                 memo.forceNearestFiltering = forceNearestFiltering;
+                memo.viewLevelCount = viewLevelCount;
                 memo.sampler = resolvedSampler;
                 memo.valid = true;
             }
         } else {
-            resolvedSampler =
-                m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
+            resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering,
+                                                                   resource->sampledLevelCount);
         }
         outImageInfo = {
             .sampler = resolvedSampler,
@@ -408,6 +447,106 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return outImageInfo.sampler != VK_NULL_HANDLE;
     }
 
+    namespace {
+        // fp16 carries an 11-bit mantissa, so an 8-bit normalized channel round-trips exactly.
+        // Anything wider - 16-bit normalized, half float, full float, and every packed HDR
+        // encoding - holds precision or range that relaxing the arithmetic would throw away.
+        Bool IsLowPrecisionNormalizedFormat(VkFormat format) {
+            if (format == VK_FORMAT_UNDEFINED) return false;
+            if (!vkuFormatIsUNORM(format) && !vkuFormatIsSNORM(format) && !vkuFormatIsSRGB(format)) {
+                return false;
+            }
+            const struct VKU_FORMAT_INFO info = vkuGetFormatInfo(format);
+            for (Uint32 i = 0; i < info.component_count; ++i) {
+                if (info.components[i].size > 8) return false;
+            }
+            return info.component_count > 0;
+        }
+    } // namespace
+
+    Bool UniformManager::DrawTargetIsLowPrecision(const MG_State::GLState::FramebufferObject* drawFramebuffer) {
+        // Default framebuffer: the swapchain is an 8-bit normalized surface.
+        if (drawFramebuffer == nullptr) return true;
+
+        Bool sawColour = false;
+        for (Int i = static_cast<Int>(FramebufferAttachmentType::Color0);
+             i < static_cast<Int>(FramebufferAttachmentType::FramebufferAttachmentTypeCount);
+             ++i) {
+            const auto& attachment =
+                drawFramebuffer->GetAttachment(static_cast<FramebufferAttachmentType>(i));
+            VkFormat format = VK_FORMAT_UNDEFINED;
+            if (const auto& texture = attachment.GetTexture()) {
+                format = MG_Util::ConvertTextureInternalFormatToVkEnum(texture->GetFormat());
+            } else if (const auto& renderbuffer = attachment.GetRenderbuffer()) {
+                format = MG_Util::ConvertTextureInternalFormatToVkEnum(
+                    renderbuffer->GetInternalFormat());
+            } else {
+                continue;
+            }
+            if (!IsLowPrecisionNormalizedFormat(format)) return false;
+            sawColour = true;
+        }
+        return sawColour;
+    }
+
+    Bool UniformManager::ProgramSamplesOnlyLowPrecisionTextures(
+        const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj) {
+        for (Uint32 binding = 0; binding < programObj.bindingKinds.size(); ++binding) {
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding);
+            // An unresolvable binding is unknown territory, not licence to relax.
+            if (texture == nullptr) return false;
+            const VkFormat format =
+                MG_Util::ConvertTextureInternalFormatToVkEnum(texture->GetFormat());
+            if (!IsLowPrecisionNormalizedFormat(format)) return false;
+        }
+        return true;
+    }
+
+    Bool UniformManager::ProgramSamplesOnlySingleLevelTextures(
+        const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj) {
+        Bool sawSampler = false;
+        for (Uint32 binding = 0; binding < programObj.bindingKinds.size(); ++binding) {
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding);
+            if (texture == nullptr) return false;
+            const auto& levelRange = texture->GetLevelRange();
+            if (levelRange.x() != levelRange.y()) return false;
+
+            // An explicit-LOD sample is a single filtered tap, so it also gives up anisotropic
+            // filtering - which a single-level view can still have. Resolve the sampler exactly
+            // the way ResolveSamplerDescriptor does and bail if anisotropy would apply.
+            const Int location = programObj.samplerUniformLocationByBinding[binding];
+            const Int unit = ResolveSamplerUnitIndex(program, location, binding);
+            const auto& samplerOverride = MG_State::pGLContext->GetTextureUnitObject(unit).GetSamplerObject();
+            const auto* effectiveSampler =
+                samplerOverride ? samplerOverride.get() : texture->GetSamplerObject().get();
+            if (effectiveSampler == nullptr) return false;
+            if (effectiveSampler->GetMaxAnisotropy() > 1.0f &&
+                effectiveSampler->GetMinFilter() == SamplerFilterMode::Linear &&
+                effectiveSampler->GetMagFilter() == SamplerFilterMode::Linear) {
+                return false;
+            }
+
+            // An explicit LOD 0 makes lambda exactly 0, which is the magnification side of the
+            // min/mag decision. That only matches the implicit form when lambda could not have been
+            // positive anyway (the LOD clamp already pins it at or below 0), or when the two
+            // filters are the same and the choice cannot be observed.
+            const Float effectiveMaxLod = effectiveSampler->GetMipmapMode() == SamplerMipmapMode::None
+                                              ? 0.0f
+                                              : effectiveSampler->GetMaxLod();
+            if (effectiveMaxLod > 0.0f && effectiveSampler->GetMinFilter() != effectiveSampler->GetMagFilter()) {
+                return false;
+            }
+            sawSampler = true;
+        }
+        return sawSampler;
+    }
+
     Bool UniformManager::ResolveSamplerTexture(const MG_State::GLState::ProgramObject& program,
                                                          const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                                          SharedPtr<MG_State::GLState::ITextureObject>& outTexture) {
@@ -911,6 +1050,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkDescriptorPoolCreateInfo poolInfo{};
         poolInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
+        // FREE_DESCRIPTOR_SET_BIT lets a destroyed layout's cached sets be freed back
+        // (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.maxSets = maxSets;
         poolInfo.poolSizeCount = static_cast<Uint32>(std::size(poolSizes));
         poolInfo.pPoolSizes = poolSizes;
@@ -990,7 +1134,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto& frame = m_frames[frameIndex];
         auto& cache = frame.descriptorSetCacheByLayout[programObj.descriptorSetLayout];
         if (cache.cursor < cache.sets.size()) {
-            outDescriptorSet = cache.sets[cache.cursor++];
+            outDescriptorSet = cache.sets[cache.cursor++].set;
         } else {
             VkResult allocResult = AllocateDescriptorSetsFromActivePool(frameIndex, programObj, outDescriptorSet);
             if (allocResult == VK_ERROR_OUT_OF_POOL_MEMORY || allocResult == VK_ERROR_FRAGMENTED_POOL) {
@@ -1004,7 +1148,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 return allocResult;
             }
 
-            cache.sets.push_back(outDescriptorSet);
+            // The successful allocation came from the bucket the alloc helper left
+            // active; record it so a layout-destroyed purge can free the set back.
+            cache.sets.push_back({outDescriptorSet, frame.descriptorPools[frame.activeDescriptorPoolIndex].handle});
             ++cache.cursor;
             MGLOG_D("UniformDescriptorBinder: cached descriptor set count for frame=%u grew to %zu", frameIndex,
                     cache.sets.size());

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -39,6 +39,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void Shutdown();
 
         void BeginFrame(Uint32 frameIndex);
+        // A ProgramFactory eviction just destroyed this layout: purge every frame
+        // slot's cached descriptor sets for it, so a recycled handle value can never
+        // stale-hit sets written for the dead layout's bindings. The sets are
+        // vkFreeDescriptorSets'd back to their pools (created with
+        // FREE_DESCRIPTOR_SET_BIT) and the pool accounting is credited, so program
+        // churn recycles pool capacity instead of abandoning it. GPU-safe: the layout
+        // only dies after >1024 idle frame boundaries, so no in-flight command buffer
+        // references its sets. This is the only eviction path for the per-layout
+        // caches - a live layout's entry must never be purged (its sets would be
+        // unreachable pool slots), so there is deliberately no age-based sweep here.
+        void OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout);
         Bool CollectSampledTextures(const MG_State::GLState::ProgramObject& program,
                                     const ProgramFactory::VkProgramObject& programObj,
                                     Vector<MG_State::GLState::ITextureObject*>& outTextures);
@@ -58,6 +69,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         static VkFormat ResolveStorageImageViewFormat(VkFormat reflectedFormat, GLenum bindingFormat,
                                                       VkFormat resourceFormat, Bool useBindingFormat);
 
+        // True when the program reads at least one sampler and every one of them is bound to a
+        // texture whose GL level range is a single level. Such a sampler resolves to
+        // minLod = maxLod = 0 (see VkSamplerManager::GetOrCreateSampler), so an implicit-LOD sample
+        // and an explicit LOD 0 sample must read the same texel - which is what makes the
+        // ExplicitLod0Sampling SPIR-V rewrite safe to request. Deliberately conservative: it reads
+        // only GL state, so a texture that ends up single-level for another reason (one uploaded
+        // level under a wide level range) merely misses the rewrite.
+        // True when every texture this program samples is an 8-bit-or-less normalized format, so
+        // relaxing the fragment stage to fp16 cannot lose a bit the texel ever carried. Says
+        // nothing about the render target - the caller must check that too.
+        static Bool ProgramSamplesOnlyLowPrecisionTextures(const MG_State::GLState::ProgramObject& program,
+                                                          const ProgramFactory::VkProgramObject& programObj);
+        // True when every colour attachment the draw writes is an 8-bit-or-less normalized
+        // format (nullptr = default framebuffer, which is). Blending happens at attachment
+        // precision, so a wider target must keep the fragment stage at full precision.
+        static Bool DrawTargetIsLowPrecision(const MG_State::GLState::FramebufferObject* drawFramebuffer);
+        static Bool ProgramSamplesOnlySingleLevelTextures(const MG_State::GLState::ProgramObject& program,
+                                                          const ProgramFactory::VkProgramObject& programObj);
+
     private:
         struct DescriptorPoolBucket {
             VkDescriptorPool handle = VK_NULL_HANDLE;
@@ -65,8 +95,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint32 allocatedSets = 0;
         };
 
+        // A cached descriptor set together with the pool it was allocated from, so a
+        // layout-destroyed purge can vkFreeDescriptorSets it back and credit the
+        // owning bucket's accounting.
+        struct CachedDescriptorSet {
+            VkDescriptorSet set = VK_NULL_HANDLE;
+            VkDescriptorPool pool = VK_NULL_HANDLE;
+        };
+
         struct DescriptorSetCacheEntry {
-            Vector<VkDescriptorSet> sets;
+            Vector<CachedDescriptorSet> sets;
             Uint32 cursor = 0;
         };
 
@@ -172,6 +210,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 samplerLifetimeId = 0;
             Uint64 textureLifetimeId = 0;
             VkSampler sampler = VK_NULL_HANDLE;
+            Uint32 viewLevelCount = 0;
             Uint16 samplerVersion = 0;
             Uint16 textureParamsVersion = 0;
             Bool forceNearestFiltering = false;

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp

@@ -32,6 +32,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             XXHASH_VERIFY(XXH64_update(m_hashState, &attr.IsBgra, sizeof(attr.IsBgra)));
             XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Divisor, sizeof(attr.Divisor)));
 
+            // The buffer's heap address is an identity component of the key: a freed
+            // buffer's reused address can alias an old cache entry, but only under a
+            // byte-identical attribute layout - and the entry payload is a pure function
+            // of the hashed inputs, with the draw path re-resolving bindingBufferKeys
+            // against the live VAO attribute pointers, so an aliased hit returns exactly
+            // what a rebuild would. Address drift only grows the map; the OnFrameBoundary
+            // aging sweep bounds that.
             const SizeT bufferKey = reinterpret_cast<SizeT>(attr.Buffer.get());
             XXHASH_VERIFY(XXH64_update(m_hashState, &bufferKey, sizeof(bufferKey)));
         }
@@ -58,6 +65,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const MG_State::GLState::VertexArrayObject& vao, HashType hash) {
         auto it = m_cache.find(hash);
         if (it != m_cache.end()) {
+            it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
             return it->second;
         }
 
@@ -166,6 +174,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         auto& entry = m_cache[hash];
         entry.hash = hash;
+        entry.lastUsedFrameBoundary = m_frameBoundaryCounter;
         entry.bindings = builder.GetBindings();
         entry.attributes = builder.GetAttributes();
         entry.bindingBufferKeys = std::move(bindingBufferKeys);
@@ -180,6 +189,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return entry;
     }
 
+    void VertexInputStateFactory::OnFrameBoundary() {
+        ++m_frameBoundaryCounter;
+
+        // Sweep occasionally; evict entries whose last hit is far in the past.
+        // Erasure happens only here, never mid-frame: the draw path holds a
+        // reference into the current entry across its setup, and unordered_map
+        // erase would invalidate it. Entries are CPU-side only, so no GPU-idle
+        // proof is needed; an evicted entry that is used again is simply rebuilt
+        // from the VAO state (same hash, same content).
+        constexpr Uint64 kSweepInterval = 256;
+        constexpr Uint64 kRetireAgeBoundaries = 1024;
+        if ((m_frameBoundaryCounter % kSweepInterval) != 0) {
+            return;
+        }
+
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (m_frameBoundaryCounter - it->second.lastUsedFrameBoundary > kRetireAgeBoundaries) {
+                it = m_cache.erase(it);
+            } else {
+                ++it;
+            }
+        }
+    }
+
     VkFormat VertexInputStateFactory::ToVkVertexFormat(DataType type, Int size, Bool normalized, Bool isInteger,
                                                        Bool isBgra) {
         if (isBgra) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -27,6 +27,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         struct BackendVertexInputState {
             HashType hash = 0;
+            // Frame boundary of the last cache hit; entries idle past the
+            // OnFrameBoundary retirement age are evicted (CPU heap only).
+            Uint64 lastUsedFrameBoundary = 0;
             Vector<VkVertexInputBindingDescription> bindings;
             Vector<VkVertexInputAttributeDescription> attributes;
             Vector<SizeT> bindingBufferKeys;
@@ -55,6 +58,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const BackendVertexInputState& GetOrCreateVertexInputState(
             const MG_State::GLState::VertexArrayObject& vao, HashType hash);
         const BackendVertexInputState& GetOrCreateVertexInputState(const MG_State::GLState::VertexArrayObject& vao);
+        // Frame boundary hook: ages the cache and evicts entries not hit for many
+        // frames. The key mixes buffer heap addresses, so buffer/VAO churn keeps
+        // minting fresh keys; without eviction the map grows for the whole session.
+        // Entries hold no Vulkan handles (pipeline creation copies the descriptions)
+        // and the draw path's entry reference never spans a frame boundary, so
+        // eviction here needs no GPU-idle proof. Self-gated: one counter bump and
+        // compare except on sweep boundaries.
+        void OnFrameBoundary();
         static SizeT GetComponentSize(DataType type);
         // Tightly-packed byte size of one vertex element for this attribute: componentSize * size for
         // normal types, and 4 (one packed word) for the 2_10_10_10 types and GL_BGRA. Returns 0 for
@@ -70,6 +81,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VulkanRendererConfig& m_config;
         VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
         UnorderedMap<HashType, BackendVertexInputState> m_cache;
+        // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+        Uint64 m_frameBoundaryCounter = 0;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.cpp

@@ -141,6 +141,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_transientUploadArena.BeginFrame(frameIndex);
     }
 
+    void VkBufferManager::CollectAllDeferredReleases() {
+        for (Uint32 frameIndex = 0; frameIndex < m_deferredBufferReleases.size(); ++frameIndex) {
+            CollectDeferredReleases(frameIndex);
+        }
+        for (Uint32 frameIndex = 0; frameIndex < m_transientUploadArena.GetFrameCount(); ++frameIndex) {
+            m_transientUploadArena.CollectDeferredReleases(frameIndex);
+        }
+    }
+
     void VkBufferManager::NotifyDeviceIdle() {
         // Everything submitted so far has completed. Work recorded for the
         // current frame has not been submitted yet, so the current serial

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.h

@@ -77,6 +77,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Recreate all per-frame transient arenas
         Bool RecreateTransientArenas(Uint32 frameCount);
         void BeginFrame(Uint32 frameIndex);
+        // Drains every frame slot's deferred buffer/resource releases (and the
+        // transient arena's parked superseded blocks). Only valid when the
+        // caller has proven every queue submission complete; used by the
+        // present-less frame-boundary drain.
+        void CollectAllDeferredReleases();
         // All previously submitted GPU work has completed (vkDeviceWaitIdle).
         void NotifyDeviceIdle();
         // A frame slot's submission fence has been waited: every serial up to

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.cpp

@@ -180,6 +180,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         sampleCount = VK_SAMPLE_COUNT_1_BIT;
         internalFormat = TextureInternalFormat::Unknown;
         samples = 0;
+        deadSinceFrame = kNeverObservedDead;
     }
 
     VkRenderPassManager::VkRenderPassManager(VkDevice device,
@@ -206,6 +207,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.Destroy(m_device, m_allocator);
         }
         m_renderbufferResources.clear();
+        CollectDeferredRenderbufferReleases(/*destroyAll=*/true); // caller guarantees device idle
         m_pendingRenderbufferClears.clear();
         RenderPassEntry::s_textureResourcesScratch.clear();
         s_activeRenderPass = {};
@@ -213,22 +215,75 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_rpFastValid = false;
     }
 
+    Uint64 VkRenderPassManager::RetireAgeFrames() const {
+        // MaxFramesInFlight + 2 covers the frame ring plus one boundary for the
+        // recording-to-submit gap and one because OnPresent runs ahead of Present's
+        // fence wait; the floor of 8 keeps a margin over the default ring of 3 while
+        // still releasing multi-MB attachment memory promptly (the render-pass cache's
+        // 1024-frame retirement would pin it for no additional safety).
+        return std::max<Uint64>(8, static_cast<Uint64>(m_config.MaxFramesInFlight) + 2);
+    }
+
+    void VkRenderPassManager::DeferRenderbufferBackingRelease(RenderbufferResource& resource) {
+        // The superseded backing may still be referenced by in-flight command buffers
+        // (glRenderbufferStorage can respecify a renderbuffer drawn this very frame),
+        // so it is parked and destroyed only after RetireAgeFrames() boundaries.
+        if (resource.image == VK_NULL_HANDLE && resource.view == VK_NULL_HANDLE) {
+            return;
+        }
+        m_deferredRenderbufferReleases.push_back({resource.image, resource.allocation, resource.view, m_frameCounter});
+        resource.image = VK_NULL_HANDLE;
+        resource.allocation = nullptr;
+        resource.view = VK_NULL_HANDLE;
+    }
+
+    void VkRenderPassManager::CollectDeferredRenderbufferReleases(Bool destroyAll) {
+        if (m_deferredRenderbufferReleases.empty()) {
+            return;
+        }
+        const Uint64 retireAgeFrames = RetireAgeFrames();
+        std::erase_if(m_deferredRenderbufferReleases, [&](DeferredRenderbufferRelease& release) {
+            if (!destroyAll && m_frameCounter - release.deferredAtFrame < retireAgeFrames) {
+                return false;
+            }
+            if (release.view != VK_NULL_HANDLE) {
+                vkDestroyImageView(m_device, release.view, nullptr);
+            }
+            if (release.image != VK_NULL_HANDLE) {
+                vmaDestroyImage(m_allocator, release.image, release.allocation);
+            }
+            return true;
+        });
+    }
+
     void VkRenderPassManager::CollectRenderbufferGarbage() {
-        Vector<MG_State::GLState::RenderbufferObject*> deadRenderbuffers;
-        deadRenderbuffers.reserve(m_renderbufferResources.size());
-        for (auto& [renderbuffer, resource] : m_renderbufferResources) {
+        // Two-phase reclamation: a dead renderbuffer's VkImage may still be referenced by
+        // command buffers submitted up to frames-in-flight frames ago (it was legally
+        // attached and drawn right up to its deletion), so the first observation of an
+        // expired weak reference only stamps the current frame counter; Destroy runs once
+        // enough frame boundaries have passed that the stamping frame's submission fence
+        // has provably been waited (see RetireAgeFrames).
+        const Uint64 retireAgeFrames = RetireAgeFrames();
+        for (auto it = m_renderbufferResources.begin(); it != m_renderbufferResources.end();) {
+            auto& resource = it->second;
             const auto liveRenderbuffer = resource.renderbuffer.lock();
-            if (!liveRenderbuffer || liveRenderbuffer.get() != renderbuffer) {
-                deadRenderbuffers.emplace_back(renderbuffer);
+            if (liveRenderbuffer && liveRenderbuffer.get() == it->first) {
+                resource.deadSinceFrame = RenderbufferResource::kNeverObservedDead;
+                ++it;
+                continue;
             }
+            if (resource.deadSinceFrame == RenderbufferResource::kNeverObservedDead) {
+                resource.deadSinceFrame = m_frameCounter;
+                ++it;
+                continue;
             }
-        for (auto* renderbuffer : deadRenderbuffers) {
-            auto resourceIt = m_renderbufferResources.find(renderbuffer);
-            if (resourceIt != m_renderbufferResources.end()) {
-                resourceIt->second.Destroy(m_device, m_allocator);
-                m_renderbufferResources.erase(resourceIt);
+            if (m_frameCounter - resource.deadSinceFrame < retireAgeFrames) {
+                ++it;
+                continue;
             }
-            m_pendingRenderbufferClears.erase(renderbuffer);
+            m_pendingRenderbufferClears.erase(it->first);
+            resource.Destroy(m_device, m_allocator);
+            it = m_renderbufferResources.erase(it);
         }
     }
 
@@ -270,9 +325,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.samples != renderbuffer->GetSamples();
         if (!needsCreate) {
             resource.renderbuffer = renderbuffer;
+            // A new renderbuffer at a recycled address may adopt a compatible entry that
+            // was already stamped dead; it is alive again, so cancel the aging.
+            resource.deadSinceFrame = RenderbufferResource::kNeverObservedDead;
             return &resource;
         }
 
+        // Respecify: park the old backing for aged destruction instead of destroying
+        // inline - it may still be referenced by in-flight command buffers.
+        DeferRenderbufferBackingRelease(resource);
         resource.Destroy(m_device, m_allocator);
         resource.renderbuffer = renderbuffer;
 
@@ -1178,6 +1239,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void VkRenderPassManager::OnPresent() {
         ++m_frameCounter;
 
+        // Runs every frame boundary, ahead of the render-pass sweep gate below: the walk
+        // is O(#renderbuffer resources) — single digits in practice — and per-frame
+        // invocation keeps dead-resource reclaim latency at the aging bound instead of
+        // coupling it to renderbuffer *use* (the GetOrCreateRenderbufferResource call
+        // site never runs again once an app stops using renderbuffers).
+        CollectRenderbufferGarbage();
+        CollectDeferredRenderbufferReleases(/*destroyAll=*/false);
+
         // Sweep occasionally; evict entries whose last use is far past every
         // in-flight frame so their VkRenderPass/VkFramebuffer can be destroyed
         // safely (RenderPassEntry's destructor releases the handles).
@@ -1187,6 +1256,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return;
         }
 
+        // Collect the dying handles and notify once after the loop: pipelines hashed
+        // on them share the entries' >kRetireAgeFrames idleness (they are only bound
+        // by draws that hit those entries), so the observer may destroy them
+        // immediately - and a single batched notification costs one pipeline-cache
+        // scan instead of one per evicted pass.
+        Vector<VkRenderPass> destroyedRenderPasses;
         const Uint64 activeHash = s_hasActiveRenderPass ? s_activeRenderPass.hash : 0;
         for (auto it = m_renderPasses.begin(); it != m_renderPasses.end();) {
             const Bool isActive = s_hasActiveRenderPass && it->first == activeHash;
@@ -1194,11 +1269,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 if (m_rpFastValid && m_rpFastRenderPassHash == it->first) {
                     m_rpFastValid = false;
                 }
+                destroyedRenderPasses.push_back(it->second.renderPass);
                 it = m_renderPasses.erase(it);
             } else {
                 ++it;
             }
         }
+        if (!destroyedRenderPasses.empty() && m_evictionObserver != nullptr) {
+            m_evictionObserver->OnRenderPassesDestroyed(destroyedRenderPasses);
+        }
     }
 
     Bool VkRenderPassManager::BeginRenderPass(VkCommandBuffer commandBuffer, RenderPassEntry& renderPassEntry) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.h

@@ -157,11 +157,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     class VkRenderPassManager {
     public:
         using HashType = Uint64;
+
+        // Notified once per OnPresent sweep with every aged-out entry's VkRenderPass
+        // value: pipelines are hashed on the raw handle, and once destroyed the value
+        // may be recycled for an incompatible pass, so dependent caches must purge
+        // everything keyed on them before any new pass can be created (the sweep and
+        // the notification run back-to-back with no creation in between; observers
+        // compare the values, never dereference them). Batched so a mass-idle cohort
+        // (shader-pack switch, dimension exit) costs the observer one pipeline-cache
+        // scan, not one per dying pass. The wholesale paths
+        // (Shutdown/RecreateSwapchain) do not notify - their callers already drop
+        // every pipeline outright.
+        class IEvictionObserver {
+        public:
+            virtual ~IEvictionObserver() = default;
+            virtual void OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) = 0;
+        };
+
         VkRenderPassManager(VkDevice device,
             VkPhysicalDevice physicalDevice, VmaAllocator allocator, const VulkanRendererConfig& config,
             VkClearManager& clearManager, VkTextureManager& textureManager, SwapchainObject& swapchainObject);
         ~VkRenderPassManager();
 
+        // Observer may be null (no notifications). Not owned.
+        void SetEvictionObserver(IEvictionObserver* observer) { m_evictionObserver = observer; }
+
         Bool Initialize();
         void Shutdown();
 
@@ -192,6 +212,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         UnorderedMap<Uint64, RenderPassEntry> m_renderPasses;
         // Monotonic frame counter (bumped in OnPresent) for render-pass cache aging.
         Uint64 m_frameCounter = 0;
+        IEvictionObserver* m_evictionObserver = nullptr;
 
         // Bumped whenever a renderbuffer VkImage is (re)created; together with the texture
         // manager's image epoch this invalidates the render-pass fast path on any attachment
@@ -213,6 +234,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     public:
         struct RenderbufferResource {
+            // deadSinceFrame sentinel: the owning weak reference has not been observed
+            // expired. Dead resources age past every in-flight frame before Destroy
+            // (see CollectRenderbufferGarbage); the GPU may still reference the image
+            // for frames-in-flight frames after the GL object dies.
+            static constexpr Uint64 kNeverObservedDead = UINT64_MAX;
+
             WeakPtr<MG_State::GLState::RenderbufferObject> renderbuffer;
             VkImage image = VK_NULL_HANDLE;
             VmaAllocation allocation = nullptr;
@@ -224,6 +251,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
             TextureInternalFormat internalFormat = TextureInternalFormat::Unknown;
             Int samples = 0;
+            // m_frameCounter value at which the weak reference was first seen expired.
+            Uint64 deadSinceFrame = kNeverObservedDead;
 
             void Destroy(VkDevice device, VmaAllocator allocator);
         };
@@ -241,12 +270,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             ClearAttachmentPayload payload{};
         };
 
+        // A superseded renderbuffer backing (glRenderbufferStorage respecify) parked
+        // until enough frame boundaries have passed that no in-flight command buffer
+        // can still reference it; destroyed in OnPresent (see RetireAgeFrames).
+        struct DeferredRenderbufferRelease {
+            VkImage image = VK_NULL_HANDLE;
+            VmaAllocation allocation = nullptr;
+            VkImageView view = VK_NULL_HANDLE;
+            Uint64 deferredAtFrame = 0;
+        };
+
         UnorderedMap<MG_State::GLState::RenderbufferObject*, RenderbufferResource> m_renderbufferResources;
         UnorderedMap<MG_State::GLState::RenderbufferObject*, PendingRenderbufferClear> m_pendingRenderbufferClears;
+        Vector<DeferredRenderbufferRelease> m_deferredRenderbufferReleases;
 
         Bool HasPendingRenderbufferClear(
             const MG_State::GLState::FramebufferAttachmentObject& attachment) const;
         void CollectRenderbufferGarbage();
+        // Frame-boundary margin after which a resource last referenced by a retired
+        // GL object (or superseded backing) is provably past every in-flight frame.
+        Uint64 RetireAgeFrames() const;
+        void DeferRenderbufferBackingRelease(RenderbufferResource& resource);
+        void CollectDeferredRenderbufferReleases(Bool destroyAll);
 
         static inline XXH64_state_t* m_hashState = XXH64_createState();
         static inline ActiveRenderPassInfo s_activeRenderPass{};

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.cpp

@@ -51,6 +51,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Float ResolveEffectiveMinLod(const MG_State::GLState::SamplerObject& sampler, Float effectiveMaxLod) {
             return std::min(sampler.GetMinLod(), effectiveMaxLod);
         }
+
+        // A single-level view can only ever deliver the base level, but the LOD clamp must not be
+        // collapsed to exactly 0: both GL and Vulkan pick magFilter over minFilter from the
+        // *clamped* lambda, so maxLod = 0 would make every fragment magnify and quietly retire the
+        // min filter. 0.25 is the value VkSamplerCreateInfo's own note prescribes for emulating
+        // GL's non-mipmapped minification - large enough for lambda to stay positive, small enough
+        // that a NEAREST mip mode still rounds down to level 0. Clamped rather than assigned, so a
+        // texture whose GL_TEXTURE_MAX_LOD really is 0 keeps magnifying as GL says it must.
+        Float ResolveSingleLevelMaxLod(const MG_State::GLState::SamplerObject& sampler, Bool singleLevelView) {
+            const Float maxLod = ResolveEffectiveMaxLod(sampler);
+            return singleLevelView ? std::min(maxLod, 0.25f) : maxLod;
+        }
     } // namespace
 
     Bool VkSamplerManager::Initialize(const InitInfo& initInfo) {
@@ -89,15 +101,43 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         m_device = VK_NULL_HANDLE;
         m_config = nullptr;
+        m_frameBoundaryCounter = 0;
+    }
+
+    void VkSamplerManager::OnFrameBoundary() {
+        ++m_frameBoundaryCounter;
+
+        // Sweep occasionally; destroy samplers whose last use is far past every
+        // in-flight frame. Destroy and erase must stay atomic, or Shutdown would
+        // double-free the handle; an evicted key that recurs simply re-creates
+        // its sampler on the next miss.
+        constexpr Uint64 kSweepInterval = 256;
+        constexpr Uint64 kRetireAgeBoundaries = 1024;
+        if ((m_frameBoundaryCounter % kSweepInterval) != 0) {
+            return;
+        }
+
+        for (auto it = m_samplers.begin(); it != m_samplers.end();) {
+            auto& entry = it->second;
+            if (m_frameBoundaryCounter - entry.lastUsedFrameBoundary > kRetireAgeBoundaries) {
+                if (m_device != VK_NULL_HANDLE && entry.handle != VK_NULL_HANDLE) {
+                    vkDestroySampler(m_device, entry.handle, nullptr);
+                }
+                it = m_samplers.erase(it);
+            } else {
+                ++it;
+            }
+        }
     }
 
     Uint64 VkSamplerManager::BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
                                              const MG_State::GLState::ITextureObject& texture,
-                                             Bool forceNearestFiltering) const {
+                                             Bool forceNearestFiltering, Bool singleLevelView) const {
         MOBILEGL_ASSERT(m_config != nullptr, "VkSamplerManager::BuildSamplerKey: m_config is null");
         XXHASH_VERIFY(XXH64_reset(m_hashState, m_config->CacheVersion));
 
         XXHASH_VERIFY(XXH64_update(m_hashState, &forceNearestFiltering, sizeof(forceNearestFiltering)));
+        XXHASH_VERIFY(XXH64_update(m_hashState, &singleLevelView, sizeof(singleLevelView)));
 
         const auto minFilter = sampler.GetMinFilter();
         XXHASH_VERIFY(XXH64_update(m_hashState, &minFilter, sizeof(minFilter)));
@@ -111,7 +151,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         XXHASH_VERIFY(XXH64_update(m_hashState, &wrapT, sizeof(wrapT)));
         const auto wrapR = sampler.GetWrapR();
         XXHASH_VERIFY(XXH64_update(m_hashState, &wrapR, sizeof(wrapR)));
-        const auto maxLod = ResolveEffectiveMaxLod(sampler);
+        const auto maxLod = ResolveSingleLevelMaxLod(sampler, singleLevelView);
         const auto minLod = ResolveEffectiveMinLod(sampler, maxLod);
         XXHASH_VERIFY(XXH64_update(m_hashState, &minLod, sizeof(minLod)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &maxLod, sizeof(maxLod)));
@@ -133,10 +173,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     VkSampler VkSamplerManager::GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
                                                    const MG_State::GLState::ITextureObject& texture,
-                                                   Bool forceNearestFiltering) {
-        const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering);
+                                                   Bool forceNearestFiltering, Uint32 viewLevelCount) {
+        // A view that exposes a single mip level has no second level to blend with, so GL's
+        // *_MIPMAP_* minification filters degenerate to plain filtering on the base level -
+        // sampling is unchanged by pinning the Vulkan sampler to NEAREST mip mode at LOD 0.
+        // It is not cosmetic: MobileGL backs such a view with a fully allocated mip chain whose
+        // tail is never written, and a LINEAR mip mode lets the texture unit issue the level+1
+        // fetch anyway. On Adreno that fetch lands in uninitialized UBWC pages (or past the
+        // allocation for a genuinely single-level image) and faults the GPU - the same failure
+        // the default-framebuffer blit shader had to work around with an explicit-LOD sample.
+        const Bool singleLevelView = viewLevelCount == 1;
+        const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering, singleLevelView);
         auto it = m_samplers.find(key);
         if (it != m_samplers.end()) {
+            it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
             return it->second.handle;
         }
 
@@ -144,7 +194,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         samplerInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
         samplerInfo.magFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMagFilter());
         samplerInfo.minFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMinFilter());
-        samplerInfo.mipmapMode = forceNearestFiltering ? VK_SAMPLER_MIPMAP_MODE_NEAREST
+        samplerInfo.mipmapMode = (forceNearestFiltering || singleLevelView)
+                                     ? VK_SAMPLER_MIPMAP_MODE_NEAREST
                                      : ToVkMipmapMode(sampler.GetMipmapMode());
         samplerInfo.addressModeU = ToVkAddressMode(sampler.GetWrapS());
         samplerInfo.addressModeV = ToVkAddressMode(sampler.GetWrapT());
@@ -157,7 +208,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         samplerInfo.maxAnisotropy = maxAnisotropy;
         samplerInfo.compareEnable = sampler.GetCompareMode() == SamplerCompareMode::CompareToTexture ? VK_TRUE : VK_FALSE;
         samplerInfo.compareOp = ToVkCompareOp(ResolveCompareFunc(sampler, texture));
-        samplerInfo.maxLod = ResolveEffectiveMaxLod(sampler);
+        // Must match BuildSamplerKey's resolution exactly.
+        samplerInfo.maxLod = ResolveSingleLevelMaxLod(sampler, singleLevelView);
         samplerInfo.minLod = ResolveEffectiveMinLod(sampler, samplerInfo.maxLod);
         samplerInfo.borderColor = ResolveVkBorderColor(sampler, texture);
         samplerInfo.unnormalizedCoordinates = VK_FALSE;
@@ -169,6 +221,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         entry.handle = vkSampler;
         entry.externalIndex = sampler.GetExternalIndex();
         entry.version = sampler.GetVersion();
+        entry.lastUsedFrameBoundary = m_frameBoundaryCounter;
         m_samplers[key] = entry;
         return vkSampler;
     }

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.h

@@ -33,20 +33,38 @@ public:
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
 
+    // viewLevelCount is the mip-level count of the image view this sampler will be paired
+    // with; 0 means "unknown, do not narrow". See GetOrCreateSampler for why it matters.
     VkSampler GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
                                  const MG_State::GLState::ITextureObject& texture,
-                                 Bool forceNearestFiltering = false);
+                                 Bool forceNearestFiltering = false,
+                                 Uint32 viewLevelCount = 0);
+    // Frame boundary hook: ages the sampler cache and destroys samplers not used
+    // for many frames. The key hashes continuous float state (lodBias, LOD clamps,
+    // anisotropy), so an app animating those would otherwise mint an unbounded
+    // stream of never-destroyed VkSamplers and eventually exhaust the device's
+    // maxSamplerAllocationCount. A sampler idle for over a thousand frame
+    // boundaries cannot be referenced by any in-flight command buffer (frames in
+    // flight are single digits), and every descriptor set the GPU consumes is
+    // written that same frame with live handles (the per-binding resolve memo and
+    // descriptor-set reuse are both frame-reset), so destruction here needs no
+    // fence wait. Self-gated: one counter bump and compare except on sweep
+    // boundaries.
+    void OnFrameBoundary();
 
 private:
     struct SamplerCacheEntry {
         VkSampler handle = VK_NULL_HANDLE;
         Uint externalIndex = 0;
         Uint16 version = 0;
+        // Frame boundary of the last cache hit; entries idle past the
+        // OnFrameBoundary retirement age have their VkSampler destroyed.
+        Uint64 lastUsedFrameBoundary = 0;
     };
 
     Uint64 BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
                            const MG_State::GLState::ITextureObject& texture,
-                           Bool forceNearestFiltering) const;
+                           Bool forceNearestFiltering, Bool singleLevelView) const;
     static VkFilter ToVkFilter(SamplerFilterMode mode);
     static VkSamplerMipmapMode ToVkMipmapMode(SamplerMipmapMode mode);
     static VkSamplerAddressMode ToVkAddressMode(SamplerWrapMode mode);
@@ -67,6 +85,8 @@ private:
     Bool m_samplerAnisotropySupported = false;
     Float m_maxSamplerAnisotropy = 1.0f;
     UnorderedMap<Uint64, SamplerCacheEntry> m_samplers;
+    // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+    Uint64 m_frameBoundaryCounter = 0;
     static inline XXH64_state_t* m_hashState = XXH64_createState();
 };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -587,6 +587,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_allocator = initInfo.allocator;
         m_commandPool = initInfo.commandPool;
         m_graphicsQueue = initInfo.graphicsQueue;
+        m_imageFormatListSupported = initInfo.imageFormatListSupported;
         m_currentFrameIndex = 0;
         m_deferredReleases.clear();
         m_deferredReleases.resize(initInfo.frameCount);
@@ -609,6 +610,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         DestroyDeferredReleases();
         m_textureResources.clear();
         m_aliveObjects.clear();
+        m_storageImageTextures.clear();
 
         m_device = VK_NULL_HANDLE;
         m_physicalDevice = VK_NULL_HANDLE;
@@ -627,6 +629,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         frameIndex, m_deferredViewReleases.size());
         m_currentFrameIndex = frameIndex;
         CollectDeferredReleases(frameIndex);
+
+        // Frame-boundary GC: every 64 frame boundaries (~1 s at 60 fps) bounds the reclaim
+        // latency for dead textures regardless of draw traffic — workloads that churn
+        // textures through clears/readbacks alone never reach the draw-gated
+        // CollectGarbage. Must run after CollectDeferredReleases above: the prune defers
+        // its releases into this frame's slot, which was just drained, so they are
+        // destroyed only after the slot's fence has been waited again one full frame-ring
+        // cycle from now (never while an in-flight frame may still reference them).
+        constexpr Uint32 kGcFrameInterval = 64;
+        ++m_gcFrameCounter;
+        if (m_gcFrameCounter % kGcFrameInterval == 0) {
+            PruneDeadTextures();
+        }
+    }
+
+    void VkTextureManager::CollectAllDeferredReleases() {
+        const SizeT frameCount = std::min(m_deferredReleases.size(), m_deferredViewReleases.size());
+        for (SizeT frameIndex = 0; frameIndex < frameCount; ++frameIndex) {
+            CollectDeferredReleases(static_cast<Uint32>(frameIndex));
+        }
     }
 
     void VkTextureManager::EraseTrackedTexture(const TextureIdentity& identity) {
@@ -636,6 +658,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_textureResources.erase(resourceIt);
         }
         m_aliveObjects.erase(identity);
+        m_storageImageTextures.erase(identity);
     }
 
     void VkTextureManager::PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture) {
@@ -702,9 +725,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // construction introduces a new identity. Doing this unconditionally made every
         // sampled-texture sync scan the entire alive-texture map per draw.
         if (aliveIt == m_aliveObjects.end()) {
+            WeakPtr<MG_State::GLState::ITextureObject> aliveTexture;
             const auto& liveTexture = MG_State::pGLContext->GetTextureObject(texture.GetExternalIndex());
             if (liveTexture && liveTexture.get() == &texture) {
-                m_aliveObjects[identity] = WeakPtr<MG_State::GLState::ITextureObject>(liveTexture);
+                aliveTexture = liveTexture;
+            } else {
+                // The name lookup legally fails while the object is alive: the name was
+                // deleted with the texture still attached to an FBO (the attachment's
+                // SharedPtr keeps it alive), or the name was reused by a new texture, or
+                // this is a default texture object (name 0 lives outside the name map).
+                // Register through the object's own control block so the resource created
+                // below still participates in weak-expiry GC instead of becoming an
+                // orphan no reclamation path can reach until Shutdown.
+                aliveTexture = texture.weak_from_this();
+            }
+            if (!aliveTexture.expired()) {
+                m_aliveObjects[identity] = Move(aliveTexture);
                 PruneStaleTextureAliases(&texture);
             }
         }
@@ -1156,8 +1192,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ok;
     }
 
+    void VkTextureManager::MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture) {
+        m_storageImageTextures.insert(MakeTextureIdentity(&texture));
+    }
+
+    Bool VkTextureManager::NeedsStorageUsageUpgrade(MG_State::GLState::ITextureObject& texture) const {
+        const TextureIdentity identity = MakeTextureIdentity(&texture);
+        if (m_storageImageTextures.find(identity) == m_storageImageTextures.end()) {
+            return false;
+        }
+        const auto it = m_textureResources.find(identity);
+        // No image yet: the first sync creates it with STORAGE straight away, so there is nothing
+        // to preserve and nothing to order against.
+        return it != m_textureResources.end() && it->second.image != VK_NULL_HANDLE &&
+               !it->second.storageUsageResolved;
+    }
+
     Bool VkTextureManager::NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const {
-        const auto it = m_textureResources.find(MakeTextureIdentity(&texture));
+        const TextureIdentity identity = MakeTextureIdentity(&texture);
+        const auto it = m_textureResources.find(identity);
         if (it == m_textureResources.end()) {
             return true;
         }
@@ -1165,6 +1218,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (resource.image == VK_NULL_HANDLE || resource.layout != VK_IMAGE_LAYOUT_GENERAL) {
             return true;
         }
+        // The image predates this texture's first image-unit binding, so it was created without
+        // STORAGE usage and has to be recreated - which is illegal inside a render pass.
+        if (!resource.storageUsageResolved &&
+            m_storageImageTextures.find(identity) != m_storageImageTextures.end()) {
+            return true;
+        }
         // Mirror SyncTexture's cross-draw skip condition: any version drift means the sync
         // path may upload or rebuild, both of which need the render pass ended first.
         const auto* mipTexture = MG_State::GLState::AsMipmapTexture(&texture);
@@ -1212,10 +1271,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     SizeT VkTextureManager::CollectGarbage() {
+        // Draw-gated stagger (1 in 256 calls): keeps the per-draw cost at one counter
+        // bump. The guaranteed reclaim path is the frame-boundary prune in BeginFrame;
+        // this remains as a cheap assist so draw-heavy workloads reclaim sooner.
         m_gcCounter++;
         if (m_gcCounter != 0) {
             return 0;
         }
+        return PruneDeadTextures();
+    }
+
+    SizeT VkTextureManager::PruneDeadTextures() {
+        // Erasing entries would dangle the raw TextureResource pointers memoized for the
+        // current draw; every call path (BeginFrame, and CollectGarbage at the top of a
+        // freshly opened draw-sync scope) runs before any memo entry is recorded.
+        MOBILEGL_ASSERT(m_drawSyncedThisDraw.empty(),
+                        "PruneDeadTextures: draw-sync memo holds raw resource pointers an erase would dangle");
 
         Vector<MG_State::GLState::ITextureObject*> expiredTextures;
         expiredTextures.reserve(m_aliveObjects.size());
@@ -1227,7 +1298,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         for (auto* texture : expiredTextures) {
             PruneStaleTextureAliases(texture);
         }
-        return expiredTextures.size();
+        SizeT prunedCount = expiredTextures.size();
+
+        // Orphan sweep: after the pass above, m_aliveObjects holds only live entries.
+        // Registration in SyncTextureAndGetDescriptor cannot fail for a SharedPtr-owned
+        // texture (weak_from_this fallback), so a resource whose identity has no alive
+        // entry has no trackable owner: its GL-side object is gone, or was never
+        // shared-owned, in which case recreation on a later sync is the safe fallback.
+        // Destruction goes through the per-frame deferred queues, never immediate.
+        Vector<TextureIdentity> orphanIdentities;
+        for (auto it = m_textureResources.begin(); it != m_textureResources.end(); ++it) {
+            if (m_aliveObjects.find(it->first) == m_aliveObjects.end()) {
+                orphanIdentities.emplace_back(it->first);
+            }
+        }
+        for (const auto& identity : orphanIdentities) {
+            EraseTrackedTexture(identity);
+        }
+        prunedCount += orphanIdentities.size();
+        return prunedCount;
     }
 
     Bool VkTextureManager::SyncTexture(MG_State::GLState::ITextureObject &texture,
@@ -1241,7 +1330,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const auto* syncingMipTexture = MG_State::GLState::AsMipmapTexture(&texture);
         const Uint32 syncingMipLevelCount =
             syncingMipTexture != nullptr ? syncingMipTexture->GetMipmapLevelCount() : 0u;
-        if (outResource.image != VK_NULL_HANDLE &&
+        // A pending storage-usage upgrade also has to bust the skip: nothing about the texture's
+        // content or params changed, but the image itself must be recreated with STORAGE usage
+        // before it can back an image-unit descriptor.
+        const Bool storageUpgradePending =
+            !outResource.storageUsageResolved &&
+            m_storageImageTextures.find(MakeTextureIdentity(&texture)) != m_storageImageTextures.end();
+        if (outResource.image != VK_NULL_HANDLE && !storageUpgradePending &&
             outResource.syncedContentVersion == syncingContentVersion &&
             outResource.syncedTextureParamsVersion == texture.GetTextureParamsVersion() &&
             outResource.syncedMipLevelCount == syncingMipLevelCount) {
@@ -1352,16 +1447,44 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkImageAspectFlags aspect = GetAspectMaskForFormat(format);
         VkFormatProperties formatProperties{};
         vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
-        const Bool supportsStorageImage =
+        // Only textures that have actually been bound to a GL image unit get STORAGE usage (and
+        // the MUTABLE_FORMAT it drags in for format-reinterpreting image views). Requesting it
+        // for every storage-capable colour texture costs real bandwidth: Adreno cannot keep UBWC
+        // compression on an image that may be written through a storage descriptor, so the whole
+        // render target - MC's included - runs uncompressed. MarkStorageImageTexture upgrades a
+        // texture before its first image-unit draw, and the usage below feeds the compatibility
+        // check so the upgrade recreates the image.
+        const Bool markedAsStorageImage =
+            m_storageImageTextures.find(MakeTextureIdentity(
+                const_cast<MG_State::GLState::ITextureObject*>(&texture))) != m_storageImageTextures.end();
+        // Storage-image CAPABILITY (does the format allow it at all) is deliberately separate from
+        // whether this texture actually needs the usage. MUTABLE_FORMAT keys off capability, as
+        // before: format-reinterpreting views are not a storage-only concern - the SAMPLED path
+        // needs them too (GetOrCreateSampledImageView bails out without it, see ~line 892), so
+        // tying MUTABLE_FORMAT to the image-unit mark would break sampled format reinterpretation
+        // for every texture that never becomes a storage image.
+        const Bool storageImageCapable =
             !isMultisampleTexture &&
             (aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0 &&
             (formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT) != 0;
+        const Bool supportsStorageImage = storageImageCapable && markedAsStorageImage;
         VkImageCreateFlags imageCreateFlags = shapeInfo.imageFlags;
-        if (supportsStorageImage && IsMutableStorageImageFormat(format) &&
+        if (storageImageCapable && IsMutableStorageImageFormat(format) &&
             m_mutableFormatUnsupported.find(format) == m_mutableFormatUnsupported.end()) {
             imageCreateFlags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
         }
 
+        VkImageUsageFlags desiredUsage =
+            VK_IMAGE_USAGE_SAMPLED_BIT |
+            (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
+            ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
+            (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT)) ?
+                 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT :
+                 0);
+        if (!isMultisampleTexture) {
+            desiredUsage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+        }
+
         const Bool compatible = resource.image != VK_NULL_HANDLE && resource.format == format &&
                                 resource.extent.width == static_cast<Uint32>(texelSize.x()) &&
                                 resource.extent.height == static_cast<Uint32>(texelSize.y()) &&
@@ -1370,6 +1493,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                 resource.viewType == shapeInfo.viewType &&
                                 resource.sampleCount == resolvedSampleCount &&
                                 resource.imageCreateFlags == imageCreateFlags &&
+                                resource.usageFlags == desiredUsage &&
                                 resource.mipLevels == backingMipLevels;
         if (compatible) {
             if (resource.perMipViews.size() != backingMipLevels) {
@@ -1378,6 +1502,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (resource.perMipSampledViews.size() != backingMipLevels) {
                 resource.perMipSampledViews.resize(backingMipLevels, VK_NULL_HANDLE);
             }
+            // Keeping the image is itself the answer to the mark: either it already carries
+            // STORAGE, or this format can never carry it. Either way there is nothing left to
+            // recreate, so stop reporting the texture as needing preparation.
+            resource.storageUsageResolved = markedAsStorageImage;
             return true;
         }
 
@@ -1392,7 +1520,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.sampleCount == resolvedSampleCount &&
             resource.imageCreateFlags == imageCreateFlags &&
             resolvedSampleCount == VK_SAMPLE_COUNT_1_BIT &&
-            resource.mipLevels < backingMipLevels &&
+            // '<=' rather than '<': a storage-usage upgrade recreates the image with an
+            // unchanged mip count, and its contents (a render target's pixels live only on the
+            // GPU) still have to survive. The vkCmdCopyImage below copies min(mipLevels).
+            resource.mipLevels <= backingMipLevels &&
             resource.layout != VK_IMAGE_LAYOUT_UNDEFINED;
 
         std::unique_ptr<TextureResource> preservedResource;
@@ -1414,16 +1545,37 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         imageInfo.format = format;
         imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
         imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
-        imageInfo.usage = VK_IMAGE_USAGE_SAMPLED_BIT |
-                          (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
-                          ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
-                          (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT)) ?
-                               VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT :
-                               0);
-        if (!isMultisampleTexture) {
-            imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
-        }
+        imageInfo.usage = desiredUsage;
         imageInfo.samples = resolvedSampleCount;
+
+        // Bound the mutability. A blindly-mutable image has to be laid out so that ANY format in
+        // its compatibility class can be viewed, which costs bandwidth compression on tilers;
+        // naming the exact set instead lets the driver keep it. Only safe when that set really is
+        // exhaustive, so it is restricted to textures that are not image-unit bound: sampled views
+        // can only ever ask for ResolveSampledImageViewFormat's output, whereas glBindImageTexture
+        // may name any compatible format, which nothing here can enumerate ahead of time.
+        Vector<VkFormat> viewFormats;
+        VkImageFormatListCreateInfo formatListInfo{};
+        if (m_imageFormatListSupported && !supportsStorageImage &&
+            (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+            viewFormats.push_back(format);
+            for (const SamplerNumericDomain domain : {SamplerNumericDomain::Float,
+                                                      SamplerNumericDomain::SignedInteger,
+                                                      SamplerNumericDomain::UnsignedInteger}) {
+                const VkFormat viewFormat = ResolveSampledImageViewFormat(format, domain);
+                if (viewFormat == VK_FORMAT_UNDEFINED) {
+                    continue;
+                }
+                if (std::find(viewFormats.begin(), viewFormats.end(), viewFormat) == viewFormats.end()) {
+                    viewFormats.push_back(viewFormat);
+                }
+            }
+            formatListInfo.sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO;
+            formatListInfo.viewFormatCount = static_cast<Uint32>(viewFormats.size());
+            formatListInfo.pViewFormats = viewFormats.data();
+            imageInfo.pNext = &formatListInfo;
+        }
+
         if (isMultisampleTexture || (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
             VkImageFormatProperties imageFormatProperties{};
             VkResult imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
@@ -1480,6 +1632,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         resource.viewType = shapeInfo.viewType;
         resource.sampleCount = resolvedSampleCount;
         resource.imageCreateFlags = imageCreateFlags;
+        resource.usageFlags = imageInfo.usage;
+        resource.storageUsageResolved = markedAsStorageImage;
         resource.syncedTextureParamsVersion = 0;
 
         if (preservedResource) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -53,6 +53,9 @@ public:
         VkCommandPool commandPool = VK_NULL_HANDLE;
         VkQueue graphicsQueue = VK_NULL_HANDLE;
         Uint32 frameCount = 0;
+        // VK_KHR_image_format_list is enabled: MUTABLE_FORMAT images can name the exact set of
+        // formats they will be viewed as, which is what lets a tiler keep them compressed.
+        Bool imageFormatListSupported = false;
     };
 
     struct TextureResource {
@@ -157,6 +160,17 @@ public:
         VkImageViewType viewType = VK_IMAGE_VIEW_TYPE_2D;
         VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
         VkImageCreateFlags imageCreateFlags = 0;
+        // Usage the live image was created with. STORAGE is only requested for textures that
+        // have actually been bound to a GL image unit, because on Adreno a storage-capable
+        // image loses UBWC bandwidth compression; a later image binding upgrades the usage
+        // and recreates the image, so the resolved usage has to be part of the compatibility
+        // check that decides whether the existing image can be kept.
+        VkImageUsageFlags usageFlags = 0;
+        // True once this image was (re)resolved while the texture was already marked as an
+        // image-unit texture. Distinguishes "not upgraded yet" from "cannot be upgraded"
+        // (a format whose optimalTilingFeatures lack STORAGE_IMAGE never gains the bit), so
+        // NeedsStorageImagePreparation cannot ask for a recreate that will never happen.
+        Bool storageUsageResolved = false;
         Uint16 syncedTextureParamsVersion = 0;
         // Snapshot of ITextureObject::GetContentVersion() at the last successful sync;
         // lets SyncTexture skip the whole re-check/re-upload when content is unchanged.
@@ -190,6 +204,8 @@ public:
             std::swap(this->viewType, that.viewType);
             std::swap(this->sampleCount, that.sampleCount);
             std::swap(this->imageCreateFlags, that.imageCreateFlags);
+            std::swap(this->usageFlags, that.usageFlags);
+            std::swap(this->storageUsageResolved, that.storageUsageResolved);
             std::swap(this->syncedTextureParamsVersion, that.syncedTextureParamsVersion);
             std::swap(this->syncedContentVersion, that.syncedContentVersion);
             std::swap(this->syncedMipLevelCount, that.syncedMipLevelCount);
@@ -251,6 +267,8 @@ public:
             viewType = VK_IMAGE_VIEW_TYPE_2D;
             sampleCount = VK_SAMPLE_COUNT_1_BIT;
             imageCreateFlags = 0;
+            usageFlags = 0;
+            storageUsageResolved = false;
             syncedTextureParamsVersion = 0;
             syncedContentVersion = 0;
             syncedMipLevelCount = 0;
@@ -267,6 +285,10 @@ public:
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
     void BeginFrame(Uint32 frameIndex);
+    // Drains every frame slot's deferred image/view releases. Only valid when
+    // the caller has proven every queue submission complete; used by the
+    // present-less frame-boundary drain.
+    void CollectAllDeferredReleases();
 
     TextureResource* SyncTextureAndGetDescriptor(
         MG_State::GLState::ITextureObject& texture);
@@ -285,6 +307,17 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+    // Records that this texture is bound to a GL image unit, so its image must carry
+    // VK_IMAGE_USAGE_STORAGE_BIT. Must be called before NeedsStorageImagePreparation, and
+    // therefore before the render pass is committed: an image that has to be upgraded is
+    // recreated, which is illegal inside a render pass. Sticky for the texture's lifetime -
+    // GL lets an image binding come and go, and re-creating the image every time it does
+    // would cost far more than the compression it wins back.
+    void MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture);
+    // True when this texture is marked but its live image predates the mark, i.e. the next sync
+    // will recreate it with STORAGE usage and copy the old contents forward. Callers use this to
+    // submit their pending recording first, so that copy cannot read pre-flush content.
+    Bool NeedsStorageUsageUpgrade(MG_State::GLState::ITextureObject& texture) const;
     // Non-mutating probe for the per-draw storage-image fast path: true when preparing this
     // texture as a storage image may need work that is illegal inside a render pass (resource
     // creation, dirty-content upload, or a layout transition to GENERAL). Unknown state reports
@@ -364,15 +397,20 @@ private:
     static TextureIdentity MakeTextureIdentity(MG_State::GLState::ITextureObject* texture);
     void EraseTrackedTexture(const TextureIdentity& identity);
     void PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture);
+    SizeT PruneDeadTextures();
 
     VkDevice m_device = VK_NULL_HANDLE;
     VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
     VmaAllocator m_allocator = nullptr;
     VkCommandPool m_commandPool = VK_NULL_HANDLE;
     VkQueue m_graphicsQueue = VK_NULL_HANDLE;
+    Bool m_imageFormatListSupported = false;
     Uint32 m_currentFrameIndex = 0;
 
     Uint8 m_gcCounter = 0;
+    // Frame-boundary GC gate: counts BeginFrame calls, not draws, so texture churn
+    // through non-draw paths (FBO clears, readbacks) still reaches the prune.
+    Uint32 m_gcFrameCounter = 0;
     // Active only between BeginDrawSyncScope/EndDrawSyncScope; identities of
     // textures already fully synced in the current draw (small N -> flat scan).
     Bool m_drawSyncScopeActive = false;
@@ -390,6 +428,8 @@ private:
     std::unordered_set<VkFormat> m_mutableFormatUnsupported;
     std::unordered_map<TextureIdentity, WeakPtr<MG_State::GLState::ITextureObject>, TextureIdentityHash> m_aliveObjects;
     std::unordered_map<TextureIdentity, TextureResource, TextureIdentityHash> m_textureResources;
+    // Textures that have been bound to a GL image unit (see MarkStorageImageTexture).
+    std::unordered_set<TextureIdentity, TextureIdentityHash> m_storageImageTextures;
     Vector<Vector<TextureResource>> m_deferredReleases;
     Vector<Vector<VkImageView>> m_deferredViewReleases;
 };

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -1016,7 +1016,18 @@ layout(location = 0) in vec2 vTexCoord;
 layout(location = 0) out vec4 outColor;
 
 void main() {
-    outColor = texture(uSource, vTexCoord);
+    // Explicit LOD, not texture(): a blit reads exactly the selected level, so
+    // derivative-based mip selection has no business here. It is also load-bearing:
+    // on Adreno 650 (driver 512.502) an implicit-LOD sample of this single-mip
+    // UBWC render target through the pre-rotation (ROTATE_90) mapping reads past
+    // the image's allocation - despite the sampler's maxLod=0 and a nominal 1:1
+    // texel mapping whose LOD is 0, so the driver's implicit-LOD path itself is at
+    // fault - and page-faults the GPU once the neighbouring memory is returned to
+    // the kernel (frame 2 of Minecraft 26.2's resource reload; the kernel then
+    // invalidates the context and the next submit dies with EDEADLK ->
+    // VK_ERROR_DEVICE_LOST at Present). Verified on device: texture() faults on
+    // the second frame every run, textureLod survives with identical state.
+    outColor = textureLod(uSource, vTexCoord, 0.0);
 }
 )";
 
@@ -1910,8 +1921,10 @@ void main() {
                 case VK_FORMAT_R16G16_UNORM:          out = {ReadbackSourceClass::Float, 2, 16}; return true;
                 case VK_FORMAT_R16G16B16A16_UNORM:    out = {ReadbackSourceClass::Float, 4, 16}; return true;
                 // --- SRGB (decode to linear like GL readback of sRGB textures) ---
-                case VK_FORMAT_R8G8B8A8_SRGB:         out = {ReadbackSourceClass::Float, 4, 8, false, true}; return true;
-                case VK_FORMAT_B8G8R8A8_SRGB:         out = {ReadbackSourceClass::Float, 4, 8, false, true, true}; return true;
+                // GL GetTexImage/ReadPixels of sRGB textures return the raw sRGB-encoded
+                // bytes (GL 3.3 has no FRAMEBUFFER_SRGB read decode) - do NOT linearize.
+                case VK_FORMAT_R8G8B8A8_SRGB:         out = {ReadbackSourceClass::Float, 4, 8}; return true;
+                case VK_FORMAT_B8G8R8A8_SRGB:         out = {ReadbackSourceClass::Float, 4, 8, false, false, true}; return true;
                 // --- SNORM ---
                 case VK_FORMAT_R8_SNORM:              out = {ReadbackSourceClass::Float, 1, 8, true};  return true;
                 case VK_FORMAT_R8G8_SNORM:            out = {ReadbackSourceClass::Float, 2, 8, true};  return true;
@@ -1949,6 +1962,8 @@ void main() {
                 // --- packed / special ---
                 case VK_FORMAT_A2B10G10R10_UNORM_PACK32:
                 case VK_FORMAT_A2B10G10R10_UINT_PACK32:
+                case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
+                case VK_FORMAT_A2R10G10B10_UINT_PACK32:
                 case VK_FORMAT_B10G11R11_UFLOAT_PACK32:
                 case VK_FORMAT_E5B9G9R9_UFLOAT_PACK32:
                 case VK_FORMAT_R5G6B5_UNORM_PACK16:
@@ -1958,7 +1973,8 @@ void main() {
                 case VK_FORMAT_B5G5R5A1_UNORM_PACK16:
                 case VK_FORMAT_R4G4B4A4_UNORM_PACK16:
                 case VK_FORMAT_B4G4R4A4_UNORM_PACK16:
-                    out.sourceClass = format == VK_FORMAT_A2B10G10R10_UINT_PACK32 ?
+                    out.sourceClass = (format == VK_FORMAT_A2B10G10R10_UINT_PACK32 ||
+                                       format == VK_FORMAT_A2R10G10B10_UINT_PACK32) ?
                         ReadbackSourceClass::UnsignedInt : ReadbackSourceClass::Float;
                     out.special = format;
                     return true;
@@ -2034,6 +2050,15 @@ void main() {
                     rgba[3] = static_cast<Float>((word >> 30) & 0x3u) / 3.0f;
                     return;
                 }
+                case VK_FORMAT_A2R10G10B10_UNORM_PACK32: {
+                    Uint32 word = 0;
+                    Memcpy(&word, source, sizeof(word));
+                    rgba[2] = static_cast<Float>(word & 0x3FFu) / 1023.0f;
+                    rgba[1] = static_cast<Float>((word >> 10) & 0x3FFu) / 1023.0f;
+                    rgba[0] = static_cast<Float>((word >> 20) & 0x3FFu) / 1023.0f;
+                    rgba[3] = static_cast<Float>((word >> 30) & 0x3u) / 3.0f;
+                    return;
+                }
                 case VK_FORMAT_B10G11R11_UFLOAT_PACK32: {
                     Uint32 word = 0;
                     Memcpy(&word, source, sizeof(word));
@@ -2185,6 +2210,13 @@ void main() {
                     rgba[1] = (word >> 10) & 0x3FFu;
                     rgba[2] = (word >> 20) & 0x3FFu;
                     rgba[3] = (word >> 30) & 0x3u;
+                } else if (srcFormat == VK_FORMAT_A2R10G10B10_UINT_PACK32) {
+                    Uint32 word = 0;
+                    Memcpy(&word, source, sizeof(word));
+                    rgba[2] = word & 0x3FFu;
+                    rgba[1] = (word >> 10) & 0x3FFu;
+                    rgba[0] = (word >> 20) & 0x3FFu;
+                    rgba[3] = (word >> 30) & 0x3u;
                 } else {
                     for (Int c = 0; c < desc.channels; ++c) {
                         if (desc.componentBits == 8) {
@@ -2216,8 +2248,9 @@ void main() {
         }
 
         static Bool PackReadbackToClientOrPbo(const Uint8* srcPixels, VkFormat srcFormat, GLsizei width,
-                                              GLsizei height, GLenum format, GLenum type, void* pixels) {
-            if (width <= 0 || height <= 0) {
+                                              GLsizei sliceHeight, GLsizei sliceCount, GLenum format, GLenum type,
+                                              void* pixels, Bool applyPackImageParams) {
+            if (width <= 0 || sliceHeight <= 0 || sliceCount <= 0) {
                 return true;
             }
 
@@ -2230,7 +2263,8 @@ void main() {
 
             Vector<Uint8> wide;
             GLenum wideType = GL_FLOAT;
-            if (!DecodeReadbackRowsToWide(srcPixels, srcFormat, width, height, wide, wideType)) {
+            if (!DecodeReadbackRowsToWide(srcPixels, srcFormat, width,
+                                          sliceHeight * sliceCount, wide, wideType)) {
                 MGLOG_E("DirectVulkan readback skipped: unsupported source format=%d",
                         static_cast<Int>(srcFormat));
                 return false;
@@ -2243,9 +2277,9 @@ void main() {
                 return false;
             }
 
-            return DirectGLES::ReadbackImpl::StoreWideRowsToClient(wide.data(), wideType, width, height,
-                                                                   /*sliceCount=*/1, mapping, type, pixels,
-                                                                   /*applyPackImageParams=*/false);
+            return DirectGLES::ReadbackImpl::StoreWideRowsToClient(wide.data(), wideType, width, sliceHeight,
+                                                                   sliceCount, mapping, type, pixels,
+                                                                   applyPackImageParams);
         }
     } // namespace
 
@@ -2457,7 +2491,7 @@ void main() {
         MOBILEGL_ASSERT(m_textureManager != nullptr, "VkTextureManager creation failed.");
         succeeded = m_textureManager->Initialize(
             {m_device, m_physicalDevice.handle, m_allocator, m_commandPool, m_graphicsQueue,
-             m_frameContext.GetFrameCount()});
+             m_frameContext.GetFrameCount(), m_imageFormatListExtensionEnabled});
         MOBILEGL_ASSERT(succeeded, "VkTextureManager initialization failed.");
         m_clearManager = MakeUnique<VkClearManager>();
         MOBILEGL_ASSERT(m_clearManager != nullptr, "VkClearManager creation failed.");
@@ -2506,6 +2540,11 @@ void main() {
                                                       m_shaderDrawParametersFeatureEnabled,
                                                       m_unformattedFloatStorageImagesEnabled);
         MOBILEGL_ASSERT(m_programFactory != nullptr, "ProgramFactory creation failed.");
+        // Aging evictions (render passes and program entries) must purge the dependent
+        // pipeline / compute-pipeline / descriptor-set caches in the same step; both
+        // sweeps only run from the frame-boundary seams, long after initialization.
+        m_renderPassManager->SetEvictionObserver(this);
+        m_programFactory->SetEvictionObserver(this);
 
         m_samplerManager = MakeUnique<VkSamplerManager>();
         MOBILEGL_ASSERT(m_samplerManager != nullptr, "VkSamplerManager creation failed.");
@@ -2528,15 +2567,32 @@ void main() {
         MOBILEGL_ASSERT(m_vertexInputStateFactory != nullptr, "VertexInputStateFactory creation failed.");
 
         // Prime the first frame so Render() always targets an acquired swapchain image.
+        // A zero-area window (GLFW's hidden helper window during the WGL bootstrap, or a
+        // window that is already minimized) legitimately yields no swapchain here; defer
+        // the first acquire to Present in that case instead of acquiring from a null
+        // swapchain handle.
+        if (m_swapchainObject.GetHandle() != VK_NULL_HANDLE) {
             VkResult acquireResult =
                 m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
-        if (acquireResult == VK_ERROR_OUT_OF_DATE_KHR || acquireResult == VK_SUBOPTIMAL_KHR) {
+            if (acquireResult == VK_ERROR_OUT_OF_DATE_KHR) {
+                // Nothing was acquired and no semaphore signal was armed, so
+                // rebuilding and re-acquiring on the same semaphore is safe.
                 MGLOG_D("Initialize, vkAcquireNextImageKHR got %d, recreating swapchain", acquireResult);
                 RecreateSwapchain();
                 acquireResult =
                     m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
+            } else if (acquireResult == VK_SUBOPTIMAL_KHR) {
+                // The image is usable, and its acquire signal is already armed on
+                // imageAvailableSemaphore. Re-acquiring here would arm a second signal on a
+                // binary semaphore whose first one nobody has waited on yet; keep the image.
+                // Only a real surface change schedules a rebuild.
+                m_swapchainResizeRequested = m_swapchainResizeRequested || SwapchainIsOutOfDate();
+                acquireResult = VK_SUCCESS;
             }
             VK_VERIFY(acquireResult, "Initialize, WaitAndAcquireNextImage");
+        } else {
+            MGLOG_W("DirectVulkan: no swapchain at initialization (zero-area window); deferring first acquire");
+        }
         m_textureManager->BeginFrame(m_frameContext.GetCurrentFrameIndex());
         m_bufferManager.BeginFrame(m_frameContext.GetCurrentFrameIndex());
         m_convertedVertexStreams.clear();
@@ -2558,6 +2614,15 @@ void main() {
         DestroyDeferredDepthMipmapCleanup();
         DestroyComputePipelines();
 
+        // No sweep runs during teardown, but the observers point at this renderer
+        // and the factories die at different times below; disconnect them first.
+        if (m_renderPassManager) {
+            m_renderPassManager->SetEvictionObserver(nullptr);
+        }
+        if (m_programFactory) {
+            m_programFactory->SetEvictionObserver(nullptr);
+        }
+
         m_pipelineFactory.reset();
         ShutdownBlitResources();
         ShutdownDepthMipmapResources();
@@ -2648,6 +2713,7 @@ void main() {
             DestroyDebugMessenger();
             m_debugMessenger = VK_NULL_HANDLE;
         }
+        DestroyDebugReportCallback();
 
         if (m_instance != VK_NULL_HANDLE) {
             vkDestroyInstance(m_instance, nullptr);
@@ -4127,7 +4193,7 @@ void main() {
     }
 
     Bool VulkanRenderer::PrepareStorageImageTextures(
-        VkCommandBuffer commandBuffer,
+        FrameContext::FrameData& frame,
         const MG_State::GLState::ProgramObject& program,
         const ProgramFactory::VkProgramObject& programObj) {
         if (!programObj.hasStorageImages) {
@@ -4148,9 +4214,18 @@ void main() {
         // keep the render pass alive instead of splitting it on every storage-image draw (on
         // tiled GPUs each split is a full tile load/store). GL makes cross-draw image-store
         // coherence the app's job (glMemoryBarrier), so no implicit barrier is owed here.
-        Bool anyNeedsPreparation = false;
+        // Record every image-unit binding before probing anything: a texture whose image was
+        // created without STORAGE usage (the default - it costs UBWC compression on Adreno)
+        // needs a recreate, and the probe below is what ends the render pass so that recreate
+        // lands here rather than mid-pass. This cannot be folded into the probe loop, which
+        // stops at the first texture that needs work and would leave the rest unmarked.
         for (auto* texture : storageTextures) {
             MOBILEGL_ASSERT(texture != nullptr, "%s: collected a null storage texture", __func__);
+            m_textureManager->MarkStorageImageTexture(*texture);
+        }
+
+        Bool anyNeedsPreparation = false;
+        for (auto* texture : storageTextures) {
             if (m_textureManager->NeedsStorageImagePreparation(*texture) ||
                 m_clearManager->HasPendingClear(texture)) {
                 anyNeedsPreparation = true;
@@ -4161,21 +4236,51 @@ void main() {
             return true;
         }
 
+        // A first-time storage-usage upgrade recreates the image and carries the old contents
+        // forward with an out-of-band, immediately-submitted copy (PreserveTextureContentsOnRecreate).
+        // Whatever this frame already recorded into the old image is still sitting unsubmitted in
+        // this command buffer, so that copy would read pre-frame content and this frame's rendering
+        // into the texture would be lost - precisely the render-target-then-image-unit case this
+        // whole path exists for. Submit what is recorded first; the copy then queues behind it.
+        Bool anyNeedsStorageUpgrade = false;
+        for (auto* texture : storageTextures) {
+            if (m_textureManager->NeedsStorageUsageUpgrade(*texture)) {
+                anyNeedsStorageUpgrade = true;
+                break;
+            }
+        }
+        if (anyNeedsStorageUpgrade && HasPendingRecordedWork()) {
+            if (FlushPendingCommands()) {
+                // Fresh command buffer: the sampled-descriptor-set memo describes bindings that
+                // only existed in the retired one. FlushPendingCommands drops the pipeline memo
+                // itself; this is the other command-buffer-scoped cache.
+                m_lastSampledSetValid = false;
+            } else {
+                // Best effort: the upgrade still produces a correct image, only its preserved
+                // contents may predate this frame's writes. Dropping the draw would be worse.
+                MGLOG_E("%s: flush before a storage-usage image upgrade failed; preserved contents "
+                        "may be stale for one frame", __func__);
+            }
+        }
+        if (!frame.isCommandRecording) {
+            m_frameContext.BeginCommandRecording();
+        }
+
         // Image uploads, deferred-clear materialization, and layout barriers are illegal inside
         // a classic render pass. Do this before sampler preparation as well: a texture used by
         // both a sampler and an image must stay in GENERAL, and both descriptors must name that
         // same layout independent of SPIR-V reflection/binding order.
         if (VkRenderPassManager::GetActiveRenderPass() != nullptr) {
-            VkRenderPassManager::EndRenderPass(commandBuffer);
+            VkRenderPassManager::EndRenderPass(frame.commandBuffer);
         }
 
         for (auto* texture : storageTextures) {
-            if (!MaterializePendingClearForTexture(commandBuffer, *texture)) {
+            if (!MaterializePendingClearForTexture(frame.commandBuffer, *texture)) {
                 MGLOG_E("%s: failed to materialize pending clear for storage textureId=%d",
                         __func__, texture->GetExternalIndex());
                 return false;
             }
-            if (!m_textureManager->TransitionTextureForStorageImage(commandBuffer, *texture)) {
+            if (!m_textureManager->TransitionTextureForStorageImage(frame.commandBuffer, *texture)) {
                 MGLOG_E("%s: failed to prepare storage textureId=%d",
                         __func__, texture->GetExternalIndex());
                 return false;
@@ -4201,7 +4306,25 @@ void main() {
         const auto& vao = *MG_State::pGLContext->GetBoundVertexArray();
         const auto& program = *MG_State::pGLContext->GetCurrentProgram();
         ProgramFactory::CompileOptionFlags transformFlags = GetShaderTransformFlags(m_swapchainObject.GetPreTransform());
-        const auto& programObj = m_programFactory->GetOrCreateProgram(program, transformFlags);
+        const auto* programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
+        // Sampling a colour render target through the driver's implicit-LOD path faults the GPU on
+        // Adreno 650 (see ForceExplicitLod0SamplePass); ask for the explicit-LOD variant when doing
+        // so cannot change a texel, i.e. when every sampler this program reads is pinned to a
+        // single mip level.
+        if (UniformManager::ProgramSamplesOnlySingleLevelTextures(program, *programObjPtr)) {
+            transformFlags |= ProgramFactory::CompileOptionBit::ExplicitLod0Sampling;
+            programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
+        }
+        // fp16 fragment arithmetic is only sound when nothing this draw reads or writes carries
+        // more than 8 normalized bits per channel. A shaderpack's HDR gbuffer, or a data texture
+        // holding positions, must keep full precision - and SPIR-V cannot tell, since sampler2D
+        // yields vec4 whatever the bound format is, so the decision has to be made here.
+        if (UniformManager::ProgramSamplesOnlyLowPrecisionTextures(program, *programObjPtr) &&
+            UniformManager::DrawTargetIsLowPrecision(drawFbo.get())) {
+            transformFlags |= ProgramFactory::CompileOptionBit::RelaxedFragmentPrecision;
+            programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
+        }
+        const auto& programObj = *programObjPtr;
 
         // Begin command recording if not yet
         if (!frame.isCommandRecording) {
@@ -4211,7 +4334,7 @@ void main() {
             m_lastSampledSetValid = false;
         }
 
-        if (!PrepareStorageImageTextures(frame.commandBuffer, program, programObj)) {
+        if (!PrepareStorageImageTextures(frame, program, programObj)) {
             MGLOG_E("SetupDraw skipped: storage image preparation failed");
             return false;
         }
@@ -4436,7 +4559,7 @@ void main() {
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
         }
 
-        if (!PrepareStorageImageTextures(frame.commandBuffer, program, programObj)) {
+        if (!PrepareStorageImageTextures(frame, program, programObj)) {
             MGLOG_E("DispatchCompute skipped: storage image preparation failed");
             return;
         }
@@ -4476,7 +4599,7 @@ void main() {
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
         }
 
-        if (!PrepareStorageImageTextures(frame.commandBuffer, program, programObj)) {
+        if (!PrepareStorageImageTextures(frame, program, programObj)) {
             MGLOG_E("DispatchComputeIndirect skipped: storage image preparation failed");
             return;
         }
@@ -6173,12 +6296,12 @@ void main() {
 
         frame.hasCommandBufferRecorded = false;
         frame.isCommandRecording = false;
-        // The wait proved every descriptor set this slot has in flight idle;
-        // rewind the reuse cursors so present-less readback loops stay
-        // bounded (Present is the only other rewind point).
-        if (m_uniformManager) {
-            m_uniformManager->BeginFrame(m_frameContext.GetCurrentFrameIndex());
-        }
+        // The wait proved every submission complete, so the full frame-boundary
+        // drain applies: descriptor cursors, transient arenas, deferred
+        // texture/buffer releases, retired command buffers and the converted
+        // vertex-stream cache all rewind here, keeping present-less readback
+        // loops bounded (Present is the only other drain point).
+        TryDrainFrameTransients();
         return true;
     }
 
@@ -6325,7 +6448,8 @@ void main() {
                 if (RemapDefaultFboReadbackToGLOrientation(mapped, swapchainExtent, preTransform,
                                                            sourceTexelSize,
                                                            remapped.data())) {
-                    PackReadbackToClientOrPbo(remapped.data(), srcFormat, width, height, format, type, pixels);
+                    PackReadbackToClientOrPbo(remapped.data(), srcFormat, width, height, 1, format, type, pixels,
+                                              /*applyPackImageParams=*/false);
                     return;
                 }
             }
@@ -6334,7 +6458,8 @@ void main() {
                     width, height, swapchainExtent.width, swapchainExtent.height,
                     static_cast<Int>(preTransform));
         }
-        PackReadbackToClientOrPbo(mapped, srcFormat, width, height, format, type, pixels);
+        PackReadbackToClientOrPbo(mapped, srcFormat, width, height, 1, format, type, pixels,
+                                  /*applyPackImageParams=*/false);
     }
 
     void VulkanRenderer::GetTexImage(GLenum target, GLint level, GLenum format, GLenum type, GLvoid* pixels) {
@@ -6387,6 +6512,17 @@ void main() {
         if (width <= 0 || height <= 0) {
             return;
         }
+        // GetTexImage returns every slice of a 3D level and every layer of an array
+        // level; GL_PACK_IMAGE_HEIGHT / GL_PACK_SKIP_IMAGES apply to the 3D/array
+        // destination layout (GL 3.3 section 6.1.4).
+        const auto imageTextureTarget = textureObject->GetTarget();
+        const Bool is3dImage = imageTextureTarget == TextureTarget::Texture3D;
+        const Bool isArrayImage = imageTextureTarget == TextureTarget::Texture1DArray ||
+                                  imageTextureTarget == TextureTarget::Texture2DArray ||
+                                  imageTextureTarget == TextureTarget::TextureCubeMapArray;
+        const GLsizei depthSlices = is3dImage ? std::max<GLsizei>(texelSize.z(), 1) : 1;
+        const GLsizei arrayLayers = isArrayImage ? static_cast<GLsizei>(resource->arrayLayers) : 1;
+        const GLsizei sliceCount = std::max<GLsizei>(depthSlices * arrayLayers, 1);
         if (bufSize >= 0) {
             const Int dstChannels = GetReadbackChannelCount(format);
             if ((type == GL_UNSIGNED_BYTE || type == GL_FLOAT) && dstChannels > 0) {
@@ -6407,7 +6543,8 @@ void main() {
             return;
         }
         const VkDeviceSize readbackSize = static_cast<VkDeviceSize>(width) *
-                                          static_cast<VkDeviceSize>(height) * sourceTexelSize;
+                                          static_cast<VkDeviceSize>(height) *
+                                          static_cast<VkDeviceSize>(sliceCount) * sourceTexelSize;
         VkBufferObject readback;
         if (!readback.Create({
                 .allocator = m_allocator,
@@ -6435,8 +6572,9 @@ void main() {
         copyRegion.imageSubresource.aspectMask = resource->aspect;
         copyRegion.imageSubresource.mipLevel = static_cast<Uint32>(level);
         copyRegion.imageSubresource.baseArrayLayer = 0;
-        copyRegion.imageSubresource.layerCount = 1;
-        copyRegion.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
+        copyRegion.imageSubresource.layerCount = static_cast<Uint32>(arrayLayers);
+        copyRegion.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height),
+                                  static_cast<Uint32>(depthSlices)};
         vkCmdCopyImageToBuffer(frame.commandBuffer, resource->image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                                readback.GetHandle(), 1, &copyRegion);
 
@@ -6462,7 +6600,8 @@ void main() {
             MGLOG_E("DirectVulkan::GetTextureImage skipped: failed to invalidate readback buffer");
             return;
         }
-        PackReadbackToClientOrPbo(mapped, resource->format, width, height, format, type, pixels);
+        PackReadbackToClientOrPbo(mapped, resource->format, width, height, sliceCount, format, type, pixels,
+                                  /*applyPackImageParams=*/is3dImage || isArrayImage);
     }
 
     void VulkanRenderer::GenerateMipmap(GLenum target) {
@@ -7076,6 +7215,9 @@ void main() {
         }
         m_bufferManager.NotifyDeviceIdle();
         OnSubmitsCompletedUpTo(m_submitCounter);
+        // The queue was just drained; take the free frame-boundary drain when
+        // nothing is recorded (present-less timer-query loops). No-op otherwise.
+        TryDrainFrameTransients();
         return true;
     }
 
@@ -7144,6 +7286,85 @@ void main() {
                 vkDestroyFence(m_device, record.fence, nullptr);
             }
         }
+        // Mid-frame-flushed command buffers whose submission just completed can
+        // be freed now; present-less flush loops have no other reclaim point.
+        m_frameContext.FreeRetiredCommandBuffersCompletedUpTo(m_completedSubmitCounter);
+    }
+
+    Bool VulkanRenderer::TryDrainFrameTransients() {
+        if (m_device == VK_NULL_HANDLE || m_frameContext.GetFrameCount() == 0) {
+            return false;
+        }
+        if (m_completedSubmitCounter != m_submitCounter) {
+            RefreshCompletedSubmits();
+            if (m_completedSubmitCounter != m_submitCounter) {
+                return false;
+            }
+        }
+        if (HasPendingRecordedWork()) {
+            return false;
+        }
+
+        // Every submission is complete and nothing recorded references the
+        // per-frame transients. Pure-reclaim work runs on every drain: it only
+        // releases memory that is provably dead, never invalidates anything a
+        // later draw would have to rebuild. Raise the buffer manager's
+        // completed floor first so busy-tracking reflects the proven idleness.
+        m_bufferManager.NotifyDeviceIdle();
+
+        const Uint32 frameIndex = m_frameContext.GetCurrentFrameIndex();
+        m_frameContext.FreeAllRetiredCommandBuffers();
+        for (Uint32 slot = 0; slot < m_deferredDepthMipmapCleanup.size(); ++slot) {
+            CollectDeferredDepthMipmapCleanup(slot);
+        }
+        if (m_textureManager) {
+            m_textureManager->CollectAllDeferredReleases();
+        }
+        m_bufferManager.CollectAllDeferredReleases();
+        // Descriptor cursors rewind on every drain (the pre-drain readback path
+        // already did exactly this), keeping fence/readback loops' set usage bounded.
+        if (m_uniformManager) {
+            m_uniformManager->BeginFrame(frameIndex);
+        }
+
+        // Frame-boundary-equivalent work - transient arena rewind (which invalidates
+        // the conversion cache) and the cache-aging clocks - is gated to every 8th
+        // drain since the last Present: a presenting app's mid-frame readbacks/waits
+        // must neither force re-conversion/re-upload churn for the rest of the frame
+        // nor multiply the aging rate (which would shrink the 1024-boundary retire
+        // window and thrash periodically-used pipelines/programs), while present-less
+        // loops still rewind the arena and age their caches every 8 iterations -
+        // bounded by 8 iterations' transient usage.
+        ++m_drainsSinceLastPresent;
+        if ((m_drainsSinceLastPresent % 8) != 0) {
+            return true;
+        }
+        if (m_textureManager) {
+            m_textureManager->BeginFrame(frameIndex);
+        }
+        m_bufferManager.BeginFrame(frameIndex);
+        // The cached conversion slices point into the transient arena the
+        // BeginFrame above just rewound; drop them together.
+        m_convertedVertexStreams.clear();
+        if (m_renderPassManager) {
+            m_renderPassManager->OnPresent();
+        }
+        // The pipeline memo can survive across these boundaries (no per-frame reset
+        // on this path), so it must drop whenever the sweep destroys anything.
+        if (m_programFactory) {
+            m_programFactory->OnFrameBoundary();
+        }
+        if (m_pipelineFactory && m_pipelineFactory->OnFrameBoundary() > 0) {
+            m_lastPipelineValid = false;
+            m_lastPipelineResult = VK_NULL_HANDLE;
+        }
+        if (m_vertexInputStateFactory) {
+            m_vertexInputStateFactory->OnFrameBoundary();
+        }
+        if (m_samplerManager) {
+            m_samplerManager->OnFrameBoundary();
+        }
+        return true;
     }
 
     VkFence VulkanRenderer::AcquirePooledSubmitFence() {
@@ -7206,6 +7427,10 @@ void main() {
         if (m_device == VK_NULL_HANDLE || m_graphicsQueue == VK_NULL_HANDLE || m_frameContext.GetFrameCount() == 0) {
             return false;
         }
+        // Non-blocking completion poll: gives flush-only workloads (no sync
+        // objects, no present) a point where finished submissions retire their
+        // pooled fences and mid-frame command buffers.
+        RefreshCompletedSubmits();
         auto& frame = m_frameContext.GetCurrent();
         if (!frame.isCommandRecording && !frame.hasCommandBufferRecorded) {
             return false;
@@ -7230,6 +7455,14 @@ void main() {
             return false;
         }
 
+        // Command-buffer boundary: the pipeline memo must not survive it, or a
+        // pipeline bound only through memo hits is never re-stamped in the factory
+        // cache and the aging sweep could destroy it while the flushed submission
+        // still references it. Mirrors the drops at the readback and Present
+        // boundaries; costs one full pipeline lookup on the next draw.
+        m_lastPipelineValid = false;
+        m_lastPipelineResult = VK_NULL_HANDLE;
+
         // The submitted command buffer may still be executing; recording must
         // restart on a fresh one. If none can be allocated, fall back to
         // draining this submission so reusing the buffer stays legal.
@@ -7281,6 +7514,10 @@ void main() {
                 const VkResult result = vkWaitForFences(m_device, 1, &record.fence, VK_TRUE, timeoutNs);
                 if (result == VK_SUCCESS) {
                     OnSubmitsCompletedUpTo(record.submitIndex);
+                    // The wait already stalled the pipeline; if it happens to
+                    // have drained everything (present-less fence loops), take
+                    // the free frame-boundary drain. No-op otherwise.
+                    TryDrainFrameTransients();
                     return true;
                 }
                 if (result != VK_TIMEOUT) {
@@ -7353,24 +7590,86 @@ void main() {
     }
 
     void VulkanRenderer::Present() {
+        if (m_swapchainObject.GetHandle() == VK_NULL_HANDLE || m_presentSuspended) {
+            // No usable swapchain: the window was zero-area at initialization, or
+            // presentation was suspended when the window minimized. Try to bring a
+            // swapchain up now that the window may have a real size; until then, drop
+            // this frame's recording instead of submitting - a submit would wait on a
+            // never-signaled acquire semaphore and reuse a still-signaled fence.
+            if (!RecreateSwapchain()) {
+                auto& suspendedFrame = m_frameContext.GetCurrent();
+                if (VkRenderPassManager::GetActiveRenderPass()) {
+                    VkRenderPassManager::EndRenderPass(suspendedFrame.commandBuffer);
+                }
+                if (suspendedFrame.isCommandRecording) {
+                    m_frameContext.EndCommandRecording();
+                }
+                suspendedFrame.isCommandRecording = false;
+                suspendedFrame.hasCommandBufferRecorded = false;
+                m_lastPipelineValid = false;
+                // The dropped recording is never submitted, so once the fence
+                // poll shows the pre-suspension submissions complete the frame
+                // transients (descriptor sets, transient arenas, deferred
+                // releases, conversion caches) can rewind; without this a
+                // minimized-window app accumulates them for the whole
+                // suspension.
+                TryDrainFrameTransients();
+                MGLOG_D("Present skipped: no usable swapchain (zero-area window)");
+                return;
+            }
+            m_presentSuspended = false;
+            const VkResult acquireResult =
+                m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
+            if (acquireResult == VK_SUBOPTIMAL_KHR) {
+                // Usable image with its acquire signal already armed; a rebuild is scheduled
+                // only if the surface genuinely no longer matches (see step 4 of Present).
+                m_swapchainResizeRequested = m_swapchainResizeRequested || SwapchainIsOutOfDate();
+            } else {
+                VK_VERIFY(acquireResult, "Present, deferred first WaitAndAcquireNextImage");
+            }
+        }
         MOBILEGL_ASSERT(m_imageIndexAcquired < m_swapchainObject.GetImageCount(),
                         "Present, acquired image index out of range");
         m_renderPassManager->OnPresent();
+        // A real presented frame is the canonical aging cadence; mid-frame drains
+        // count against this and only age when presents stop coming.
+        m_drainsSinceLastPresent = 0;
+        // Age the content-addressed caches on the same frame-boundary cadence. Each
+        // keeps its own internal 256-sweep gate, so the per-frame cost is one counter
+        // increment and compare per cache; entries used by this frame's still-
+        // unsubmitted recording were stamped this boundary (every command-buffer
+        // boundary drops the pipeline memo, so the first draw of each recording
+        // performs a real, stamping lookup) and can never age out.
+        m_programFactory->OnFrameBoundary();
+        if (m_pipelineFactory->OnFrameBoundary() > 0) {
+            m_lastPipelineValid = false; // an aged-out pipeline may still be memoized
+            m_lastPipelineResult = VK_NULL_HANDLE;
+        }
+        m_vertexInputStateFactory->OnFrameBoundary();
+        m_samplerManager->OnFrameBoundary();
         auto& frame = m_frameContext.GetCurrent();
         auto* activeRenderPass = VkRenderPassManager::GetActiveRenderPass();
         if (activeRenderPass)
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
 
+        // Transition while this frame's recording is still open. A frame that
+        // rendered only into FBOs has no default-framebuffer render pass, and that
+        // pass's finalLayout is the only other thing that carries the swapchain
+        // image to PRESENT_SRC_KHR - so closing the buffer first, which made
+        // TransitionToPresent refuse to record, handed the image to
+        // vkQueuePresentKHR in the layout it was acquired in (UNDEFINED on a fresh
+        // swapchain). The SetImageLayout below then made the tracker's
+        // disagreement with reality permanent for that image index.
+        const auto acquiredImageLayout = m_swapchainObject.GetImageLayout(m_imageIndexAcquired);
+        m_frameContext.TransitionToPresent(m_swapchainObject.GetImage(m_imageIndexAcquired), acquiredImageLayout);
+
         if (frame.isCommandRecording) {
             m_frameContext.EndCommandRecording();
             frame.hasCommandBufferRecorded = true;
             m_lastPipelineValid = false; // command-buffer boundary: drop the pipeline memo
         }
 
-        const auto acquiredImageLayout = m_swapchainObject.GetImageLayout(m_imageIndexAcquired);
-        const Bool needsLayoutTransitionForPresent =
-            m_frameContext.TransitionToPresent(m_swapchainObject.GetImage(m_imageIndexAcquired), acquiredImageLayout);
-        const Bool shouldSubmitCommandBuffer = frame.hasCommandBufferRecorded || needsLayoutTransitionForPresent;
+        const Bool shouldSubmitCommandBuffer = frame.hasCommandBufferRecorded;
 
         // 1) Submit current frame work.
         auto submitPacket = m_frameContext.GetSubmitInfo(shouldSubmitCommandBuffer, m_imageIndexAcquired);
@@ -7384,16 +7683,43 @@ void main() {
         // 2) Present current frame.
         auto presentPacket = m_frameContext.GetPresentInfo(m_swapchainObject.GetHandle(), m_imageIndexAcquired);
         auto result = vkQueuePresentKHR(m_presentQueue, &presentPacket.presentInfo);
-        if (result == VK_ERROR_OUT_OF_DATE_KHR || result == VK_SUBOPTIMAL_KHR) {
+        if (result == VK_SUBOPTIMAL_KHR) {
+            // Suboptimal is not a reason to rebuild on its own: a driver may report it for a
+            // surface whose size and orientation still match what we built from (Android does
+            // this routinely), and rebuilding on it alone destroys every pipeline and
+            // reallocates the default framebuffer once per frame - flicker, then garbage.
+            // Defer to the surface-capabilities comparison below.
+            result = VK_SUCCESS;
+        }
+        if (result == VK_ERROR_OUT_OF_DATE_KHR) {
             MGLOG_D("Present, vkQueuePresentKHR got %d, recreating swapchain", result);
-            RecreateSwapchain();
+            if (!RecreateSwapchain()) {
+                // Window went zero-area (minimize) with the swapchain out of date:
+                // stop submitting/acquiring until it has a size again.
+                m_presentSuspended = true;
+                m_swapchainResizeRequested = false;
+                MGLOG_D("Present, zero-area window with out-of-date swapchain; suspending presentation");
+                return;
+            }
             m_swapchainResizeRequested = false;
             result = VK_SUCCESS;
         }
         VK_VERIFY(result, "Present, vkQueuePresentKHR");
+        // The authoritative check, done here - after the frame is presented, before the next
+        // acquire. This is what makes a launcher-side resolution change take effect: shrinking
+        // the window's buffer (SurfaceHolder.setFixedSize) moves currentExtent, the swapchain
+        // follows, and the compositor scales the smaller image up to the view for free.
+        if (!m_swapchainResizeRequested && SwapchainIsOutOfDate()) {
+            m_swapchainResizeRequested = true;
+        }
         if (m_swapchainResizeRequested) {
             MGLOG_D("Present, processing requested swapchain resize");
-            RecreateSwapchain();
+            if (!RecreateSwapchain()) {
+                m_presentSuspended = true;
+                m_swapchainResizeRequested = false;
+                MGLOG_D("Present, zero-area window on requested resize; suspending presentation");
+                return;
+            }
             m_swapchainResizeRequested = false;
         }
 
@@ -7402,9 +7728,23 @@ void main() {
 
         // 4) Wait/reset/acquire for next frame.
         result = m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
-        if (result == VK_ERROR_OUT_OF_DATE_KHR || result == VK_SUBOPTIMAL_KHR) {
+        if (result == VK_SUBOPTIMAL_KHR) {
+            // An image WAS acquired and its signal is armed on this slot's
+            // imageAvailableSemaphore, so the frame proceeds normally. Whether a rebuild is
+            // actually needed is decided by the surface-capabilities comparison at the next
+            // Present - suboptimal alone must not schedule one, or a driver that reports it
+            // every frame would rebuild every frame.
+            m_swapchainResizeRequested = m_swapchainResizeRequested || SwapchainIsOutOfDate();
+            result = VK_SUCCESS;
+        } else if (result == VK_ERROR_OUT_OF_DATE_KHR) {
+            // Nothing acquired, nothing signaled: safe to rebuild and re-acquire.
             MGLOG_D("Present, vkAcquireNextImageKHR got %d, recreating swapchain", result);
-            RecreateSwapchain();
+            if (!RecreateSwapchain()) {
+                m_presentSuspended = true;
+                m_swapchainResizeRequested = false;
+                MGLOG_D("Present, zero-area window on next-frame acquire; suspending presentation");
+                return;
+            }
             m_swapchainResizeRequested = false;
             result =
                 m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
@@ -7445,6 +7785,24 @@ void main() {
 
         m_validationLayersEnabled = m_config.EnableValidationLayers && validationLayerAvailable;
 
+        // The debug messenger is a VK_EXT_debug_utils object, but a driver can ship
+        // the validation layers while exposing only the older VK_EXT_debug_report
+        // (Adreno 650 / Vulkan 1.1.128 does exactly that). Requesting the extension
+        // unconditionally tripped the required-extension assert below, aborting every
+        // validation-enabled build in CreateInstance. Keep the layers - they still
+        // validate, and on Android they report to logcat on their own - and drop only
+        // the messenger.
+        const Bool debugUtilsAvailable =
+            m_validationLayersEnabled && IsExtensionSupported(m_extensions, VK_EXT_DEBUG_UTILS_EXTENSION_NAME);
+        // Without a reporting channel the layers validate but say nothing, so fall
+        // back to VK_EXT_debug_report when debug_utils is missing.
+        const Bool debugReportAvailable = m_validationLayersEnabled && !debugUtilsAvailable &&
+                                          IsExtensionSupported(m_extensions, VK_EXT_DEBUG_REPORT_EXTENSION_NAME);
+        if (m_validationLayersEnabled && !debugUtilsAvailable) {
+            MGLOG_I("%s not available; validation reports via %s instead.", VK_EXT_DEBUG_UTILS_EXTENSION_NAME,
+                    debugReportAvailable ? VK_EXT_DEBUG_REPORT_EXTENSION_NAME : "(no channel)");
+        }
+
         // ---------------- App info -------------------
         VkApplicationInfo appInfo = {};
         appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
@@ -7495,8 +7853,10 @@ void main() {
         }
 #endif
 
-        if (m_validationLayersEnabled) {
+        if (debugUtilsAvailable) {
             exts.push_back(VK_EXT_DEBUG_UTILS_EXTENSION_NAME);
+        } else if (debugReportAvailable) {
+            exts.push_back(VK_EXT_DEBUG_REPORT_EXTENSION_NAME);
         }
 
         MGLOG_I("Enabling %d Vulkan instance extensions:", exts.size());
@@ -7521,7 +7881,8 @@ void main() {
             MGLOG_I("Enabling validation layer...");
             instanceInfo.enabledLayerCount = static_cast<uint32_t>(std::size(s_validationLayerNames));
             instanceInfo.ppEnabledLayerNames = s_validationLayerNames;
-            instanceInfo.pNext = &debugMessengerCreateInfo;
+            // Chaining the messenger create-info is only legal with the extension on.
+            instanceInfo.pNext = debugUtilsAvailable ? &debugMessengerCreateInfo : nullptr;
         } else {
             instanceInfo.enabledLayerCount = 0;
             instanceInfo.pNext = nullptr;
@@ -7529,7 +7890,40 @@ void main() {
 
         VK_VERIFY(vkCreateInstance(&instanceInfo, nullptr, &m_instance), "vkCreateInstance failed");
 
-        if (m_validationLayersEnabled) VK_VERIFY(SetupDebugMessenger());
+        if (debugUtilsAvailable) {
+            VK_VERIFY(SetupDebugMessenger());
+        } else if (debugReportAvailable) {
+            VK_VERIFY(SetupDebugReportCallback());
+        }
+    }
+
+    static VKAPI_ATTR VkBool32 VKAPI_CALL DebugReportCallback(VkDebugReportFlagsEXT flags, VkDebugReportObjectTypeEXT,
+                                                              Uint64, size_t, Int32 messageCode, const char* pLayerPrefix,
+                                                              const char* pMessage, void*) {
+        if ((flags & (VK_DEBUG_REPORT_ERROR_BIT_EXT | VK_DEBUG_REPORT_WARNING_BIT_EXT |
+                      VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT)) != 0) {
+            MGLOG_F("[Vulkan %s %d] %s", pLayerPrefix ? pLayerPrefix : "?", messageCode, pMessage ? pMessage : "");
+        }
+        return VK_FALSE;
+    }
+
+    VkResult VulkanRenderer::SetupDebugReportCallback() {
+        auto vkCreateDebugReportCallbackEXT =
+            (PFN_vkCreateDebugReportCallbackEXT)vkGetInstanceProcAddr(m_instance, "vkCreateDebugReportCallbackEXT");
+        if (!vkCreateDebugReportCallbackEXT) return VK_ERROR_EXTENSION_NOT_PRESENT;
+        VkDebugReportCallbackCreateInfoEXT createInfo{VK_STRUCTURE_TYPE_DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT};
+        createInfo.flags = VK_DEBUG_REPORT_ERROR_BIT_EXT | VK_DEBUG_REPORT_WARNING_BIT_EXT |
+                           VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT;
+        createInfo.pfnCallback = &DebugReportCallback;
+        return vkCreateDebugReportCallbackEXT(m_instance, &createInfo, nullptr, &m_debugReportCallback);
+    }
+
+    void VulkanRenderer::DestroyDebugReportCallback() {
+        if (m_debugReportCallback == VK_NULL_HANDLE) return;
+        auto func = (PFN_vkDestroyDebugReportCallbackEXT)vkGetInstanceProcAddr(m_instance,
+                                                                               "vkDestroyDebugReportCallbackEXT");
+        if (func != nullptr) func(m_instance, m_debugReportCallback, nullptr);
+        m_debugReportCallback = VK_NULL_HANDLE;
     }
 
     VkResult VulkanRenderer::SetupDebugMessenger() {
@@ -7811,6 +8205,18 @@ void main() {
 
         const Vector<VkExtensionProperties> availableExtensions = EnumerateDeviceExtensions(m_physicalDevice.handle);
         ResolveOptionalDeviceExtensions(availableExtensions, enabledDeviceExtensions);
+
+        // VK_KHR_image_format_list lets a MUTABLE_FORMAT image declare exactly which formats it
+        // may be viewed as. Adreno drops UBWC bandwidth compression on a blindly-mutable image
+        // (measured: 65 -> 80 fps in MC 26.2 once mutability is not requested); an explicit,
+        // compression-compatible format list is the portable way to keep both.
+        m_imageFormatListExtensionEnabled =
+            IsExtensionSupported(availableExtensions, VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME);
+        if (m_imageFormatListExtensionEnabled) {
+            enabledDeviceExtensions.push_back(VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME);
+        }
+        MGLOG_I("VK_KHR_image_format_list enabled: %s",
+                m_imageFormatListExtensionEnabled ? "true" : "false");
         MGLOG_I("VK_KHR_draw_indirect_count enabled: %s", m_drawIndirectCountExtensionEnabled ? "true" : "false");
 
         m_indexTypeUint8ExtensionEnabled = false;
@@ -8056,6 +8462,7 @@ void main() {
         MOBILEGL_ASSERT(hwnd, "HWND is null");
 
         VkWin32SurfaceCreateInfoKHR sci{VK_STRUCTURE_TYPE_WIN32_SURFACE_CREATE_INFO_KHR};
+        sci.hinstance = GetModuleHandleW(nullptr);
         sci.hwnd = hwnd;
         VK_VERIFY(vkCreateWin32SurfaceKHR(m_instance, &sci, nullptr, &m_surface), "vkCreateWin32SurfaceKHR failed");
 #elif defined VK_USE_PLATFORM_METAL_EXT
@@ -8217,13 +8624,13 @@ void main() {
         m_swapchainObject.Shutdown(m_device);
     }
 
-    void VulkanRenderer::RecreateSwapchain() {
+    Bool VulkanRenderer::RecreateSwapchain() {
         // Handle cases like minimize on Windows, where swapchain could return a 0x0 extent
         const auto swapchainCapabilities =
             SwapchainObject::GetSwapchainCapabilities(m_physicalDevice.handle, m_surface);
         if (swapchainCapabilities.capabilities.currentExtent.width == 0 ||
             swapchainCapabilities.capabilities.currentExtent.height == 0) {
-            return;
+            return false;
         }
 
         vkDeviceWaitIdle(m_device);
@@ -8267,12 +8674,45 @@ void main() {
             m_bufferManager.BeginFrame(m_frameContext.GetCurrentFrameIndex());
             m_convertedVertexStreams.clear();
         }
+        return true;
     }
 
     const PhysicalDevice& VulkanRenderer::GetPhysicalDevice() const {
         return m_physicalDevice;
     }
 
+    Bool VulkanRenderer::SwapchainIsOutOfDate() {
+        if (m_surface == VK_NULL_HANDLE || m_swapchainObject.GetHandle() == VK_NULL_HANDLE) {
+            return false;
+        }
+        VkSurfaceCapabilitiesKHR surfaceCaps{};
+        if (vkGetPhysicalDeviceSurfaceCapabilitiesKHR(m_physicalDevice.handle, m_surface, &surfaceCaps) !=
+            VK_SUCCESS) {
+            return false;
+        }
+        // A driver-defined currentExtent (UINT32_MAX) means the surface takes its size from the
+        // swapchain, so there is nothing to compare against - the app's requested size wins and
+        // only an explicit RequestSwapchainResize can change it.
+        if (surfaceCaps.currentExtent.width == UINT32_MAX || surfaceCaps.currentExtent.height == UINT32_MAX) {
+            return false;
+        }
+        // Compare in SURFACE space against the extent the live swapchain was created from. Using
+        // the swapchain's own (quarter-turn swapped) extent here would report a difference on
+        // every rotated frame and rebuild forever.
+        const VkExtent2D builtFrom = m_swapchainObject.GetSurfaceExtent();
+        const Bool extentChanged = surfaceCaps.currentExtent.width != builtFrom.width ||
+                                   surfaceCaps.currentExtent.height != builtFrom.height;
+        const Bool transformChanged = surfaceCaps.currentTransform != m_swapchainObject.GetPreTransform();
+        if (!extentChanged && !transformChanged) {
+            return false;
+        }
+        MGLOG_I("Swapchain out of date: surface %ux%u transform %u -> %ux%u transform %u",
+                builtFrom.width, builtFrom.height, static_cast<Uint32>(m_swapchainObject.GetPreTransform()),
+                surfaceCaps.currentExtent.width, surfaceCaps.currentExtent.height,
+                static_cast<Uint32>(surfaceCaps.currentTransform));
+        return true;
+    }
+
     void VulkanRenderer::RequestSwapchainResize(Uint32 width, Uint32 height) {
         width = std::max<Uint32>(width, 1);
         height = std::max<Uint32>(height, 1);
@@ -8369,6 +8809,43 @@ void main() {
         m_computePipelines.clear();
     }
 
+    void VulkanRenderer::OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) {
+        if (m_pipelineFactory == nullptr) {
+            return;
+        }
+        // The render-pass sweep's >1024-boundary idle guarantee covers these pipelines
+        // too (they are only bound by draws that hit the dying entries), so the factory
+        // destroys them immediately. The memo must drop as well: it can hand out a
+        // cached handle without touching the factory.
+        if (m_pipelineFactory->EvictByRenderPasses(renderPasses) > 0) {
+            m_lastPipelineValid = false;
+            m_lastPipelineResult = VK_NULL_HANDLE;
+        }
+    }
+
+    void VulkanRenderer::OnProgramEvicted(ProgramFactory::HashType programHash,
+                                          VkDescriptorSetLayout descriptorSetLayout) {
+        // Same >1024-boundary idleness as the program entry: its compute pipeline is
+        // only dispatched, and its graphics pipelines only bound, through paths that
+        // stamp the entry, so immediate destruction is GPU-safe. (The graphics memo
+        // never holds compute pipelines; it only needs invalidating for the factory
+        // eviction below.)
+        const auto computeIt = m_computePipelines.find(programHash);
+        if (computeIt != m_computePipelines.end()) {
+            if (computeIt->second != VK_NULL_HANDLE && m_device != VK_NULL_HANDLE) {
+                vkDestroyPipeline(m_device, computeIt->second, nullptr);
+            }
+            m_computePipelines.erase(computeIt);
+        }
+        if (m_pipelineFactory != nullptr && m_pipelineFactory->EvictByProgramHash(programHash) > 0) {
+            m_lastPipelineValid = false;
+            m_lastPipelineResult = VK_NULL_HANDLE;
+        }
+        if (m_uniformManager != nullptr) {
+            m_uniformManager->OnDescriptorSetLayoutDestroyed(descriptorSetLayout);
+        }
+    }
+
     VkPipeline VulkanRenderer::GetOrCreateComputePipeline(const ProgramFactory::VkProgramObject& programObj) {
         const auto it = m_computePipelines.find(programObj.hash);
         if (it != m_computePipelines.end()) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -114,7 +114,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     };
 
-    class VulkanRenderer : public IBufferCopyCommandProvider, public FrameContext::IRecordingObserver {
+    class VulkanRenderer : public IBufferCopyCommandProvider,
+                           public FrameContext::IRecordingObserver,
+                           public VkRenderPassManager::IEvictionObserver,
+                           public ProgramFactory::IEvictionObserver {
     public:
         VulkanRenderer(NativeWindowType window, const VulkanRendererConfig& cfg = {});
         ~VulkanRenderer();
@@ -131,6 +134,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // recording, before any render pass.
         void OnFrameCommandRecordingBegan(VkCommandBuffer commandBuffer) override;
 
+        // VkRenderPassManager::IEvictionObserver: the render-pass aging sweep just
+        // destroyed these VkRenderPasses; evict every graphics pipeline hashed on a
+        // dying handle (they share its >1024-boundary idleness, so immediate
+        // destruction is safe) and drop the last-pipeline memo if any went.
+        void OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) override;
+
+        // ProgramFactory::IEvictionObserver: an aged-out program entry was
+        // destroyed; evict its compute pipeline and graphics pipelines (same
+        // idleness guarantee - they are only bound through draws/dispatches that
+        // stamp the program entry) and purge the descriptor-set cache entries
+        // keyed by its now-recyclable VkDescriptorSetLayout handle.
+        void OnProgramEvicted(ProgramFactory::HashType programHash,
+                              VkDescriptorSetLayout descriptorSetLayout) override;
+
         Bool SetupDraw(FrameContext::FrameData& frame, GLenum mode, Flags<DrawSetupAspect> aspects,
                        const DrawCmdParam& drawParams,
                        const IndexBufferView* pIndexBufferView = nullptr);
@@ -257,7 +274,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint64 GetTimerQueryTimestampNs(const VkTimerQueryManager::TimestampRecord& record) const;
 
         void RequestSwapchainResize(Uint32 width, Uint32 height);
-        void RecreateSwapchain();
+        // Re-query the surface and report whether the live swapchain no longer matches it
+        // (size or orientation). This - not a VK_SUBOPTIMAL_KHR result - is what decides a
+        // rebuild, so a surface the driver merely considers suboptimal cannot thrash.
+        Bool SwapchainIsOutOfDate();
+        // Returns false when the surface is zero-area (minimized/hidden window):
+        // no new swapchain is installed and presentation must stay suspended.
+        Bool RecreateSwapchain();
 
     private:
         struct BlitUniformData {
@@ -343,11 +366,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkFence AcquirePooledSubmitFence();
         void DestroySubmitFencePool();
         Bool HasPendingRecordedWork() const;
+        // Frame-boundary housekeeping for paths that never reach Present's
+        // tail (present-less readback loops, suspended presentation, blocking
+        // sync waits): runs the same per-frame drains Present performs, but
+        // only when every queue submission has been observed complete AND no
+        // recorded-but-unsubmitted commands exist - i.e. when CPU-GPU overlap
+        // is provably already zero. Never blocks (non-blocking fence poll
+        // only), so the presenting path's frames-in-flight pipelining is
+        // untouched. Returns true when the drain ran.
+        Bool TryDrainFrameTransients();
 
         Vector<SubmitRecord> m_inFlightSubmits;
         Vector<VkFence> m_freeSubmitFences;
         Uint64 m_submitCounter = 0;
         Uint64 m_completedSubmitCounter = 0;
+        // Drains since the last Present, gating the drain's frame-boundary-equivalent
+        // work (arena rewind + cache aging): a presenting app's mid-frame
+        // readbacks/waits must neither churn the transient caches nor accelerate the
+        // aging clocks, while present-less loops still cross a boundary every few
+        // iterations. Reset in Present.
+        Uint32 m_drainsSinceLastPresent = 0;
 
         NativeWindowType m_window = 0;
         void* m_platformDisplay = nullptr;
@@ -355,12 +393,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void* m_platformCloseDisplay = nullptr;
         VulkanRendererConfig m_config;
         Bool m_swapchainResizeRequested = false;
+        // Presentation is suspended while the window is zero-area (minimized): the
+        // swapchain is unusable/out of date, so Present drops frames instead of
+        // submitting on a signaled fence / presenting never-acquired images.
+        Bool m_presentSuspended = false;
 
         // Vulkan objects
         Bool m_validationLayersEnabled = false;
         Vector<VkExtensionProperties> m_extensions;
         VkInstance m_instance = VK_NULL_HANDLE;
         VkDebugUtilsMessengerEXT m_debugMessenger = VK_NULL_HANDLE;
+        // Fallback reporting channel for drivers that ship the validation layers but
+        // only expose the older VK_EXT_debug_report (Adreno 650 / Vulkan 1.1.128).
+        VkDebugReportCallbackEXT m_debugReportCallback = VK_NULL_HANDLE;
         PhysicalDevice m_physicalDevice;
         VkDevice m_device = VK_NULL_HANDLE;
         VmaAllocator m_allocator = nullptr;
@@ -511,6 +556,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void CreateInstance();
         VkResult SetupDebugMessenger();
         VkResult DestroyDebugMessenger();
+        VkResult SetupDebugReportCallback();
+        void DestroyDebugReportCallback();
         VkDebugUtilsMessengerCreateInfoEXT PopulateDebugMessengerCreateInfo();
         void CreateSurface();
         void PickPhysicalDevice();
@@ -529,8 +576,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const RenderPassEntry& renderPassEntry);
         VkPipeline GetOrCreateComputePipeline(const ProgramFactory::VkProgramObject& programObj);
         void DestroyComputePipelines();
+        // Takes the frame rather than a command buffer: a first-time storage-usage upgrade has to
+        // flush the pending recording (see the body), which retires the current command buffer.
         Bool PrepareStorageImageTextures(
-            VkCommandBuffer commandBuffer,
+            FrameContext::FrameData& frame,
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj);
 
@@ -592,6 +641,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  const PhysicalDevice& compareWithDevice,
                                                  PhysicalDevice& outBetterDevice);
         static constexpr const char* s_validationLayerNames[] = {"VK_LAYER_KHRONOS_validation"};
+        // VK_KHR_image_format_list: lets MUTABLE_FORMAT images declare their exact view-format
+        // set so the driver can keep bandwidth compression (see CreateLogicalDeviceAndQueues).
+        Bool m_imageFormatListExtensionEnabled = false;
+
         static constexpr const char* s_deviceExtensionNames[] = {VK_KHR_SWAPCHAIN_EXTENSION_NAME};
         static Bool CheckValidationLayerSupport();
 

MobileGL/MG_Impl/CGLImpl/CGLImpl.cpp

@@ -24,6 +24,7 @@ namespace MobileGL::MG_Impl::CGLImpl {
             GLint Samples = 0;
             GLint Profile = kCGLOGLPVersion_3_2_Core;
             GLint RendererId = 0x4d474c;
+            GLint DisplayMask = 0;
         };
 
         struct ContextObject {
@@ -134,6 +135,9 @@ namespace MobileGL::MG_Impl::CGLImpl {
             case kCGLPFARendererID:
                 pixelFormat.RendererId = value;
                 break;
+            case kCGLPFADisplayMask:
+                pixelFormat.DisplayMask = value;
+                break;
             default:
                 break;
             }
@@ -343,6 +347,9 @@ namespace MobileGL::MG_Impl::CGLImpl {
         case kCGLPFARendererID:
             *value = pixelFormat->RendererId;
             return kCGLNoError;
+        case kCGLPFADisplayMask:
+            *value = pixelFormat->DisplayMask;
+            return kCGLNoError;
         case kCGLPFAOpenGLProfile:
             *value = pixelFormat->Profile;
             return kCGLNoError;
@@ -481,6 +488,32 @@ namespace MobileGL::MG_Impl::CGLImpl {
         return it == currentContexts.end() ? nullptr : it->second;
     }
 
+    CGLError SetVirtualScreen(CGLContextObj ctx, GLint screen) {
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(ctx);
+        if (!object) {
+            return kCGLBadContext;
+        }
+        if (screen != 0) {
+            return kCGLBadValue;
+        }
+        object->VirtualScreen = screen;
+        return kCGLNoError;
+    }
+
+    CGLError GetVirtualScreen(CGLContextObj ctx, GLint* screen) {
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(ctx);
+        if (!object) {
+            return kCGLBadContext;
+        }
+        if (!screen) {
+            return kCGLBadAddress;
+        }
+        *screen = object->VirtualScreen;
+        return kCGLNoError;
+    }
+
     CGLError SetParameter(CGLContextObj ctx, CGLContextParameter pname, const GLint* params) {
         const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
         auto* object = TryGetContext(ctx);

MobileGL/MG_Impl/CGLImpl/CGLImpl.h

@@ -32,6 +32,8 @@ namespace MobileGL::MG_Impl::CGLImpl {
 
     CGLError SetCurrentContext(CGLContextObj ctx);
     CGLContextObj GetCurrentContext();
+    CGLError SetVirtualScreen(CGLContextObj ctx, GLint screen);
+    CGLError GetVirtualScreen(CGLContextObj ctx, GLint* screen);
     CGLError SetParameter(CGLContextObj ctx, CGLContextParameter pname, const GLint* params);
     CGLError GetParameter(CGLContextObj ctx, CGLContextParameter pname, GLint* params);
     CGLError UpdateContext(CGLContextObj ctx);

MobileGL/MG_Impl/CGLImpl/Exporting/Definitions.cpp

@@ -71,6 +71,14 @@ MOBILEGL_CGL_API CGLContextObj CGLGetCurrentContext(void) {
     return MobileGL::MG_Impl::CGLImpl::GetCurrentContext();
 }
 
+MOBILEGL_CGL_API CGLError CGLSetVirtualScreen(CGLContextObj ctx, GLint screen) {
+    return MobileGL::MG_Impl::CGLImpl::SetVirtualScreen(ctx, screen);
+}
+
+MOBILEGL_CGL_API CGLError CGLGetVirtualScreen(CGLContextObj ctx, GLint* screen) {
+    return MobileGL::MG_Impl::CGLImpl::GetVirtualScreen(ctx, screen);
+}
+
 MOBILEGL_CGL_API CGLError CGLSetParameter(CGLContextObj ctx, CGLContextParameter pname, const GLint* params) {
     return MobileGL::MG_Impl::CGLImpl::SetParameter(ctx, pname, params);
 }

MobileGL/MG_Impl/DyldInterpose/DyldInterpose.cpp

@@ -10,8 +10,12 @@
 
 #if defined(__APPLE__)
 
+#include "MG_Impl/CGLImpl/CGLImpl.h"
 #include "MG_Impl/GetProcAddress.h"
 
+#include <CoreGraphics/CoreGraphics.h>
+#include <CoreVideo/CVDisplayLink.h>
+#include <cstdint>
 #include <dlfcn.h>
 
 namespace {
@@ -47,10 +51,52 @@ namespace {
         return dlsym(handle, symbol);
     }
 
+    CGDirectDisplayID DisplayForMask(GLint displayMask) {
+        constexpr std::uint32_t MaxDisplays = sizeof(CGOpenGLDisplayMask) * 8;
+        CGDirectDisplayID displays[MaxDisplays] = {};
+        std::uint32_t displayCount = 0;
+        if (displayMask != 0 &&
+            CGGetActiveDisplayList(MaxDisplays, displays, &displayCount) == kCGErrorSuccess) {
+            const auto mask = static_cast<CGOpenGLDisplayMask>(displayMask);
+            for (std::uint32_t i = 0; i < displayCount; ++i) {
+                if ((CGDisplayIDToOpenGLDisplayMask(displays[i]) & mask) != 0) {
+                    return displays[i];
+                }
+            }
+        }
+        return CGMainDisplayID();
+    }
+
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wdeprecated-declarations"
+    CVReturn MobileGLCVDisplayLinkSetCurrentCGDisplayFromOpenGLContext(
+        CVDisplayLinkRef displayLink,
+        CGLContextObj context,
+        CGLPixelFormatObj pixelFormat) {
+        GLint virtualScreen = 0;
+        if (MobileGL::MG_Impl::CGLImpl::GetVirtualScreen(context, &virtualScreen) == kCGLNoError) {
+            GLint displayMask = 0;
+            if (!displayLink ||
+                MobileGL::MG_Impl::CGLImpl::DescribePixelFormat(
+                    pixelFormat, virtualScreen, kCGLPFADisplayMask, &displayMask) != kCGLNoError) {
+                return kCVReturnInvalidArgument;
+            }
+            return CVDisplayLinkSetCurrentCGDisplay(displayLink, DisplayForMask(displayMask));
+        }
+
+        using OriginalFunction = CVReturn (*)(CVDisplayLinkRef, CGLContextObj, CGLPixelFormatObj);
+        static const auto original = reinterpret_cast<OriginalFunction>(
+            dlsym(RTLD_NEXT, "CVDisplayLinkSetCurrentCGDisplayFromOpenGLContext"));
+        return original ? original(displayLink, context, pixelFormat) : kCVReturnError;
+    }
+
     __attribute__((used)) static const DyldInterposeEntry kMobileGLDyldInterpose[]
         __attribute__((section("__DATA,__interpose"))) = {
             {reinterpret_cast<const void*>(MobileGLDlsym), reinterpret_cast<const void*>(dlsym)},
+            {reinterpret_cast<const void*>(MobileGLCVDisplayLinkSetCurrentCGDisplayFromOpenGLContext),
+             reinterpret_cast<const void*>(CVDisplayLinkSetCurrentCGDisplayFromOpenGLContext)},
         };
+#pragma clang diagnostic pop
 } // namespace
 
 #endif

MobileGL/MG_Impl/DyldInterpose/ExportedSymbols.txt

@@ -0,0 +1,10 @@
+# Public CGL entry points.
+_CGL*
+
+# Public EGL entry points.
+_egl*
+
+# Public OpenGL and GLX entry points. OpenGL function names always use an
+# uppercase letter or digit after the "gl" prefix; excluding lowercase here
+# deliberately prevents glslang_* from matching this pattern.
+_gl[A-Z0-9]*

MobileGL/MG_Impl/EGLImpl/EGLImpl.cpp

@@ -8,6 +8,7 @@
 
 #include "EGLImpl.h"
 #include "../GetProcAddress.h"
+#include <Init.h>
 #include <MG_Backend/BackendObjects.h>
 #include <MG_State/EGLState/Core.h>
 #include <mutex>
@@ -25,6 +26,17 @@ namespace MobileGL::MG_Impl::EGLImpl {
             return MG_State::pEGLContext.get();
         }
 
+        // Entry points that can legitimately be an application's FIRST EGL
+        // call (display/proc-address/string queries) lazily bring MobileGL
+        // up here, so the library needs no static constructor and can
+        // re-initialize after the last eglTerminate tore everything down.
+        // Teardown-ish entry points keep using GetState() and fail benignly
+        // when MobileGL is not initialized.
+        EGLStateContext* GetStateEnsureInitialized() {
+            MobileGL::EnsureInitialized();
+            return GetState();
+        }
+
         MG_Backend::BackendObject* GetBackendObject(EGLStateContext* state) {
             auto* backendObject = MG_Backend::pActiveBackendObject.get();
             if (!backendObject && state) {
@@ -49,6 +61,8 @@ namespace MobileGL::MG_Impl::EGLImpl {
             return MG_Backend::WindowBackend::Android;
 #elif defined(__APPLE__)
             return MG_Backend::WindowBackend::MetalLayer;
+#elif defined(_WIN32)
+            return MG_Backend::WindowBackend::Win32;
 #elif defined(__linux__)
             return MG_Backend::WindowBackend::X11;
 #else
@@ -187,7 +201,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     EGLBoolean Initialize(EGLDisplay dpy, EGLint* major, EGLint* minor) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_FALSE;
         }
@@ -208,7 +222,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     EGLDisplay GetDisplay(NativeDisplayType display) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_NO_DISPLAY;
         }
@@ -313,6 +327,14 @@ namespace MobileGL::MG_Impl::EGLImpl {
         if (auto* backendObject = MG_Backend::pActiveBackendObject.get()) {
             backendObject->ReleaseEGLResources();
         }
+        // The last initialized display is gone and nothing is current on any
+        // thread: tear the whole library down deterministically inside the
+        // EGL lifecycle (backend, GL/EGL state, glslang). A later EGL call
+        // re-initializes lazily via GetStateEnsureInitialized(); process exit
+        // then has nothing left to destroy.
+        if (!state->HasAnyInitializedDisplay() && !state->HasAnyCurrentContext()) {
+            MobileGL::Destroy();
+        }
         return EGL_TRUE;
     }
 
@@ -345,7 +367,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     EGLBoolean BindAPI(EGLenum api) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_FALSE;
         }
@@ -378,7 +400,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     char const* QueryString(EGLDisplay display, EGLint name) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return nullptr;
         }
@@ -641,7 +663,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     EGLDisplay GetPlatformDisplay(EGLenum platform, void* native_display, const EGLAttrib* attrib_list) {
         (void)attrib_list;
 
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_NO_DISPLAY;
         }
@@ -737,6 +759,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         if (!name) {
             return nullptr;
         }
+        MobileGL::EnsureInitialized();
 
         MGLOG_D("eglGetProcAddress(%s)", name);
         void* proc = MG_Impl::GetProcAddress(name);

MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp

@@ -2144,6 +2144,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     namespace FramebufferImpl {
-        UniquePtr<DefaultFramebufferInfo> pDefaultFramebufferInfo;
+        // Leak-at-exit storage; see GlobalObjects.cpp.
+        UniquePtr<DefaultFramebufferInfo>& pDefaultFramebufferInfo = *new UniquePtr<DefaultFramebufferInfo>();
     } // namespace FramebufferImpl
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h

@@ -78,6 +78,6 @@ namespace MobileGL::MG_Impl::GLImpl {
             SharedPtr<MG_State::GLState::ITextureObject> stencilAttachment;
         };
 
-        extern UniquePtr<DefaultFramebufferInfo> pDefaultFramebufferInfo;
+        extern UniquePtr<DefaultFramebufferInfo>& pDefaultFramebufferInfo;
     } // namespace FramebufferImpl
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Sync/GL_Sync.cpp

@@ -133,4 +133,33 @@ namespace MobileGL::MG_Impl::GLImpl {
             values[0] = value;
         }
     }
+
+    void DestroyAllSyncObjects() {
+        // Detach the registry under the lock, release outside it. Entries the app
+        // already deleted were erased by DeleteSync, so nothing here double-frees;
+        // a DeleteSync racing this sweep finds an empty registry and returns. A
+        // thread still blocked inside ClientWaitSync/GetSynciv during teardown
+        // holds a raw SyncObject* these deletes invalidate - the same undefined
+        // race an app-driven DeleteSync already has.
+        UnorderedMap<GLsync, SyncObject*> orphans;
+        {
+            const std::lock_guard<std::mutex> lock(g_syncObjectsMutex);
+            orphans.swap(g_liveSyncObjects);
+        }
+        if (orphans.empty()) {
+            return;
+        }
+        // Both backends' DeleteSync only free the heap wrapper once their GL
+        // context/renderer is gone (generation/current-thread guards), so this is
+        // safe after the backend has released its EGL resources - but not after
+        // the function table itself is cleared.
+        const auto backendDeleteSync = MG_Backend::gBackendFunctionsTable.GL.DeleteSync;
+        for (const auto& [_, syncObject] : orphans) {
+            if (backendDeleteSync && syncObject->backendHandle) {
+                backendDeleteSync(syncObject->backendHandle);
+            }
+            delete syncObject;
+        }
+        MGLOG_D("DestroyAllSyncObjects: reclaimed %zu sync object(s) the app left undeleted", orphans.size());
+    }
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Sync/GL_Sync.h

@@ -16,4 +16,12 @@ namespace MobileGL::MG_Impl::GLImpl {
     void WaitSync(GLsync sync, GLbitfield flags, GLuint64 timeout);
     void DeleteSync(GLsync sync);
     void GetSynciv(GLsync sync, GLenum pname, GLsizei bufSize, GLsizei* length, GLint* values);
+    // Destroys every still-registered sync object exactly as DeleteSync would.
+    // GL requires syncs 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.
+    void DestroyAllSyncObjects();
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Texture/ProxyTexture.cpp

@@ -14,7 +14,8 @@
 #include <MG_State/GLState/TextureState/TextureObjectStubs.h>
 
 namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
-    UniquePtr<ProxyTextureManager> pProxyTextureManager;
+    // Leak-at-exit storage; see GlobalObjects.cpp.
+    UniquePtr<ProxyTextureManager>& pProxyTextureManager = *new UniquePtr<ProxyTextureManager>();
 
     Bool IsProxyTextureTarget(TextureUploadTarget target) {
         switch (target) {

MobileGL/MG_Impl/GLImpl/Texture/ProxyTexture.h

@@ -23,5 +23,5 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         UnorderedMap<TextureUploadTarget, SharedPtr<MG_State::GLState::ITextureObject>> m_proxyTexturesMap;
     };
 
-    extern UniquePtr<ProxyTextureManager> pProxyTextureManager;
+    extern UniquePtr<ProxyTextureManager>& pProxyTextureManager;
 } // namespace MobileGL::MG_Impl::GLImpl::TextureImpl

MobileGL/MG_Impl/GetProcAddress.cpp

@@ -85,6 +85,8 @@ namespace MobileGL::MG_Impl {
         GETPROC(CGLGetPixelFormat, name);
         GETPROC(CGLSetCurrentContext, name);
         GETPROC(CGLGetCurrentContext, name);
+        GETPROC(CGLSetVirtualScreen, name);
+        GETPROC(CGLGetVirtualScreen, name);
         GETPROC(CGLSetParameter, name);
         GETPROC(CGLGetParameter, name);
         GETPROC(CGLUpdateContext, name);

MobileGL/MG_Impl/NSOpenGLImpl/NSOpenGLImpl.cpp

@@ -29,10 +29,19 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
         char kContextViewKey;
         char kContextLayerKey;
 
-        std::once_flag g_installOnce;
         IMP g_pixelFormatDealloc = nullptr;
         IMP g_contextDealloc = nullptr;
 
+        std::mutex& HookInstallMutex() {
+            static auto* mutex = new std::mutex();
+            return *mutex;
+        }
+
+        Bool& HooksInstalled() {
+            static auto* installed = new Bool(false);
+            return *installed;
+        }
+
         template <typename Fn>
         Fn ObjcMsgSend() {
             return reinterpret_cast<Fn>(objc_msgSend);
@@ -431,12 +440,12 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             method_setImplementation(method, replacement);
         }
 
-        void InstallHooksOnce() {
+        Bool InstallHooksOnce() {
             Class pixelFormatClass = objc_getClass("NSOpenGLPixelFormat");
             Class contextClass = objc_getClass("NSOpenGLContext");
             if (!pixelFormatClass || !contextClass) {
                 MGLOG_W("NSOpenGLImpl: NSOpenGL classes are not loaded; hooks not installed");
-                return;
+                return false;
             }
 
             ReplaceInstanceMethod(pixelFormatClass, "initWithAttributes:",
@@ -471,11 +480,34 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             ReplaceInstanceMethod(contextClass, "dealloc", reinterpret_cast<IMP>(ContextDealloc), &g_contextDealloc);
 
             MGLOG_I("NSOpenGLImpl hooks installed");
+            return true;
         }
     } // namespace
 
     void InstallHooks() {
-        std::call_once(g_installOnce, InstallHooksOnce);
+        const std::lock_guard<std::mutex> lock(HookInstallMutex());
+        if (!HooksInstalled()) {
+            // Do not permanently consume the install attempt when the OpenGL
+            // framework has not registered its Objective-C classes yet. The
+            // dyld bootstrap normally runs after framework dependencies, but
+            // an explicitly loaded/static-linked MobileGL can arrive earlier.
+            HooksInstalled() = InstallHooksOnce();
+        }
     }
 } // namespace MobileGL::MG_Impl::NSOpenGLImpl
+
+namespace {
+    // SDL's Cocoa backend creates NSOpenGLPixelFormat/NSOpenGLContext before
+    // its first dlsym("glGetString") or other MobileGL host-API call. Install
+    // only the lightweight Objective-C dispatch hooks while the injected dylib
+    // is loading so those first Cocoa objects are routed through CGLImpl. The
+    // hooked context constructor reaches EGLImpl::GetDisplay(), which performs
+    // the full, thread-safe MobileGL initialization outside this bootstrap.
+    //
+    // There is intentionally no matching destructor: backend teardown remains
+    // owned by the EGL lifecycle and process-exit globals remain leak-at-exit.
+    __attribute__((constructor)) void BootstrapNSOpenGLHooks() {
+        MobileGL::MG_Impl::NSOpenGLImpl::InstallHooks();
+    }
+} // namespace
 #endif

MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.cpp

@@ -0,0 +1,156 @@
+// MobileGL - MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.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
+
+// wingdi.h declares most wgl* entry points as WINGDIAPI (__declspec(dllimport)),
+// which would reject our definitions. _GDI32_ is the SDK's "I am the module that
+// implements these" switch: it turns WINGDIAPI into a plain declaration. It must
+// be defined before the first windows.h inclusion in this translation unit.
+#if defined(_WIN32) && !defined(_GDI32_)
+#define _GDI32_ 1
+#endif
+
+#include <Includes.h>
+
+#if defined(_WIN32)
+#include "../WGLImpl.h"
+
+namespace WGL = MobileGL::MG_Impl::WGLImpl;
+
+// ---- Pixel-format entry points (gdi32 forwards ChoosePixelFormat/SetPixelFormat/
+// ---- DescribePixelFormat/GetPixelFormat/SwapBuffers into these exports) ----
+
+extern "C" int WINAPI wglChoosePixelFormat(HDC hdc, CONST PIXELFORMATDESCRIPTOR* ppfd) {
+    return WGL::ChoosePixelFormat(hdc, ppfd);
+}
+
+extern "C" int WINAPI wglDescribePixelFormat(HDC hdc, int iPixelFormat, UINT nBytes,
+                                                   LPPIXELFORMATDESCRIPTOR ppfd) {
+    return WGL::DescribePixelFormat(hdc, iPixelFormat, nBytes, ppfd);
+}
+
+extern "C" int WINAPI wglGetPixelFormat(HDC hdc) {
+    return WGL::GetPixelFormat(hdc);
+}
+
+extern "C" BOOL WINAPI wglSetPixelFormat(HDC hdc, int iPixelFormat, CONST PIXELFORMATDESCRIPTOR* ppfd) {
+    return WGL::SetPixelFormat(hdc, iPixelFormat, ppfd);
+}
+
+extern "C" BOOL WINAPI wglSwapBuffers(HDC hdc) {
+    return WGL::SwapBuffers(hdc);
+}
+
+// ---- Context management ----
+
+extern "C" HGLRC WINAPI wglCreateContext(HDC hdc) {
+    return WGL::CreateContext(hdc);
+}
+
+extern "C" HGLRC WINAPI wglCreateLayerContext(HDC hdc, int iLayerPlane) {
+    return iLayerPlane == 0 ? WGL::CreateContext(hdc) : nullptr;
+}
+
+extern "C" BOOL WINAPI wglCopyContext(HGLRC, HGLRC, UINT) {
+    MGLOG_W("wglCopyContext is not supported");
+    SetLastError(ERROR_NOT_SUPPORTED);
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglDeleteContext(HGLRC hglrc) {
+    return WGL::DeleteContext(hglrc);
+}
+
+extern "C" HGLRC WINAPI wglGetCurrentContext(VOID) {
+    return WGL::GetCurrentContext();
+}
+
+extern "C" HDC WINAPI wglGetCurrentDC(VOID) {
+    return WGL::GetCurrentDC();
+}
+
+extern "C" BOOL WINAPI wglMakeCurrent(HDC hdc, HGLRC hglrc) {
+    return WGL::MakeCurrent(hdc, hglrc);
+}
+
+extern "C" BOOL WINAPI wglShareLists(HGLRC hglrcShare, HGLRC hglrcDest) {
+    return WGL::ShareLists(hglrcShare, hglrcDest);
+}
+
+// ---- Proc address ----
+
+extern "C" PROC WINAPI wglGetProcAddress(LPCSTR lpszProc) {
+    return WGL::GetProcAddress(lpszProc);
+}
+
+extern "C" PROC WINAPI wglGetDefaultProcAddress(LPCSTR lpszProc) {
+    return WGL::GetProcAddress(lpszProc);
+}
+
+// ---- Layer planes and palettes (unsupported; overlay planes do not exist here) ----
+
+extern "C" BOOL WINAPI wglDescribeLayerPlane(HDC, int, int, UINT, LPLAYERPLANEDESCRIPTOR) {
+    return FALSE;
+}
+
+extern "C" int WINAPI wglSetLayerPaletteEntries(HDC, int, int, int, CONST COLORREF*) {
+    return 0;
+}
+
+extern "C" int WINAPI wglGetLayerPaletteEntries(HDC, int, int, int, COLORREF*) {
+    return 0;
+}
+
+extern "C" BOOL WINAPI wglRealizeLayerPalette(HDC, int, BOOL) {
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglSwapLayerBuffers(HDC hdc, UINT fuPlanes) {
+    if (fuPlanes & WGL_SWAP_MAIN_PLANE) {
+        return WGL::SwapBuffers(hdc);
+    }
+    return FALSE;
+}
+
+extern "C" DWORD WINAPI wglSwapMultipleBuffers(UINT n, CONST WGLSWAP* ps) {
+    if (!ps) {
+        return 0;
+    }
+    DWORD swapped = 0;
+    for (UINT i = 0; i < n; ++i) {
+        if (WGL::SwapBuffers(ps[i].hdc)) {
+            ++swapped;
+        }
+    }
+    return swapped;
+}
+
+// ---- Font rendering (legacy immediate-mode feature; not supported) ----
+
+extern "C" BOOL WINAPI wglUseFontBitmapsA(HDC, DWORD, DWORD, DWORD) {
+    MGLOG_W("wglUseFontBitmapsA is not supported");
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglUseFontBitmapsW(HDC, DWORD, DWORD, DWORD) {
+    MGLOG_W("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");
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglUseFontOutlinesW(HDC, DWORD, DWORD, DWORD, FLOAT, FLOAT, int,
+                                                 LPGLYPHMETRICSFLOAT) {
+    MGLOG_W("wglUseFontOutlinesW is not supported");
+    return FALSE;
+}
+
+#endif // _WIN32

MobileGL/MG_Impl/WGLImpl/Exporting/wgl.def

@@ -0,0 +1,30 @@
+; MobileGL WGL exports. The wgl* entry points are defined without
+; __declspec(dllexport) because wingdi.h pre-declares them (with _GDI32_ they
+; become plain declarations, and MSVC rejects adding dllexport afterwards),
+; so this .def file is what actually exports them from the DLL.
+EXPORTS
+    wglChoosePixelFormat
+    wglCopyContext
+    wglCreateContext
+    wglCreateLayerContext
+    wglDeleteContext
+    wglDescribeLayerPlane
+    wglDescribePixelFormat
+    wglGetCurrentContext
+    wglGetCurrentDC
+    wglGetDefaultProcAddress
+    wglGetLayerPaletteEntries
+    wglGetPixelFormat
+    wglGetProcAddress
+    wglMakeCurrent
+    wglRealizeLayerPalette
+    wglSetLayerPaletteEntries
+    wglSetPixelFormat
+    wglShareLists
+    wglSwapBuffers
+    wglSwapLayerBuffers
+    wglSwapMultipleBuffers
+    wglUseFontBitmapsA
+    wglUseFontBitmapsW
+    wglUseFontOutlinesA
+    wglUseFontOutlinesW

MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp

@@ -0,0 +1,732 @@
+// MobileGL - MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "WGLImpl.h"
+
+#if defined(_WIN32)
+#include "../EGLImpl/EGLImpl.h"
+#include "../GetProcAddress.h"
+#include <Init.h>
+
+namespace MobileGL::MG_Impl::WGLImpl {
+    namespace {
+        // WGL_ARB_pixel_format
+        constexpr int WGL_NUMBER_PIXEL_FORMATS_ARB = 0x2000;
+        constexpr int WGL_DRAW_TO_WINDOW_ARB = 0x2001;
+        constexpr int WGL_DRAW_TO_BITMAP_ARB = 0x2002;
+        constexpr int WGL_ACCELERATION_ARB = 0x2003;
+        constexpr int WGL_NEED_PALETTE_ARB = 0x2004;
+        constexpr int WGL_NEED_SYSTEM_PALETTE_ARB = 0x2005;
+        constexpr int WGL_SWAP_LAYER_BUFFERS_ARB = 0x2006;
+        constexpr int WGL_SWAP_METHOD_ARB = 0x2007;
+        constexpr int WGL_NUMBER_OVERLAYS_ARB = 0x2008;
+        constexpr int WGL_NUMBER_UNDERLAYS_ARB = 0x2009;
+        constexpr int WGL_TRANSPARENT_ARB = 0x200A;
+        constexpr int WGL_SHARE_DEPTH_ARB = 0x200C;
+        constexpr int WGL_SHARE_STENCIL_ARB = 0x200D;
+        constexpr int WGL_SHARE_ACCUM_ARB = 0x200E;
+        constexpr int WGL_SUPPORT_GDI_ARB = 0x200F;
+        constexpr int WGL_SUPPORT_OPENGL_ARB = 0x2010;
+        constexpr int WGL_DOUBLE_BUFFER_ARB = 0x2011;
+        constexpr int WGL_STEREO_ARB = 0x2012;
+        constexpr int WGL_PIXEL_TYPE_ARB = 0x2013;
+        constexpr int WGL_COLOR_BITS_ARB = 0x2014;
+        constexpr int WGL_RED_BITS_ARB = 0x2015;
+        constexpr int WGL_RED_SHIFT_ARB = 0x2016;
+        constexpr int WGL_GREEN_BITS_ARB = 0x2017;
+        constexpr int WGL_GREEN_SHIFT_ARB = 0x2018;
+        constexpr int WGL_BLUE_BITS_ARB = 0x2019;
+        constexpr int WGL_BLUE_SHIFT_ARB = 0x201A;
+        constexpr int WGL_ALPHA_BITS_ARB = 0x201B;
+        constexpr int WGL_ALPHA_SHIFT_ARB = 0x201C;
+        constexpr int WGL_ACCUM_BITS_ARB = 0x201D;
+        constexpr int WGL_ACCUM_RED_BITS_ARB = 0x201E;
+        constexpr int WGL_ACCUM_GREEN_BITS_ARB = 0x201F;
+        constexpr int WGL_ACCUM_BLUE_BITS_ARB = 0x2020;
+        constexpr int WGL_ACCUM_ALPHA_BITS_ARB = 0x2021;
+        constexpr int WGL_DEPTH_BITS_ARB = 0x2022;
+        constexpr int WGL_STENCIL_BITS_ARB = 0x2023;
+        constexpr int WGL_AUX_BUFFERS_ARB = 0x2024;
+        constexpr int WGL_NO_ACCELERATION_ARB = 0x2025;
+        constexpr int WGL_FULL_ACCELERATION_ARB = 0x2027;
+        constexpr int WGL_SWAP_EXCHANGE_ARB = 0x2028;
+        constexpr int WGL_TYPE_RGBA_ARB = 0x202B;
+        // WGL_ARB_multisample
+        constexpr int WGL_SAMPLE_BUFFERS_ARB = 0x2041;
+        constexpr int WGL_SAMPLES_ARB = 0x2042;
+        // WGL_ARB_create_context / _profile / _no_error
+        constexpr int WGL_CONTEXT_MAJOR_VERSION_ARB = 0x2091;
+        constexpr int WGL_CONTEXT_MINOR_VERSION_ARB = 0x2092;
+        constexpr int WGL_CONTEXT_LAYER_PLANE_ARB = 0x2093;
+        constexpr int WGL_CONTEXT_FLAGS_ARB = 0x2094;
+        constexpr int WGL_CONTEXT_PROFILE_MASK_ARB = 0x9126;
+        constexpr int WGL_CONTEXT_DEBUG_BIT_ARB = 0x0001;
+        constexpr int WGL_CONTEXT_FORWARD_COMPATIBLE_BIT_ARB = 0x0002;
+        constexpr int WGL_CONTEXT_CORE_PROFILE_BIT_ARB = 0x00000001;
+        constexpr int WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB = 0x00000002;
+        constexpr int WGL_CONTEXT_OPENGL_NO_ERROR_ARB = 0x31B3;
+        constexpr DWORD ERROR_INVALID_VERSION_ARB = 0x2095;
+        constexpr DWORD ERROR_INVALID_PROFILE_ARB = 0x2096;
+
+        struct PixelFormatInfo {
+            GLint AlphaBits;
+            GLint DepthBits;
+            GLint StencilBits;
+        };
+
+        // Mirrors the two EGLState configs (RGBA8 + depth24, stencil 8 / stencil 0).
+        constexpr PixelFormatInfo kPixelFormats[] = {
+            {8, 24, 8},
+            {8, 24, 0},
+        };
+        constexpr int kPixelFormatCount = static_cast<int>(std::size(kPixelFormats));
+
+        struct ContextObject {
+            EGLDisplay Display = EGL_NO_DISPLAY;
+            EGLConfig Config = nullptr;
+            EGLContext Context = EGL_NO_CONTEXT;
+        };
+
+        struct WindowSurface {
+            EGLDisplay Display = EGL_NO_DISPLAY;
+            EGLSurface Surface = EGL_NO_SURFACE;
+            Uint32 Width = 0;
+            Uint32 Height = 0;
+        };
+
+        std::recursive_mutex& RegistryMutex() {
+            static auto* mutex = new std::recursive_mutex();
+            return *mutex;
+        }
+
+        UnorderedMap<HGLRC, ContextObject>& Contexts() {
+            static auto* contexts = new UnorderedMap<HGLRC, ContextObject>();
+            return *contexts;
+        }
+
+        UnorderedMap<HWND, WindowSurface>& WindowSurfaces() {
+            static auto* surfaces = new UnorderedMap<HWND, WindowSurface>();
+            return *surfaces;
+        }
+
+        UnorderedMap<HWND, int>& WindowPixelFormats() {
+            static auto* formats = new UnorderedMap<HWND, int>();
+            return *formats;
+        }
+
+        Uint64& NextContextHandle() {
+            static auto* handle = new Uint64(0x10000);
+            return *handle;
+        }
+
+        struct ThreadCurrent {
+            HDC DC = nullptr;
+            HGLRC Context = nullptr;
+        };
+        thread_local ThreadCurrent t_current;
+
+        Int& SwapIntervalShadow() {
+            static auto* interval = new Int(1);
+            return *interval;
+        }
+
+        void EnsureInitialized() {
+            // Initialize() loads backend libraries and glslang, which must not run
+            // under the loader lock; first WGL call is the earliest safe moment.
+            // MobileGL::EnsureInitialized (not a local once_flag) so a fresh init
+            // can follow a full teardown from the last eglTerminate.
+            MobileGL::EnsureInitialized();
+        }
+
+        EGLDisplay EnsureDisplay() {
+            EnsureInitialized();
+            EGLDisplay display = EGLImpl::GetDisplay(EGL_DEFAULT_DISPLAY);
+            if (display == EGL_NO_DISPLAY) {
+                return EGL_NO_DISPLAY;
+            }
+            if (!EGLImpl::Initialize(display, nullptr, nullptr)) {
+                return EGL_NO_DISPLAY;
+            }
+            return display;
+        }
+
+        HGLRC EncodeContext(Uint64 handle) {
+            return reinterpret_cast<HGLRC>(static_cast<SizeT>(handle));
+        }
+
+        ContextObject* TryGetContext(HGLRC hglrc) {
+            auto& contexts = Contexts();
+            auto it = contexts.find(hglrc);
+            return it == contexts.end() ? nullptr : &it->second;
+        }
+
+        const PixelFormatInfo& PixelFormatForWindow(HWND hwnd) {
+            auto& formats = WindowPixelFormats();
+            auto it = formats.find(hwnd);
+            int index = it == formats.end() ? 1 : it->second;
+            if (index < 1 || index > kPixelFormatCount) {
+                index = 1;
+            }
+            return kPixelFormats[index - 1];
+        }
+
+        Bool QueryClientSize(HWND hwnd, Uint32& width, Uint32& height) {
+            RECT rect{};
+            if (!GetClientRect(hwnd, &rect)) {
+                return false;
+            }
+            width = static_cast<Uint32>(std::max<LONG>(rect.right - rect.left, 1));
+            height = static_cast<Uint32>(std::max<LONG>(rect.bottom - rect.top, 1));
+            return true;
+        }
+
+        // The backends never query the HWND client size themselves; the WGL layer
+        // owns size discovery and pushes changes through the internal resize hook
+        // (same contract as the macOS CGL layer).
+        void SyncSurfaceSize(HWND hwnd, WindowSurface& surface) {
+            Uint32 width = 0;
+            Uint32 height = 0;
+            if (!QueryClientSize(hwnd, width, height)) {
+                return;
+            }
+            if (width == surface.Width && height == surface.Height) {
+                return;
+            }
+            if (EGLImpl::ResizePlatformWindowSurface(surface.Display, surface.Surface,
+                                                     static_cast<EGLint>(width), static_cast<EGLint>(height))) {
+                surface.Width = width;
+                surface.Height = height;
+            }
+        }
+
+        WindowSurface* EnsureWindowSurface(HWND hwnd, const ContextObject& context) {
+            auto& surfaces = WindowSurfaces();
+            auto it = surfaces.find(hwnd);
+            if (it != surfaces.end()) {
+                SyncSurfaceSize(hwnd, it->second);
+                return &it->second;
+            }
+
+            Uint32 width = 0;
+            Uint32 height = 0;
+            if (!QueryClientSize(hwnd, width, height)) {
+                MGLOG_E("wgl: GetClientRect failed for HWND %p", hwnd);
+                return nullptr;
+            }
+
+            const EGLAttrib attribs[] = {
+                EGL_WIDTH, static_cast<EGLAttrib>(width),
+                EGL_HEIGHT, static_cast<EGLAttrib>(height),
+                EGL_NONE,
+            };
+            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);
+                return nullptr;
+            }
+
+            WindowSurface record;
+            record.Display = context.Display;
+            record.Surface = surface;
+            record.Width = width;
+            record.Height = height;
+            auto [inserted, _] = surfaces.emplace(hwnd, record);
+            return &inserted->second;
+        }
+
+        HGLRC CreateContextFromEGLAttribs(HDC hdc, HGLRC share, const EGLint* contextAttribs) {
+            const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+            EGLDisplay display = EnsureDisplay();
+            if (display == EGL_NO_DISPLAY) {
+                MGLOG_E("wgl: no EGL display");
+                return nullptr;
+            }
+            EGLImpl::BindAPI(EGL_OPENGL_API);
+
+            EGLContext shareContext = EGL_NO_CONTEXT;
+            if (share) {
+                auto* shareObject = TryGetContext(share);
+                if (!shareObject) {
+                    SetLastError(ERROR_INVALID_HANDLE);
+                    return nullptr;
+                }
+                shareContext = shareObject->Context;
+            }
+
+            HWND hwnd = WindowFromDC(hdc);
+            const PixelFormatInfo& pixelFormat = PixelFormatForWindow(hwnd);
+            const EGLint configAttribs[] = {
+                EGL_RED_SIZE, 8,
+                EGL_GREEN_SIZE, 8,
+                EGL_BLUE_SIZE, 8,
+                EGL_ALPHA_SIZE, pixelFormat.AlphaBits,
+                EGL_DEPTH_SIZE, pixelFormat.DepthBits,
+                EGL_STENCIL_SIZE, pixelFormat.StencilBits,
+                EGL_SURFACE_TYPE, EGL_WINDOW_BIT | EGL_PBUFFER_BIT,
+                EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT,
+                EGL_NONE,
+            };
+            EGLConfig config = nullptr;
+            EGLint configCount = 0;
+            if (!EGLImpl::ChooseConfig(display, configAttribs, &config, 1, &configCount) || configCount <= 0) {
+                MGLOG_E("wgl: eglChooseConfig failed");
+                return nullptr;
+            }
+
+            EGLContext eglContext = EGLImpl::CreateContext(display, config, shareContext, contextAttribs);
+            if (eglContext == EGL_NO_CONTEXT) {
+                MGLOG_E("wgl: eglCreateContext failed");
+                return nullptr;
+            }
+
+            ContextObject object;
+            object.Display = display;
+            object.Config = config;
+            object.Context = eglContext;
+            const auto handle = EncodeContext(NextContextHandle()++);
+            Contexts()[handle] = object;
+            MGLOG_I("wgl: created context %p (EGL context %p)", handle, eglContext);
+            return handle;
+        }
+
+        // ---- WGL extension entry points (resolved via wglGetProcAddress only) ----
+
+        const char* WINAPI Ext_GetExtensionsStringARB(HDC) {
+            return "WGL_ARB_create_context WGL_ARB_create_context_no_error WGL_ARB_create_context_profile "
+                   "WGL_ARB_extensions_string WGL_ARB_pixel_format WGL_EXT_extensions_string WGL_EXT_swap_control";
+        }
+
+        const char* WINAPI Ext_GetExtensionsStringEXT() {
+            return Ext_GetExtensionsStringARB(nullptr);
+        }
+
+        HGLRC WINAPI Ext_CreateContextAttribsARB(HDC hdc, HGLRC hShareContext, const int* attribList) {
+            EnsureInitialized();
+            int major = 1;
+            int minor = 0;
+            int profileMask = 0;
+            int flags = 0;
+            if (attribList) {
+                for (SizeT i = 0; attribList[i] != 0; i += 2) {
+                    const int attrib = attribList[i];
+                    const int value = attribList[i + 1];
+                    switch (attrib) {
+                    case WGL_CONTEXT_MAJOR_VERSION_ARB:
+                        major = value;
+                        break;
+                    case WGL_CONTEXT_MINOR_VERSION_ARB:
+                        minor = value;
+                        break;
+                    case WGL_CONTEXT_PROFILE_MASK_ARB:
+                        profileMask = value;
+                        break;
+                    case WGL_CONTEXT_FLAGS_ARB:
+                        flags = value;
+                        break;
+                    case WGL_CONTEXT_LAYER_PLANE_ARB:
+                        if (value != 0) {
+                            SetLastError(ERROR_INVALID_PARAMETER);
+                            return nullptr;
+                        }
+                        break;
+                    case WGL_CONTEXT_OPENGL_NO_ERROR_ARB:
+                        // Accepted and ignored: MobileGL always validates.
+                        break;
+                    default:
+                        MGLOG_D("wglCreateContextAttribsARB: ignoring attrib 0x%04x = 0x%x", attrib, value);
+                        break;
+                    }
+                }
+            }
+
+            if (major < 1 || (profileMask & ~(WGL_CONTEXT_CORE_PROFILE_BIT_ARB |
+                                              WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB))) {
+                SetLastError(profileMask ? ERROR_INVALID_PROFILE_ARB : ERROR_INVALID_VERSION_ARB);
+                return nullptr;
+            }
+
+            Vector<EGLint> attribs = {
+                EGL_CONTEXT_MAJOR_VERSION, major,
+                EGL_CONTEXT_MINOR_VERSION, minor,
+            };
+            const Bool wantsCompat = (profileMask & WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB) != 0;
+            if (major > 3 || (major == 3 && minor >= 2) || profileMask != 0) {
+                attribs.push_back(EGL_CONTEXT_OPENGL_PROFILE_MASK);
+                attribs.push_back(wantsCompat ? EGL_CONTEXT_OPENGL_COMPATIBILITY_PROFILE_BIT
+                                              : EGL_CONTEXT_OPENGL_CORE_PROFILE_BIT);
+            }
+            if (flags & WGL_CONTEXT_FORWARD_COMPATIBLE_BIT_ARB) {
+                attribs.push_back(EGL_CONTEXT_OPENGL_FORWARD_COMPATIBLE);
+                attribs.push_back(EGL_TRUE);
+            }
+            if (flags & WGL_CONTEXT_DEBUG_BIT_ARB) {
+                attribs.push_back(EGL_CONTEXT_OPENGL_DEBUG);
+                attribs.push_back(EGL_TRUE);
+            }
+            attribs.push_back(EGL_NONE);
+
+            return CreateContextFromEGLAttribs(hdc, hShareContext, attribs.data());
+        }
+
+        BOOL WINAPI Ext_SwapIntervalEXT(int interval) {
+            const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+            EGLDisplay display = EnsureDisplay();
+            if (display == EGL_NO_DISPLAY) {
+                return FALSE;
+            }
+            if (interval < 0) {
+                // Adaptive vsync is not supported; clamp to regular vsync.
+                interval = 1;
+            }
+            EGLImpl::SwapInterval(display, interval);
+            SwapIntervalShadow() = interval;
+            return TRUE;
+        }
+
+        int WINAPI Ext_GetSwapIntervalEXT() {
+            const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+            return SwapIntervalShadow();
+        }
+
+        int PixelFormatAttribValue(int format, int attrib) {
+            const PixelFormatInfo& info = kPixelFormats[format - 1];
+            switch (attrib) {
+            case WGL_NUMBER_PIXEL_FORMATS_ARB:
+                return kPixelFormatCount;
+            case WGL_SUPPORT_OPENGL_ARB:
+            case WGL_DRAW_TO_WINDOW_ARB:
+            case WGL_DOUBLE_BUFFER_ARB:
+                return 1;
+            case WGL_ACCELERATION_ARB:
+                return WGL_FULL_ACCELERATION_ARB;
+            case WGL_PIXEL_TYPE_ARB:
+                return WGL_TYPE_RGBA_ARB;
+            case WGL_COLOR_BITS_ARB:
+                return 32;
+            case WGL_RED_BITS_ARB:
+            case WGL_GREEN_BITS_ARB:
+            case WGL_BLUE_BITS_ARB:
+                return 8;
+            case WGL_RED_SHIFT_ARB:
+                return 16;
+            case WGL_GREEN_SHIFT_ARB:
+                return 8;
+            case WGL_BLUE_SHIFT_ARB:
+                return 0;
+            case WGL_ALPHA_BITS_ARB:
+                return info.AlphaBits;
+            case WGL_ALPHA_SHIFT_ARB:
+                return 24;
+            case WGL_DEPTH_BITS_ARB:
+                return info.DepthBits;
+            case WGL_STENCIL_BITS_ARB:
+                return info.StencilBits;
+            case WGL_SWAP_METHOD_ARB:
+                return WGL_SWAP_EXCHANGE_ARB;
+            case WGL_DRAW_TO_BITMAP_ARB:
+            case WGL_NEED_PALETTE_ARB:
+            case WGL_NEED_SYSTEM_PALETTE_ARB:
+            case WGL_SWAP_LAYER_BUFFERS_ARB:
+            case WGL_NUMBER_OVERLAYS_ARB:
+            case WGL_NUMBER_UNDERLAYS_ARB:
+            case WGL_TRANSPARENT_ARB:
+            case WGL_SHARE_DEPTH_ARB:
+            case WGL_SHARE_STENCIL_ARB:
+            case WGL_SHARE_ACCUM_ARB:
+            case WGL_SUPPORT_GDI_ARB:
+            case WGL_STEREO_ARB:
+            case WGL_ACCUM_BITS_ARB:
+            case WGL_ACCUM_RED_BITS_ARB:
+            case WGL_ACCUM_GREEN_BITS_ARB:
+            case WGL_ACCUM_BLUE_BITS_ARB:
+            case WGL_ACCUM_ALPHA_BITS_ARB:
+            case WGL_AUX_BUFFERS_ARB:
+            case WGL_SAMPLE_BUFFERS_ARB:
+            case WGL_SAMPLES_ARB:
+            default:
+                return 0;
+            }
+        }
+
+        BOOL WINAPI Ext_GetPixelFormatAttribivARB(HDC, int iPixelFormat, int iLayerPlane, UINT nAttributes,
+                                                  const int* piAttributes, int* piValues) {
+            if (iLayerPlane != 0 || !piAttributes || !piValues) {
+                return FALSE;
+            }
+            // Format 0 is only valid for WGL_NUMBER_PIXEL_FORMATS_ARB queries.
+            if (iPixelFormat < 0 || iPixelFormat > kPixelFormatCount) {
+                return FALSE;
+            }
+            const int format = iPixelFormat == 0 ? 1 : iPixelFormat;
+            for (UINT i = 0; i < nAttributes; ++i) {
+                piValues[i] = PixelFormatAttribValue(format, piAttributes[i]);
+            }
+            return TRUE;
+        }
+
+        BOOL WINAPI Ext_GetPixelFormatAttribfvARB(HDC hdc, int iPixelFormat, int iLayerPlane, UINT nAttributes,
+                                                  const int* piAttributes, FLOAT* pfValues) {
+            if (!pfValues) {
+                return FALSE;
+            }
+            Vector<int> values(nAttributes);
+            if (!Ext_GetPixelFormatAttribivARB(hdc, iPixelFormat, iLayerPlane, nAttributes, piAttributes,
+                                               values.data())) {
+                return FALSE;
+            }
+            for (UINT i = 0; i < nAttributes; ++i) {
+                pfValues[i] = static_cast<FLOAT>(values[i]);
+            }
+            return TRUE;
+        }
+
+        BOOL WINAPI Ext_ChoosePixelFormatARB(HDC, const int* piAttribIList, const FLOAT*, UINT nMaxFormats,
+                                             int* piFormats, UINT* nNumFormats) {
+            if (!piFormats || !nNumFormats) {
+                return FALSE;
+            }
+            int wantedStencil = 0;
+            if (piAttribIList) {
+                for (SizeT i = 0; piAttribIList[i] != 0; i += 2) {
+                    if (piAttribIList[i] == WGL_STENCIL_BITS_ARB) {
+                        wantedStencil = piAttribIList[i + 1];
+                    }
+                }
+            }
+            UINT count = 0;
+            const int preferred = wantedStencil > 0 ? 1 : 2;
+            const int fallback = wantedStencil > 0 ? 2 : 1;
+            if (count < nMaxFormats) {
+                piFormats[count++] = preferred;
+            }
+            if (count < nMaxFormats) {
+                piFormats[count++] = fallback;
+            }
+            *nNumFormats = count;
+            return TRUE;
+        }
+
+        struct WGLExtensionProc {
+            const char* Name;
+            PROC Proc;
+        };
+
+        const WGLExtensionProc kWGLExtensionProcs[] = {
+            {"wglGetExtensionsStringARB", reinterpret_cast<PROC>(Ext_GetExtensionsStringARB)},
+            {"wglGetExtensionsStringEXT", reinterpret_cast<PROC>(Ext_GetExtensionsStringEXT)},
+            {"wglCreateContextAttribsARB", reinterpret_cast<PROC>(Ext_CreateContextAttribsARB)},
+            {"wglSwapIntervalEXT", reinterpret_cast<PROC>(Ext_SwapIntervalEXT)},
+            {"wglGetSwapIntervalEXT", reinterpret_cast<PROC>(Ext_GetSwapIntervalEXT)},
+            {"wglGetPixelFormatAttribivARB", reinterpret_cast<PROC>(Ext_GetPixelFormatAttribivARB)},
+            {"wglGetPixelFormatAttribfvARB", reinterpret_cast<PROC>(Ext_GetPixelFormatAttribfvARB)},
+            {"wglChoosePixelFormatARB", reinterpret_cast<PROC>(Ext_ChoosePixelFormatARB)},
+        };
+    } // namespace
+
+    int ChoosePixelFormat(HDC hdc, const PIXELFORMATDESCRIPTOR* pfd) {
+        EnsureInitialized();
+        MGLOG_D("wglChoosePixelFormat(hdc=%p)", hdc);
+        // Format 1 (RGBA8 + depth24/stencil8) satisfies every request; a format
+        // exceeding the asked-for capabilities is a legal ChoosePixelFormat answer.
+        (void)pfd;
+        return 1;
+    }
+
+    int DescribePixelFormat(HDC hdc, int format, UINT size, PIXELFORMATDESCRIPTOR* pfd) {
+        EnsureInitialized();
+        MGLOG_D("wglDescribePixelFormat(hdc=%p, format=%d)", hdc, format);
+        if (!pfd) {
+            return kPixelFormatCount;
+        }
+        if (size < sizeof(PIXELFORMATDESCRIPTOR) || format < 1 || format > kPixelFormatCount) {
+            return 0;
+        }
+
+        const PixelFormatInfo& info = kPixelFormats[format - 1];
+        std::memset(pfd, 0, sizeof(PIXELFORMATDESCRIPTOR));
+        pfd->nSize = sizeof(PIXELFORMATDESCRIPTOR);
+        pfd->nVersion = 1;
+        pfd->dwFlags = PFD_DRAW_TO_WINDOW | PFD_SUPPORT_OPENGL | PFD_DOUBLEBUFFER | PFD_SWAP_EXCHANGE
+#if defined(PFD_SUPPORT_COMPOSITION)
+                       | PFD_SUPPORT_COMPOSITION
+#endif
+            ;
+        pfd->iPixelType = PFD_TYPE_RGBA;
+        pfd->cColorBits = 32;
+        pfd->cRedBits = 8;
+        pfd->cRedShift = 16;
+        pfd->cGreenBits = 8;
+        pfd->cGreenShift = 8;
+        pfd->cBlueBits = 8;
+        pfd->cBlueShift = 0;
+        pfd->cAlphaBits = static_cast<BYTE>(info.AlphaBits);
+        pfd->cAlphaShift = 24;
+        pfd->cDepthBits = static_cast<BYTE>(info.DepthBits);
+        pfd->cStencilBits = static_cast<BYTE>(info.StencilBits);
+        pfd->iLayerType = PFD_MAIN_PLANE;
+        return kPixelFormatCount;
+    }
+
+    int GetPixelFormat(HDC hdc) {
+        EnsureInitialized();
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            return 0;
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto& formats = WindowPixelFormats();
+        auto it = formats.find(hwnd);
+        return it == formats.end() ? 0 : it->second;
+    }
+
+    BOOL SetPixelFormat(HDC hdc, int format, const PIXELFORMATDESCRIPTOR*) {
+        EnsureInitialized();
+        MGLOG_D("wglSetPixelFormat(hdc=%p, format=%d)", hdc, format);
+        if (format < 1 || format > kPixelFormatCount) {
+            SetLastError(ERROR_INVALID_PARAMETER);
+            return FALSE;
+        }
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        WindowPixelFormats()[hwnd] = format;
+        return TRUE;
+    }
+
+    BOOL SwapBuffers(HDC hdc) {
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto& surfaces = WindowSurfaces();
+        auto it = surfaces.find(hwnd);
+        if (it == surfaces.end()) {
+            MGLOG_W("wglSwapBuffers: no surface for HWND %p", hwnd);
+            return FALSE;
+        }
+        SyncSurfaceSize(hwnd, it->second);
+        return EGLImpl::SwapBuffers(it->second.Display, it->second.Surface) == EGL_TRUE ? TRUE : FALSE;
+    }
+
+    HGLRC CreateContext(HDC hdc) {
+        EnsureInitialized();
+        MGLOG_I("wglCreateContext(hdc=%p)", hdc);
+        // A legacy WGL context is a compatibility-profile context; MobileGL keys
+        // its relaxed-semantics mode off the explicit compatibility bit.
+        const EGLint attribs[] = {
+            EGL_CONTEXT_MAJOR_VERSION, 3,
+            EGL_CONTEXT_MINOR_VERSION, 3,
+            EGL_CONTEXT_OPENGL_PROFILE_MASK, EGL_CONTEXT_OPENGL_COMPATIBILITY_PROFILE_BIT,
+            EGL_NONE,
+        };
+        return CreateContextFromEGLAttribs(hdc, nullptr, attribs);
+    }
+
+    BOOL DeleteContext(HGLRC hglrc) {
+        EnsureInitialized();
+        MGLOG_I("wglDeleteContext(%p)", hglrc);
+        if (t_current.Context == hglrc) {
+            MakeCurrent(nullptr, nullptr);
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(hglrc);
+        if (!object) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        if (object->Context != EGL_NO_CONTEXT) {
+            EGLImpl::DestroyContext(object->Display, object->Context);
+        }
+        Contexts().erase(hglrc);
+        return TRUE;
+    }
+
+    BOOL MakeCurrent(HDC hdc, HGLRC hglrc) {
+        EnsureInitialized();
+        MGLOG_D("wglMakeCurrent(hdc=%p, hglrc=%p)", hdc, hglrc);
+        if (!hglrc) {
+            if (!t_current.Context) {
+                t_current = {};
+                return TRUE;
+            }
+            const EGLBoolean released =
+                EGLImpl::MakeCurrent(EGL_NO_DISPLAY, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
+            t_current = {};
+            return released == EGL_TRUE ? TRUE : FALSE;
+        }
+
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(hglrc);
+        if (!object) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+
+        WindowSurface* surface = EnsureWindowSurface(hwnd, *object);
+        if (!surface) {
+            return FALSE;
+        }
+
+        if (!EGLImpl::MakeCurrent(object->Display, surface->Surface, surface->Surface, object->Context)) {
+            MGLOG_E("wglMakeCurrent: eglMakeCurrent failed (hdc=%p, hglrc=%p)", hdc, hglrc);
+            return FALSE;
+        }
+        t_current = {hdc, hglrc};
+        return TRUE;
+    }
+
+    HGLRC GetCurrentContext() {
+        return t_current.Context;
+    }
+
+    HDC GetCurrentDC() {
+        return t_current.DC;
+    }
+
+    BOOL ShareLists(HGLRC hglrcShare, HGLRC hglrcDest) {
+        // All MobileGL contexts alias one global GL object namespace, so every
+        // pair of contexts already shares.
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        if (!TryGetContext(hglrcShare) || !TryGetContext(hglrcDest)) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        return TRUE;
+    }
+
+    PROC GetProcAddress(const char* name) {
+        EnsureInitialized();
+        if (!name) {
+            return nullptr;
+        }
+        if (name[0] == 'w' && name[1] == 'g' && name[2] == 'l') {
+            for (const auto& entry : kWGLExtensionProcs) {
+                if (std::strcmp(entry.Name, name) == 0) {
+                    return entry.Proc;
+                }
+            }
+            MGLOG_D("wglGetProcAddress: unknown wgl entry point %s", name);
+            return nullptr;
+        }
+        return reinterpret_cast<PROC>(MG_Impl::GetProcAddress(name));
+    }
+} // namespace MobileGL::MG_Impl::WGLImpl
+
+#endif // _WIN32

MobileGL/MG_Impl/WGLImpl/WGLImpl.h

@@ -0,0 +1,34 @@
+// MobileGL - MobileGL/MG_Impl/WGLImpl/WGLImpl.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+
+#if defined(_WIN32)
+
+namespace MobileGL::MG_Impl::WGLImpl {
+    // Classic opengl32.dll surface. gdi32's ChoosePixelFormat/SetPixelFormat/
+    // DescribePixelFormat/GetPixelFormat/SwapBuffers forward into the loaded
+    // opengl32.dll's wgl* exports, so these back both call paths.
+    int ChoosePixelFormat(HDC hdc, const PIXELFORMATDESCRIPTOR* pfd);
+    int DescribePixelFormat(HDC hdc, int format, UINT size, PIXELFORMATDESCRIPTOR* pfd);
+    int GetPixelFormat(HDC hdc);
+    BOOL SetPixelFormat(HDC hdc, int format, const PIXELFORMATDESCRIPTOR* pfd);
+    BOOL SwapBuffers(HDC hdc);
+
+    HGLRC CreateContext(HDC hdc);
+    BOOL DeleteContext(HGLRC hglrc);
+    BOOL MakeCurrent(HDC hdc, HGLRC hglrc);
+    HGLRC GetCurrentContext();
+    HDC GetCurrentDC();
+    BOOL ShareLists(HGLRC hglrcShare, HGLRC hglrcDest);
+
+    PROC GetProcAddress(const char* name);
+} // namespace MobileGL::MG_Impl::WGLImpl
+
+#endif // _WIN32

MobileGL/MG_State/EGLState/Core.cpp

@@ -368,6 +368,26 @@ namespace MobileGL {
                 return true;
             }
 
+            Bool EGLContext::HasAnyInitializedDisplay() const {
+                const std::lock_guard<std::recursive_mutex> lock(m_mutex);
+                for (const auto& [handle, displayObject] : m_displays) {
+                    if (displayObject.Initialized) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
+            Bool EGLContext::HasAnyCurrentContext() const {
+                const std::lock_guard<std::recursive_mutex> lock(m_mutex);
+                for (const auto& [threadId, current] : m_threadCurrents) {
+                    if (current.Context != nullptr) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
             Bool EGLContext::ChooseConfig(EGLDisplayHandle display, const EGLint* attribList, EGLConfigHandle* configs,
                                           EGLint configSize, EGLint* numConfig) {
                 const std::lock_guard<std::recursive_mutex> lock(m_mutex);
@@ -1415,6 +1435,7 @@ namespace MobileGL {
             }
         } // namespace EGLState
 
-        UniquePtr<EGLState::EGLContext> pEGLContext;
+        // Leak-at-exit storage; see GlobalObjects.cpp.
+        UniquePtr<EGLState::EGLContext>& pEGLContext = *new UniquePtr<EGLState::EGLContext>();
     } // namespace MG_State
 } // namespace MobileGL

MobileGL/MG_State/EGLState/Core.h

@@ -39,6 +39,10 @@ namespace MobileGL {
                 Bool IsDisplayInitialized(EGLDisplayHandle display) const;
                 Bool InitializeDisplay(EGLDisplayHandle display, EGLint* major, EGLint* minor);
                 Bool TerminateDisplay(EGLDisplayHandle display);
+                // Whole-library idle checks used by EGLImpl::Terminate to decide
+                // when the last eglTerminate may tear MobileGL down entirely.
+                Bool HasAnyInitializedDisplay() const;
+                Bool HasAnyCurrentContext() const;
 
                 // Config
                 Bool ChooseConfig(EGLDisplayHandle display, const EGLint* attribList, EGLConfigHandle* configs,
@@ -262,6 +266,6 @@ namespace MobileGL {
             };
         } // namespace EGLState
 
-        extern UniquePtr<EGLState::EGLContext> pEGLContext;
+        extern UniquePtr<EGLState::EGLContext>& pEGLContext;
     } // namespace MG_State
 } // namespace MobileGL

MobileGL/MG_State/GLState/Core.cpp

@@ -721,5 +721,6 @@ namespace MobileGL::MG_State {
         }
     } // namespace GLState
 
-    UniquePtr<GLState::GLContext> pGLContext;
+    // Leak-at-exit storage; see GlobalObjects.cpp.
+    UniquePtr<GLState::GLContext>& pGLContext = *new UniquePtr<GLState::GLContext>();
 } // namespace MobileGL::MG_State

MobileGL/MG_State/GLState/Core.h

@@ -252,7 +252,7 @@ namespace MobileGL {
             };
         } // namespace GLState
 
-        extern UniquePtr<GLState::GLContext> pGLContext;
+        extern UniquePtr<GLState::GLContext>& pGLContext;
 
         // True when relaxed GL semantics apply. Strict core rules are enforced only when the
         // current EGL context explicitly requested a core profile (core bit in

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

@@ -334,16 +334,32 @@ namespace MobileGL::MG_State::GLState {
         // draw. The memo is keyed by (backendStateVersion, flags); ResetLinkArtifacts and
         // the binding setters below invalidate it by bumping m_backendStateVersion.
         Bool GetBackendHashMemo(Uint flags, Uint64& outHash) const {
-            if (m_backendHashMemoVersion != m_backendStateVersion || m_backendHashMemoFlags != flags) {
-                return false;
-            }
-            outHash = m_backendHashMemo;
+            if (m_backendHashMemoVersion != m_backendStateVersion) return false;
+            for (const auto& slot : m_backendHashMemoSlots) {
+                if (slot.valid && slot.flags == flags) {
+                    outHash = slot.hash;
                     return true;
                 }
+            }
+            return false;
+        }
         void SetBackendHashMemo(Uint flags, Uint64 hash) const {
-            m_backendHashMemo = hash;
+            if (m_backendHashMemoVersion != m_backendStateVersion) {
+                for (auto& slot : m_backendHashMemoSlots) slot.valid = false;
                 m_backendHashMemoVersion = m_backendStateVersion;
-            m_backendHashMemoFlags = flags;
+                m_backendHashMemoNextSlot = 0;
+            }
+            for (auto& slot : m_backendHashMemoSlots) {
+                if (slot.valid && slot.flags == flags) {
+                    slot.hash = hash;
+                    return;
+                }
+            }
+            auto& slot = m_backendHashMemoSlots[m_backendHashMemoNextSlot];
+            slot.flags = flags;
+            slot.hash = hash;
+            slot.valid = true;
+            m_backendHashMemoNextSlot = (m_backendHashMemoNextSlot + 1) % kBackendHashMemoSlotCount;
         }
 
         void SetUniformSamplerOrImageUnitIndex(Uint location, Int unit) {
@@ -527,10 +543,19 @@ namespace MobileGL::MG_State::GLState {
         Uint32 m_backendStateVersion = 0;
 
         // Backend-owned content-hash memo (see GetBackendHashMemo): valid only while
-        // m_backendStateVersion and the compile flags match the recorded values.
-        mutable Uint64 m_backendHashMemo = 0;
+        // m_backendStateVersion matches. Several slots, not one: a backend may resolve the same
+        // program under more than one compile-flag set within a frame (surface rotation, and the
+        // explicit-LOD sampling variant), and a single slot would then miss on every lookup and
+        // re-hash the program's whole SPIR-V once per draw.
+        static constexpr SizeT kBackendHashMemoSlotCount = 4;
+        struct BackendHashMemoSlot {
+            Uint64 hash = 0;
+            Uint flags = 0;
+            Bool valid = false;
+        };
+        mutable Array<BackendHashMemoSlot, kBackendHashMemoSlotCount> m_backendHashMemoSlots{};
+        mutable SizeT m_backendHashMemoNextSlot = 0;
         mutable Uint32 m_backendHashMemoVersion = ~0u;
-        mutable Uint m_backendHashMemoFlags = 0;
         Uint32 m_uboContentVersion = 0;
         Uint32 m_linkVersion = 0;
     };

MobileGL/MG_State/GLState/TextureState/TextureObject.h

@@ -15,7 +15,11 @@
 #include <MG_Util/Math/VectorTypes.h>
 
 namespace MobileGL::MG_State::GLState {
-    class ITextureObject {
+    // Texture objects are always SharedPtr-owned (TextureState creates every instance via
+    // MakeShared, including the per-target default objects). enable_shared_from_this lets
+    // backends that only receive a reference (e.g. syncing a name-deleted texture kept
+    // alive by an FBO attachment) still register a weak liveness reference for GC.
+    class ITextureObject : public std::enable_shared_from_this<ITextureObject> {
     public:
         using TargetEnum = TextureTarget;
         virtual ~ITextureObject() = default;

MobileGL/MG_Test/SanityTest.cpp

@@ -33,6 +33,7 @@
 #include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
 #include <MG_Util/ShaderTranspiler/ShaderSourceProcessor.h>
 #include <MG_Util/Debug/Log.h>
+#include <FastSTL/UnorderedMap.h>
 
 namespace {
     class DynamicParameterBackend final : public MobileGL::MG_Backend::BackendObject {
@@ -1736,3 +1737,61 @@ TEST(DirectGLESStateGuards, DefaultFramebufferBindGoesThroughShadow) {
     FramebufferImpl::BindFramebufferId(GL_DRAW_FRAMEBUFFER, 7); // must reach the driver again
     EXPECT_EQ(mocks.log.Count("BindFramebuffer:"), 3u);
 }
+
+// FastSTL::unordered_map::erase(iterator) regression coverage. The open-addressing
+// iterator constructor snaps forward from a tombstoned slot to the successor, so
+// erase must NOT advance the rebuilt iterator again: the old double-advance skipped
+// one live element per erase, and erasing the element in the highest occupied
+// bucket pushed the returned index past bucket_count where it never compared equal
+// to end() again - erase-while-iterating sweeps (pipeline/program cache eviction)
+// then ran off the bucket array and fed garbage handles to vkDestroyPipeline
+// (device crash on first mass eviction during world load).
+TEST(FastSTLSanity, EraseWhileIteratingVisitsEveryElementExactlyOnce) {
+    FastSTL::unordered_map<MobileGL::Uint64, MobileGL::Uint64> map;
+    constexpr MobileGL::Uint64 kCount = 1000;
+    for (MobileGL::Uint64 key = 0; key < kCount; ++key) {
+        map.emplace(key * 0x9e3779b97f4a7c15ull, key);
+    }
+    ASSERT_EQ(map.size(), kCount);
+
+    MobileGL::SizeT visited = 0;
+    for (auto it = map.begin(); it != map.end();) {
+        it = map.erase(it);
+        ++visited;
+        ASSERT_LE(visited, kCount); // old code: runaway past end / skipped entries
+    }
+    EXPECT_EQ(visited, kCount);
+    EXPECT_EQ(map.size(), 0u);
+}
+
+TEST(FastSTLSanity, EraseReturnsTheSuccessorElement) {
+    FastSTL::unordered_map<MobileGL::Uint32, MobileGL::Uint32> map;
+    for (MobileGL::Uint32 key = 1; key <= 64; ++key) {
+        map.emplace(key, key);
+    }
+
+    // Erasing every other visited element must still visit all 64 exactly once:
+    // the iterator returned by erase names the very next element, not one past it.
+    MobileGL::SizeT visited = 0;
+    MobileGL::SizeT erased = 0;
+    for (auto it = map.begin(); it != map.end();) {
+        ++visited;
+        if ((visited & 1) != 0) {
+            it = map.erase(it);
+            ++erased;
+        } else {
+            ++it;
+        }
+        ASSERT_LE(visited, 64u);
+    }
+    EXPECT_EQ(visited, 64u);
+    EXPECT_EQ(map.size(), 64u - erased);
+}
+
+TEST(FastSTLSanity, ErasingTheOnlyElementReturnsEnd) {
+    FastSTL::unordered_map<MobileGL::Uint32, MobileGL::Uint32> map;
+    map.emplace(42u, 1u);
+    auto next = map.erase(map.begin());
+    EXPECT_EQ(next, map.end());
+    EXPECT_TRUE(map.empty());
+}

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.cpp

@@ -9,7 +9,12 @@
 #include "Loader.h"
 #include "MG_Util/Types.h"
 #include <Config.h>
-#if !defined(__WIN32) && !defined(_WIN32)
+#if defined(_WIN32)
+#ifndef WIN32_LEAN_AND_MEAN
+#define WIN32_LEAN_AND_MEAN 1
+#endif
+#include <windows.h>
+#else
 #include <dlfcn.h>
 #endif
 
@@ -60,7 +65,14 @@ namespace MobileGL::MG_Util::BackendLoader {
 #endif
 
     static void* OpenLib(const Vector<String>& names) {
-#if !defined(__WIN32) && !defined(_WIN32) && (!defined(__APPLE__) || defined(MOBILEGL_IOS))
+#if defined(_WIN32)
+        for (const auto& name : names) {
+            if (HMODULE lib = LoadLibraryA(name.c_str())) {
+                MGLOG_I("Loaded GL backend library: %s", name.c_str());
+                return reinterpret_cast<void*>(lib);
+            }
+        }
+#elif !defined(__APPLE__) || defined(MOBILEGL_IOS)
         static const String LibPathPrefixes[] = {
 #if defined(MOBILEGL_IOS)
             "@rpath/", "@executable_path/Frameworks/", "@loader_path/Frameworks/",
@@ -104,7 +116,9 @@ namespace MobileGL::MG_Util::BackendLoader {
     }
 
     inline void* ProcAddress(void* lib, const char* name) {
-#if !defined(__WIN32) && !defined(_WIN32) && (!defined(__APPLE__) || defined(MOBILEGL_IOS))
+#if defined(_WIN32)
+        return reinterpret_cast<void*>(::GetProcAddress(reinterpret_cast<HMODULE>(lib), name));
+#elif !defined(__APPLE__) || defined(MOBILEGL_IOS)
         return dlsym(lib, name);
 #else
         return nullptr;
@@ -520,6 +534,15 @@ namespace MobileGL::MG_Util::BackendLoader {
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS) && defined(__ANDROID__)
         void* angleGlesLib = nullptr;
 #endif
+#if defined(_WIN32)
+        // ANGLE is the GLES provider on Windows regardless of UseAngle(). Preload
+        // libGLESv2.dll so libEGL.dll resolves its dependency from the same directory.
+        if (!OpenLib({"libGLESv2.dll"})) {
+            MGLOG_E("Failed to open ANGLE libGLESv2.dll");
+            return;
+        }
+        eglLib = OpenLib({"libEGL.dll"});
+#else
         if (UseAngle()) {
             void* glesLib = OpenLib({"libGLESv2_angle.so"});
             if (!glesLib) {
@@ -541,6 +564,7 @@ namespace MobileGL::MG_Util::BackendLoader {
             eglLib = OpenLib({"libEGL.so"});
 #endif
         }
+#endif // !_WIN32
 
         if (!eglLib) {
             MGLOG_E("Failed to open EGL library");

MobileGL/MG_Util/Converters/MGToVk/TextureEnumConverter.cpp

@@ -133,9 +133,11 @@ namespace MobileGL {
             case TextureInternalFormat::RGBA8Snorm:
                 return VK_FORMAT_R8G8B8A8_SNORM;
             case TextureInternalFormat::RGB10A2:
-                return VK_FORMAT_A2R10G10B10_UNORM_PACK32;
+                // GL_UNSIGNED_INT_2_10_10_10_REV puts R in bits 0-9, which is Vulkan's
+                // A2B10G10R10 layout - A2R10G10B10 silently swaps R and B on upload.
+                return VK_FORMAT_A2B10G10R10_UNORM_PACK32;
             case TextureInternalFormat::RGB10A2UI:
-                return VK_FORMAT_A2R10G10B10_UINT_PACK32;
+                return VK_FORMAT_A2B10G10R10_UINT_PACK32;
             case TextureInternalFormat::RGBA16:
                 return VK_FORMAT_R16G16B16A16_UNORM;
             case TextureInternalFormat::RGBA16Snorm:

tools/cts/scripts/run_cts.py

@@ -87,7 +87,7 @@ def main():
     ap.add_argument("--device-dir", default="/data/local/tmp/mgcts")
     ap.add_argument("--surface", default="fbo", help="--deqp-surface-type value")
     ap.add_argument("--max-rounds", type=int, default=4000)
-    ap.add_argument("--max-empty-streak", type=int, default=8,
+    ap.add_argument("--max-empty-streak", type=int, default=64,
                     help="abort after this many consecutive chunks that produce no log at all")
     ap.add_argument("--min-mem-kb", type=int, default=400000,
                     help="pause when the device drops below this much available memory")

tools/piglit-android/README.md

@@ -0,0 +1,157 @@
+# piglit on Android against MobileGL
+
+Run [piglit](https://gitlab.freedesktop.org/mesa/piglit) desktop-GL tests on a
+connected Android device with **MobileGL as the OpenGL implementation**, for
+both backends:
+
+- `DirectGLES` — MobileGL over the system GLES driver (or ANGLE with
+  `--use-angle`)
+- `DirectVulkan` — MobileGL over the system Vulkan driver
+
+No APK and no on-device python: piglit test binaries run as the adb shell user
+from `/data/local/tmp`, create contexts through waffle's `surfaceless_egl`
+platform, and waffle is pointed at `libMobileGL.so` (which exports the full
+`egl*`/`gl*` API under real names).
+
+## How it fits together
+
+```
+piglit test binary (aarch64, bionic)
+  └─ waffle surfaceless_egl (patched)
+       ├─ WAFFLE_EGL_LIBRARY=libMobileGL.so   → dlopen MobileGL as the EGL impl
+       ├─ WAFFLE_GL_LIBRARY=libMobileGL.so    → waffle_dl_sym resolves gl* here
+       ├─ WAFFLE_FORCE_GL_CONTEXT_VERSION=33core
+       │    upgrades low compat context requests to GL 3.3 core (never
+       │    downgrades) so piglit's supports_gl_compat_version=10 tests run
+       └─ WAFFLE_ANDROID_WINDOW=imagereader   (DirectVulkan only)
+            windows are AImageReader ANativeWindows instead of EGL pbuffers,
+            because Android ICDs lack VK_EXT_headless_surface which the
+            MobileGL pbuffer path needs
+  └─ libMobileGL.so
+       ├─ DirectGLES: dlopens the real system libEGL.so internally
+       └─ DirectVulkan: links libvulkan.so
+```
+
+Key rule: **never name MobileGL `libEGL.so`** anywhere on `LD_LIBRARY_PATH` —
+the DirectGLES backend loads the system driver with a bare-soname
+`dlopen("libEGL.so")` and would recursively pick itself up.
+
+## One-time setup (host: macOS/Linux with the Android NDK)
+
+```sh
+WORK=path/to/workdir && cd $WORK
+git clone --depth 1 https://gitlab.freedesktop.org/mesa/piglit.git
+git clone --depth 1 https://gitlab.freedesktop.org/mesa/waffle.git
+git -C waffle apply $MOBILEGL/tools/piglit-android/patches/waffle-mobilegl-android.patch
+git -C piglit apply $MOBILEGL/tools/piglit-android/patches/piglit-mobilegl-android.patch
+python3 -m venv venv && ./venv/bin/pip install mako numpy packaging
+```
+
+The piglit patch matters beyond build fixes: upstream's
+`piglit_dispatch_default_init` runs while the waffle framework is still being
+constructed (`gl_fw` is NULL), so the waffle resolvers were never installed and
+GL functions bound through the **system** libEGL's `eglGetProcAddress` — every
+test silently ran on the raw GLES driver instead of MobileGL.
+
+Build MobileGL for Android:
+
+```sh
+cmake -S $MOBILEGL -B $MOBILEGL/build-android-arm64 -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-26 \
+  -DCMAKE_BUILD_TYPE=RelWithDebInfo
+cmake --build $MOBILEGL/build-android-arm64 --target MobileGL -j
+$NDK/toolchains/llvm/prebuilt/*/bin/llvm-strip --strip-unneeded \
+  -o $WORK/libMobileGL-stripped.so $MOBILEGL/build-android-arm64/libMobileGL.so
+```
+
+Cross-build waffle (meson; a cross file and a stub `egl.pc` pointing at
+MobileGL's bundled EGL 1.5 headers are needed — see `cross-example/`):
+
+```sh
+cd $WORK/waffle
+meson setup build-android --cross-file $WORK/cross/android-arm64.ini \
+  -Dbuildtype=release -Dsurfaceless_egl=enabled \
+  -Dglx=disabled -Dx11_egl=disabled -Dgbm=disabled -Dwayland=disabled \
+  -Dbuild-tests=false -Dbuild-examples=false -Dprefix=$WORK/prefix
+ninja -C build-android && meson install -C build-android
+```
+
+Cross-build piglit (needs `PKG_CONFIG_LIBDIR` with the installed `waffle-1.pc`
+plus the stub `egl.pc`):
+
+```sh
+cd $WORK/piglit && export PKG_CONFIG_LIBDIR=$WORK/prefix/lib/pkgconfig:$WORK/cross/pkgconfig
+cmake -S . -B build-android -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-26 \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DPIGLIT_USE_WAFFLE=ON -DPIGLIT_BUILD_GL_TESTS=ON \
+  -DPIGLIT_BUILD_GLES1_TESTS=OFF -DPIGLIT_BUILD_GLES2_TESTS=OFF \
+  -DPIGLIT_BUILD_GLES3_TESTS=OFF -DPIGLIT_BUILD_EGL_TESTS=OFF \
+  -DPIGLIT_BUILD_GLX_TESTS=OFF -DPIGLIT_BUILD_WGL_TESTS=OFF \
+  -DPIGLIT_BUILD_CL_TESTS=OFF -DPIGLIT_BUILD_VK_TESTS=OFF \
+  -DPIGLIT_BUILD_DMA_BUF_TESTS=OFF -DPIGLIT_USE_GBM=OFF \
+  -DPIGLIT_USE_WAYLAND=OFF -DPIGLIT_USE_X11=OFF \
+  -DPYTHON_EXECUTABLE=$WORK/venv/bin/python \
+  -DOPENGL_INCLUDE_DIR=$MOBILEGL/include \
+  -DOPENGL_gl_LIBRARY=$SYSROOT/usr/lib/aarch64-linux-android/26/libEGL.so \
+  -DGLEXT_INCLUDE_DIR=$MOBILEGL/include
+ninja -C build-android
+```
+
+## Selecting tests
+
+Enumerate on the host with piglit's own profiles (no device needed):
+
+```sh
+cd $WORK/piglit
+for prof in opengl shader glslparser; do
+  PIGLIT_BUILD_DIR=$PWD/build-android ./venv/bin/python ./piglit print-cmd \
+    -t "spec@!opengl 3[.]" -t "spec@glsl-3[.]30" $prof
+done > /tmp/gl33.list
+```
+
+Group names use `@` separators (`spec@!opengl 3.3@minmax`). The version groups
+(`spec@!opengl 1.x…3.3`), GLSL groups (`spec@glsl-1.10…3.30`) plus the ARB
+extension groups folded into GL 3.1–3.3 core give a comprehensive "GL 3.3
+core" suite (~15k tests).
+
+## Running
+
+```sh
+python3 $MOBILEGL/tools/piglit-android/run_piglit_android.py \
+  --piglit-root $WORK/piglit --list /tmp/gl33.list \
+  --backend DirectGLES \
+  --mobilegl-lib $WORK/libMobileGL-stripped.so \
+  --waffle-lib $WORK/waffle/build-android/src/waffle/libwaffle-1.so \
+  --out results-gles
+# then the same with --backend DirectVulkan --out results-vk
+```
+
+The runner pushes binaries/libs/data (incremental; `--repush` forces), executes
+tests serially in chunked on-device shell scripts under `timeout`, parses the
+`PIGLIT: {...}` result lines, and writes `results.json` + `summary.txt` +
+`raw.log`. Exit-code semantics: parsed result wins; nonzero exit without a
+result line = `crash`; toybox timeout exits = `timeout`.
+
+Quick sanity check for the whole stack (waffle build also produces `wflinfo`):
+
+```sh
+adb shell 'cd /data/local/tmp/piglit-mgl && env LD_LIBRARY_PATH=$PWD/lib \
+  WAFFLE_EGL_LIBRARY=libMobileGL.so WAFFLE_GL_LIBRARY=libMobileGL.so \
+  MOBILEGL_BACKEND_TYPE=DirectVulkan WAFFLE_ANDROID_WINDOW=imagereader \
+  ./wflinfo --platform surfaceless_egl --api gl --version 3.3 --profile core'
+```
+
+Expect `OpenGL version string: 3.3.0 MobileGL …, Direct (Vulkan) Backend`.
+
+## Known caveats
+
+- MSAA winsys configs never match (MobileGL exposes two RGBA8888 configs,
+  samples=0); MSAA FBO tests are unaffected.
+- Tests that genuinely require compatibility-profile features will fail on the
+  forced 3.3 core context; that is honest for a core-only implementation.
+- `eglTerminate` at test exit now tears MobileGL down deterministically (see
+  the EGL-lifecycle refactor); a device-side `mobilegl.log` is written per run
+  directory for debugging.

tools/piglit-android/compare_results.py

@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+# MobileGL - tools/piglit-android/compare_results.py
+# 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
+"""Compare two run_piglit_android.py results.json files (e.g. DirectGLES vs
+DirectVulkan) and write a markdown report of totals plus categorized diffs."""
+
+import argparse
+import json
+from collections import Counter
+from pathlib import Path
+
+BAD = ('crash', 'timeout', 'fail', 'missing', 'notrun', 'warn')
+
+
+def load(path):
+    d = json.loads(Path(path).read_text())
+    return d
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument('a', help='first results.json')
+    ap.add_argument('b', help='second results.json')
+    ap.add_argument('-o', '--out', help='markdown output path')
+    args = ap.parse_args()
+
+    da, db = load(args.a), load(args.b)
+    na, nb = da['backend'], db['backend']
+    ta, tb = da['tests'], db['tests']
+    names = sorted(set(ta) | set(tb))
+
+    lines = [f'# piglit: {na} vs {nb}', '']
+    lines.append(f'| | {na} | {nb} |')
+    lines.append('|---|---|---|')
+    ca = Counter(o["status"] for o in ta.values())
+    cb = Counter(o["status"] for o in tb.values())
+    for k in sorted(set(ca) | set(cb)):
+        lines.append(f'| {k} | {ca.get(k, 0)} | {cb.get(k, 0)} |')
+    lines.append(f'| total | {len(ta)} | {len(tb)} |')
+    lines.append(f'| elapsed | {da.get("elapsed_sec")}s | {db.get("elapsed_sec")}s |')
+    lines.append('')
+
+    def bucket(pred, title):
+        rows = [n for n in names
+                if pred(ta.get(n, {}).get('status', 'absent'),
+                        tb.get(n, {}).get('status', 'absent'))]
+        if rows:
+            lines.append(f'## {title} ({len(rows)})')
+            lines.append('')
+            for n in rows:
+                sa = ta.get(n, {}).get('status', 'absent')
+                sb = tb.get(n, {}).get('status', 'absent')
+                lines.append(f'- `{n}` — {na}: {sa}, {nb}: {sb}')
+            lines.append('')
+
+    bucket(lambda a, b: a in BAD and b in BAD, 'Bad on both (likely frontend/state-tracker)')
+    bucket(lambda a, b: a in BAD and b == 'pass', f'Bad only on {na}')
+    bucket(lambda a, b: a == 'pass' and b in BAD, f'Bad only on {nb}')
+    bucket(lambda a, b: a == 'skip' and b == 'pass' or a == 'pass' and b == 'skip',
+           'Skip on one side only')
+
+    text = '\n'.join(lines) + '\n'
+    if args.out:
+        Path(args.out).write_text(text)
+        print(f'wrote {args.out}')
+    else:
+        print(text)
+
+
+if __name__ == '__main__':
+    main()

tools/piglit-android/cross-example/android-arm64.ini

@@ -0,0 +1,19 @@
+; meson cross file for waffle -> aarch64 Android
+; Replace NDK_TOOLCHAIN with e.g.
+;   $HOME/Library/Android/sdk/ndk/27.3.13750724/toolchains/llvm/prebuilt/darwin-x86_64
+; and PKGCONFIG_DIR with the directory holding the stub egl.pc.
+[binaries]
+c = 'NDK_TOOLCHAIN/bin/aarch64-linux-android26-clang'
+cpp = 'NDK_TOOLCHAIN/bin/aarch64-linux-android26-clang++'
+ar = 'NDK_TOOLCHAIN/bin/llvm-ar'
+strip = 'NDK_TOOLCHAIN/bin/llvm-strip'
+pkg-config = '/usr/bin/pkg-config'
+
+[host_machine]
+system = 'android'
+cpu_family = 'aarch64'
+cpu = 'aarch64'
+endian = 'little'
+
+[properties]
+pkg_config_libdir = 'PKGCONFIG_DIR'

tools/piglit-android/cross-example/egl.pc

@@ -0,0 +1,8 @@
+# Stub egl.pc for cross-building waffle/piglit against the NDK.
+# The NDK's own EGL headers are 1.4-era; point Cflags at MobileGL's bundled
+# EGL 1.5 headers (include/EGL) instead. -lEGL resolves to the NDK stub.
+Name: egl
+Description: EGL 1.5 headers (MobileGL bundled) + NDK libEGL stub
+Version: 1.5
+Libs: -lEGL
+Cflags: -IMOBILEGL_REPO/include

tools/piglit-android/patches/piglit-mobilegl-android.patch

@@ -0,0 +1,104 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 1a38c21..b0b6ce2 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -22,7 +22,7 @@ INCLUDE (FindPkgConfig)
+ 
+ set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+ 
+-if(${CMAKE_SYSTEM_NAME} MATCHES "Linux|FreeBSD")
++if(${CMAKE_SYSTEM_NAME} MATCHES "Linux|FreeBSD|Android")
+ 	set(DEFAULT_EGL ON)
+ 	set(DEFAULT_GLX ON)
+ 	set(DEFAULT_WGL OFF)
+@@ -499,13 +499,18 @@ if(GBM_FOUND)
+ 	endif(HAVE_LIBCACA)
+ endif(GBM_FOUND)
+ 
+-if(PIGLIT_BUILD_EGL_TESTS)
+-	pkg_check_modules(EGL REQUIRED egl)
++# EGL *support* (PIGLIT_HAS_EGL, e.g. the surfaceless_egl waffle platform) is
++# independent of building the EGL test binaries, which additionally need X11.
++pkg_check_modules(EGL egl)
++if(EGL_FOUND)
+ 	set(PIGLIT_HAS_EGL True)
+ 	add_definitions(-DPIGLIT_HAS_EGL)
+ 	include_directories(${EGL_INCLUDE_DIRS})
+ 	add_definitions (${EGL_CFLAGS_OTHER})
+ endif()
++if(PIGLIT_BUILD_EGL_TESTS AND NOT EGL_FOUND)
++	message(FATAL_ERROR "PIGLIT_BUILD_EGL_TESTS requires EGL")
++endif()
+ 
+ # Put all executables into the bin subdirectory
+ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${piglit_BINARY_DIR}/bin)
+diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
+index e3c964a..36e2109 100644
+--- a/tests/CMakeLists.txt
++++ b/tests/CMakeLists.txt
+@@ -35,9 +35,9 @@ add_subdirectory (llvmpipe)
+ add_subdirectory (perf)
+ add_subdirectory (wgl)
+ 
+-IF(EGL_FOUND)
++IF(PIGLIT_BUILD_EGL_TESTS)
+ 	add_subdirectory (egl)
+-ENDIF(EGL_FOUND)
++ENDIF(PIGLIT_BUILD_EGL_TESTS)
+ 
+ IF(PIGLIT_BUILD_CL_TESTS)
+ 	add_subdirectory (cl)
+diff --git a/tests/util/piglit-dispatch-init.c b/tests/util/piglit-dispatch-init.c
+index 3d9772d..e036eac 100644
+--- a/tests/util/piglit-dispatch-init.c
++++ b/tests/util/piglit-dispatch-init.c
+@@ -322,14 +322,19 @@ piglit_dispatch_default_init(piglit_dispatch_api api)
+ 		break;
+ 	}
+ 
+-	if (gl_fw) {
+-		piglit_dispatch_init(api,
+-				     get_wfl_core_proc,
+-				     get_wfl_ext_proc,
+-				     default_unsupported,
+-				     default_get_proc_address_failure);
+-	} else
+-#endif
++	/* MobileGL piglit harness: always use the waffle resolvers in waffle
++	 * builds. This function runs while the waffle framework is still being
++	 * constructed, so gl_fw is NULL here and the gl_fw branch never fired;
++	 * the fallback resolvers bind GL functions via the SYSTEM libEGL's
++	 * eglGetProcAddress (a DT_NEEDED symbol), which on Android is the raw
++	 * GLES driver rather than the waffle-selected EGL implementation.
++	 */
++	piglit_dispatch_init(api,
++			     get_wfl_core_proc,
++			     get_wfl_ext_proc,
++			     default_unsupported,
++			     default_get_proc_address_failure);
++#else
+ 	{
+ 
+ 		piglit_dispatch_init(api,
+@@ -338,6 +343,7 @@ piglit_dispatch_default_init(piglit_dispatch_api api)
+ 				     default_unsupported,
+ 				     default_get_proc_address_failure);
+ 	}
++#endif
+ 
+ 	already_initialized = true;
+ }
+diff --git a/tests/util/piglit-util-gl.c b/tests/util/piglit-util-gl.c
+index 16eb425..7f2796c 100644
+--- a/tests/util/piglit-util-gl.c
++++ b/tests/util/piglit-util-gl.c
+@@ -47,6 +47,9 @@ bool piglit_is_core_profile;
+ bool piglit_is_gles(void)
+ {
+ 	const char *version_string = (const char *) glGetString(GL_VERSION);
++	if (getenv("PIGLIT_DEBUG_VERSION_STRING"))
++		printf("piglit: debug: GL_VERSION = \"%s\"\n",
++		       version_string ? version_string : "(null)");
+ 	return strncmp("OpenGL ES", version_string, 9) == 0;
+ }
+ 

tools/piglit-android/patches/waffle-mobilegl-android.patch

@@ -0,0 +1,445 @@
+diff --git a/meson.build b/meson.build
+index 7060fce..dcd45ed 100644
+--- a/meson.build
++++ b/meson.build
+@@ -78,7 +78,11 @@ else
+   dep_egl = dependency('egl', required : get_option('surfaceless_egl'))
+   build_surfaceless = dep_egl.found()
+ 
+-  dep_egl = dependency('egl', required : get_option('wayland'))
++  # Don't clobber a dep_egl found for surfaceless_egl when wayland is
++  # disabled (a disabled-feature dependency() returns not-found).
++  if not dep_egl.found()
++    dep_egl = dependency('egl', required : get_option('wayland'))
++  endif
+   dep_wayland_client = dependency(
+     'wayland-client', version : '>= 1.10', required : get_option('wayland'),
+   )
+diff --git a/src/waffle/egl/wegl_context.c b/src/waffle/egl/wegl_context.c
+index 26362a9..a26c27f 100644
+--- a/src/waffle/egl/wegl_context.c
++++ b/src/waffle/egl/wegl_context.c
+@@ -2,6 +2,8 @@
+ // SPDX-License-Identifier: BSD-2-Clause
+ 
+ #include <assert.h>
++#include <stdlib.h>
++#include <string.h>
+ 
+ #include <EGL/egl.h>
+ #include <EGL/eglext.h>
+@@ -19,6 +21,25 @@
+ #define EGL_CONTEXT_OPENGL_ROBUST_ACCESS  0x31B2
+ #endif
+ 
++// MobileGL piglit harness: WAFFLE_FORCE_GL_CONTEXT_VERSION="33core" upgrades
++// every desktop-GL context request to at least that version/profile (never
++// downgrades). MobileGL implements exactly GL 3.3 core; piglit tests that ask
++// for a low compat context (supports_gl_compat_version=10) would otherwise
++// get a context whose version string reflects the raw backend, and skip.
++static bool
++force_gl_context_version(EGLint *major, EGLint *minor, bool *core)
++{
++    const char *env = getenv("WAFFLE_FORCE_GL_CONTEXT_VERSION");
++    if (!env || strlen(env) < 2)
++        return false;
++    if (env[0] < '1' || env[0] > '9' || env[1] < '0' || env[1] > '9')
++        return false;
++    *major = env[0] - '0';
++    *minor = env[1] - '0';
++    *core = strstr(env, "core") != NULL;
++    return true;
++}
++
+ static bool
+ bind_api(struct wegl_platform *plat, int32_t waffle_context_api)
+ {
+@@ -61,13 +82,30 @@ create_real_context(struct wegl_config *config,
+         context_flags |= EGL_CONTEXT_OPENGL_DEBUG_BIT_KHR;
+     }
+ 
++    // Effective version/profile for desktop GL, possibly upgraded by
++    // WAFFLE_FORCE_GL_CONTEXT_VERSION (never downgraded).
++    EGLint eff_major = attrs->context_major_version;
++    EGLint eff_minor = attrs->context_minor_version;
++    int32_t eff_profile = attrs->context_profile;
++    if (waffle_context_api == WAFFLE_CONTEXT_OPENGL) {
++        EGLint f_major, f_minor;
++        bool f_core;
++        if (force_gl_context_version(&f_major, &f_minor, &f_core) &&
++            10 * f_major + f_minor > 10 * eff_major + eff_minor) {
++            eff_major = f_major;
++            eff_minor = f_minor;
++            if (f_core)
++                eff_profile = WAFFLE_CONTEXT_CORE_PROFILE;
++        }
++    }
++
+     switch (waffle_context_api) {
+         case WAFFLE_CONTEXT_OPENGL:
+             if (dpy->KHR_create_context) {
+                 attrib_list[i++] = EGL_CONTEXT_MAJOR_VERSION_KHR;
+-                attrib_list[i++] = attrs->context_major_version;
++                attrib_list[i++] = eff_major;
+                 attrib_list[i++] = EGL_CONTEXT_MINOR_VERSION_KHR;
+-                attrib_list[i++] = attrs->context_minor_version;
++                attrib_list[i++] = eff_minor;
+             }
+             else {
+                 assert(attrs->context_major_version == 1);
+@@ -92,9 +130,9 @@ create_real_context(struct wegl_config *config,
+                 }
+             }
+ 
+-            if (wcore_config_attrs_version_ge(attrs, 32))  {
++            if (10 * eff_major + eff_minor >= 32)  {
+                 assert(dpy->KHR_create_context);
+-                switch (attrs->context_profile) {
++                switch (eff_profile) {
+                     case WAFFLE_CONTEXT_CORE_PROFILE:
+                         attrib_list[i++] = EGL_CONTEXT_OPENGL_PROFILE_MASK_KHR;
+                         attrib_list[i++] = EGL_CONTEXT_OPENGL_CORE_PROFILE_BIT_KHR;
+diff --git a/src/waffle/egl/wegl_platform.c b/src/waffle/egl/wegl_platform.c
+index 6fe72ba..c3d72b1 100644
+--- a/src/waffle/egl/wegl_platform.c
++++ b/src/waffle/egl/wegl_platform.c
+@@ -13,11 +13,24 @@
+ #include "wegl_platform.h"
+ 
+ #ifdef WAFFLE_HAS_ANDROID
+-static const char *libEGL_filename = "libEGL.so";
++static const char *libEGL_default_filename = "libEGL.so";
+ #else
+-static const char *libEGL_filename = "libEGL.so.1";
++static const char *libEGL_default_filename = "libEGL.so.1";
+ #endif
+ 
++// MobileGL piglit harness: allow overriding which library provides the EGL
++// entry points (e.g. WAFFLE_EGL_LIBRARY=libMobileGL.so). MobileGL exports the
++// full EGL API under the real egl* names, so waffle can drive it directly
++// while the system libEGL.so stays untouched for MobileGL's own backend use.
++static const char *
++egl_library_name(void)
++{
++    const char *env = getenv("WAFFLE_EGL_LIBRARY");
++    if (env && env[0])
++        return env;
++    return libEGL_default_filename;
++}
++
+ static bool
+ supports_egl_khr_display(const struct wegl_platform *plat,
+                          const char *client_extensions)
+@@ -120,7 +133,7 @@ wegl_platform_teardown(struct wegl_platform *self)
+             ok = false;
+             wcore_errorf(WAFFLE_ERROR_UNKNOWN,
+                          "dlclose(\"%s\") failed: %s",
+-                         libEGL_filename, dlerror());
++                         egl_library_name(), dlerror());
+         }
+         self->egl.handle = NULL;
+     }
+@@ -132,12 +145,19 @@ wegl_platform_teardown(struct wegl_platform *self)
+ bool
+ wegl_platform_init(struct wegl_platform *self, EGLenum egl_platform)
+ {
+-    static const char *const dso_names[] = {
++    static const char *dso_names[] = {
+         "libGL.so.1",
+         "libGLESv1_CM.so.1",
+         "libGLESv2.so.2",
+     };
+ 
++    // MobileGL piglit harness: let waffle_dl_sym(WAFFLE_DL_OPENGL, ...)
++    // resolve desktop-GL symbols from an alternate library (MobileGL exports
++    // real gl* names). Android has no libGL.so.1.
++    const char *gl_override = getenv("WAFFLE_GL_LIBRARY");
++    if (gl_override && gl_override[0])
++        dso_names[0] = gl_override;
++
+     wcore_platform_init(&self->wcore);
+     posix_platform_init(&self->wcore, dso_names);
+ 
+@@ -146,11 +166,11 @@ wegl_platform_init(struct wegl_platform *self, EGLenum egl_platform)
+     // Most Waffle platforms will call eglCreateWindowSurface.
+     self->egl_surface_type_mask = EGL_WINDOW_BIT;
+ 
+-    self->egl.handle = dlopen(libEGL_filename, RTLD_LAZY | RTLD_LOCAL);
++    self->egl.handle = dlopen(egl_library_name(), RTLD_LAZY | RTLD_LOCAL);
+     if (!self->egl.handle) {
+         wcore_errorf(WAFFLE_ERROR_FATAL,
+                      "dlopen(\"%s\") failed: %s",
+-                     libEGL_filename, dlerror());
++                     egl_library_name(), dlerror());
+         return false;
+     }
+ 
+@@ -158,7 +178,7 @@ wegl_platform_init(struct wegl_platform *self, EGLenum egl_platform)
+     self->egl.function = dlsym(self->egl.handle, "egl" #function);             \
+     if (!self->egl.function) {                                                 \
+         wcore_errorf(WAFFLE_ERROR_FATAL, "dlsym(\"%s\", \"%s\") failed: %s",   \
+-                     libEGL_filename, "egl" #function, dlerror());             \
++                     egl_library_name(), "egl" #function, dlerror());          \
+         goto error;                                                            \
+     }
+ 
+diff --git a/src/waffle/meson.build b/src/waffle/meson.build
+index 2ae4c40..e336566 100644
+--- a/src/waffle/meson.build
++++ b/src/waffle/meson.build
+@@ -22,6 +22,16 @@ deps_for_waffle = [
+   idep_threads,
+ ]
+ 
++# MobileGL piglit harness: the Android imagereader window mode in
++# surfaceless_egl needs AImageReader (libmediandk) and ANativeWindow
++# (libandroid).
++if build_surfaceless and host_machine.system() == 'android'
++  deps_for_waffle += [
++    cc.find_library('mediandk'),
++    cc.find_library('android'),
++  ]
++endif
++
+ files_libwaffle = files(
+   'api/api_priv.c',
+   'api/waffle_attrib_list.c',
+diff --git a/src/waffle/surfaceless_egl/sl_platform.c b/src/waffle/surfaceless_egl/sl_platform.c
+index 30018c1..2c0024c 100644
+--- a/src/waffle/surfaceless_egl/sl_platform.c
++++ b/src/waffle/surfaceless_egl/sl_platform.c
+@@ -101,7 +101,7 @@ static const struct wcore_platform_vtbl sl_platform_vtbl = {
+         .destroy = sl_window_destroy,
+         .show = sl_window_show,
+         .resize = sl_window_resize,
+-        .swap_buffers = wegl_surface_swap_buffers,
++        .swap_buffers = sl_window_swap_buffers,
+         .get_native = sl_window_get_native,
+     },
+ };
+diff --git a/src/waffle/surfaceless_egl/sl_platform.h b/src/waffle/surfaceless_egl/sl_platform.h
+index a97fabe..a1c91aa 100644
+--- a/src/waffle/surfaceless_egl/sl_platform.h
++++ b/src/waffle/surfaceless_egl/sl_platform.h
+@@ -5,7 +5,6 @@
+ 
+ #include <stdbool.h>
+ #include <stdlib.h>
+-#include <gbm.h>
+ 
+ #undef linux
+ 
+diff --git a/src/waffle/surfaceless_egl/sl_window.c b/src/waffle/surfaceless_egl/sl_window.c
+index 0c54613..d18f8b3 100644
+--- a/src/waffle/surfaceless_egl/sl_window.c
++++ b/src/waffle/surfaceless_egl/sl_window.c
+@@ -9,18 +9,108 @@
+ #include "wcore_tinfo.h"
+ 
+ #include "wegl_config.h"
++#include "wegl_display.h"
++#include "wegl_platform.h"
+ #include "wegl_util.h"
+ 
+ #include "sl_display.h"
+ #include "sl_platform.h"
+ #include "sl_window.h"
+ 
++#ifdef __ANDROID__
++#include <android/hardware_buffer.h>
++#include <android/native_window.h>
++#include <media/NdkImageReader.h>
++
++// MobileGL piglit harness: back "windows" with an AImageReader-provided
++// ANativeWindow instead of an EGL pbuffer. MobileGL's DirectVulkan backend
++// needs a real ANativeWindow on Android because the device ICD lacks
++// VK_EXT_headless_surface (which its pbuffer path requires).
++static bool
++sl_android_window_mode(void)
++{
++    const char *env = getenv("WAFFLE_ANDROID_WINDOW");
++    return env && strcmp(env, "imagereader") == 0;
++}
++
++static bool
++sl_imagereader_init(struct sl_window *self,
++                    struct wcore_config *wc_config,
++                    int32_t width, int32_t height)
++{
++    struct wegl_config *config = wegl_config(wc_config);
++    struct wegl_display *dpy = wegl_display(wc_config->display);
++    struct wegl_platform *plat = wegl_platform(dpy->wcore.platform);
++    AImageReader *reader = NULL;
++    ANativeWindow *window = NULL;
++
++    media_status_t st = AImageReader_newWithUsage(
++        width, height, AIMAGE_FORMAT_RGBA_8888,
++        AHARDWAREBUFFER_USAGE_GPU_COLOR_OUTPUT |
++            AHARDWAREBUFFER_USAGE_GPU_SAMPLED_IMAGE,
++        8, &reader);
++    if (st != AMEDIA_OK || !reader) {
++        wcore_errorf(WAFFLE_ERROR_UNKNOWN,
++                     "AImageReader_newWithUsage(%dx%d) failed: %d",
++                     width, height, (int)st);
++        return false;
++    }
++
++    st = AImageReader_getWindow(reader, &window);
++    if (st != AMEDIA_OK || !window) {
++        wcore_errorf(WAFFLE_ERROR_UNKNOWN,
++                     "AImageReader_getWindow failed: %d", (int)st);
++        AImageReader_delete(reader);
++        return false;
++    }
++
++    // MobileGL reads EGL_WIDTH/EGL_HEIGHT from the window-surface attrib
++    // list (non-standard); unknown attribs are ignored.
++    EGLint attrib_list[] = {
++        EGL_RENDER_BUFFER, EGL_BACK_BUFFER,
++        EGL_WIDTH, width,
++        EGL_HEIGHT, height,
++        EGL_NONE,
++    };
++
++    self->wegl.egl = plat->egl.CreateWindowSurface(
++        dpy->egl, config->egl, (EGLNativeWindowType)window, attrib_list);
++    if (!self->wegl.egl) {
++        wegl_emit_error(plat, "eglCreateWindowSurface");
++        AImageReader_delete(reader);
++        return false;
++    }
++
++    self->reader = reader;
++    return true;
++}
++
++static void
++sl_imagereader_drain(struct sl_window *self)
++{
++    if (!self->reader)
++        return;
++
++    // Keep the BufferQueue from filling up: nobody displays these frames,
++    // so consume-and-drop the latest (which also releases all older ones).
++    AImage *image = NULL;
++    if (AImageReader_acquireLatestImage(self->reader, &image) == AMEDIA_OK &&
++        image)
++        AImage_delete(image);
++}
++#endif // __ANDROID__
++
+ bool
+ sl_window_destroy(struct wcore_window *wc_self)
+ {
+     struct sl_window *self = sl_window(wegl_surface(wc_self));
+     bool ok = wegl_surface_teardown(&self->wegl);
+ 
++#ifdef __ANDROID__
++    if (self->reader)
++        AImageReader_delete(self->reader);
++#endif
++
+     free(self);
+     return ok;
+ }
+@@ -52,6 +142,11 @@ sl_window_create(struct wcore_platform *wc_plat,
+ 
+     wcore_window_init(&self->wegl.wcore, wc_config);
+ 
++#ifdef __ANDROID__
++    if (sl_android_window_mode())
++        ok = sl_imagereader_init(self, wc_config, width, height);
++    else
++#endif
+     ok = wegl_pbuffer_init(&self->wegl, wc_config, width, height);
+     if (!ok)
+         goto error;
+@@ -71,6 +166,18 @@ sl_window_show(struct wcore_window *wc_self)
+     return true;
+ }
+ 
++bool
++sl_window_swap_buffers(struct wcore_window *wc_self)
++{
++    bool ok = wegl_surface_swap_buffers(wc_self);
++
++#ifdef __ANDROID__
++    sl_imagereader_drain(sl_window(wegl_surface(wc_self)));
++#endif
++
++    return ok;
++}
++
+ bool
+ sl_window_resize(struct wcore_window *wc_self,
+                  int32_t width, int32_t height)
+@@ -83,6 +190,36 @@ sl_window_resize(struct wcore_window *wc_self,
+     struct wcore_tinfo *tinfo;
+     bool ok = true;
+ 
++#ifdef __ANDROID__
++    if (self->reader) {
++        // ImageReader-backed window: build a fresh reader + EGL window
++        // surface at the new size, make it current, then retire the old one.
++        struct sl_window new_win;
++        memset(&new_win, 0, sizeof(new_win));
++        wcore_window_init(&new_win.wegl.wcore, self->wc_config);
++
++        ok = sl_imagereader_init(&new_win, self->wc_config, width, height);
++        if (!ok)
++            return false;
++
++        tinfo = wcore_tinfo_get();
++        wc_ctx = tinfo->current_context;
++
++        ok = wegl_make_current(wc_plat, wc_dpy, &new_win.wegl.wcore, wc_ctx);
++        if (!ok) {
++            wegl_surface_teardown(&new_win.wegl);
++            AImageReader_delete(new_win.reader);
++            return false;
++        }
++
++        wegl_surface_teardown(&self->wegl);
++        AImageReader_delete(self->reader);
++        self->wegl.egl = new_win.wegl.egl;
++        self->reader = new_win.reader;
++        return true;
++    }
++#endif
++
+     // Create a new pbuffer for the resized window.
+     ok = wegl_pbuffer_init(&new_wegl, self->wc_config, width, height);
+     if (!ok)
+diff --git a/src/waffle/surfaceless_egl/sl_window.h b/src/waffle/surfaceless_egl/sl_window.h
+index 8a96239..65a86ad 100644
+--- a/src/waffle/surfaceless_egl/sl_window.h
++++ b/src/waffle/surfaceless_egl/sl_window.h
+@@ -9,9 +9,18 @@
+ 
+ struct wcore_platform;
+ 
++#ifdef __ANDROID__
++typedef struct AImageReader AImageReader;
++#endif
++
+ struct sl_window {
+     struct wegl_surface wegl;
+     struct wcore_config *wc_config;
++#ifdef __ANDROID__
++    // Non-NULL when the window is backed by an AImageReader ANativeWindow
++    // (WAFFLE_ANDROID_WINDOW=imagereader) instead of an EGL pbuffer.
++    AImageReader *reader;
++#endif
+ };
+ 
+ DEFINE_CONTAINER_CAST_FUNC(sl_window,
+@@ -31,6 +40,9 @@ sl_window_destroy(struct wcore_window *wc_self);
+ bool
+ sl_window_show(struct wcore_window *wc_self);
+ 
++bool
++sl_window_swap_buffers(struct wcore_window *wc_self);
++
+ bool
+ sl_window_resize(struct wcore_window *wc_self,
+                  int32_t width, int32_t height);

tools/piglit-android/run_piglit_android.py

@@ -0,0 +1,347 @@
+#!/usr/bin/env python3
+# MobileGL - tools/piglit-android/run_piglit_android.py
+# 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
+"""Run piglit GL tests on a connected Android device against MobileGL.
+
+Pipeline:
+  1. Read a test list produced by `piglit print-cmd` ("name ::: command" lines).
+  2. Rewrite host paths to on-device paths and collect referenced data files.
+  3. Push binaries/libs/data to the device (incremental, marker-based).
+  4. Execute tests serially on-device in chunked shell scripts under `timeout`,
+     with MobileGL served to waffle via WAFFLE_EGL_LIBRARY=libMobileGL.so.
+  5. Parse "PIGLIT: {...}" result lines, classify, and write results + summary.
+
+The device never needs python or an APK; everything runs as the adb shell user
+from /data/local/tmp.
+"""
+
+import argparse
+import json
+import os
+import re
+import shlex
+import subprocess
+import sys
+import tarfile
+import tempfile
+import time
+from pathlib import Path
+
+RESULT_LINE = re.compile(rb'PIGLIT: *({.*})')
+MARK_START = re.compile(rb'@@@T (\d+) START')
+MARK_EXIT = re.compile(rb'@@@T (\d+) EXIT (\d+)')
+
+# toybox timeout exit codes: 124 = timed out (SIGTERM), 137 = SIGKILL after -k.
+TIMEOUT_EXITS = {124, 137, 142}
+
+
+def adb(args, serial=None, **kw):
+    cmd = ['adb'] + (['-s', serial] if serial else []) + args
+    return subprocess.run(cmd, **kw)
+
+
+def adb_check(args, serial=None):
+    r = adb(args, serial=serial, capture_output=True)
+    if r.returncode != 0:
+        raise RuntimeError(f"adb {' '.join(args)} failed: {r.stderr.decode(errors='replace')}")
+    return r.stdout
+
+
+def parse_list(path):
+    tests = []
+    for line in Path(path).read_text().splitlines():
+        line = line.strip()
+        if not line or ' ::: ' not in line:
+            continue
+        name, cmd = line.split(' ::: ', 1)
+        tests.append((name.strip(), shlex.split(cmd)))
+    return tests
+
+
+def rewrite_cmd(argv, piglit_root, build_dir, device_dir):
+    """Map host paths in a test command to device paths.
+
+    Returns (device_argv, referenced_host_files).
+    """
+    build_abs = str((piglit_root / build_dir).resolve())
+    src_abs = str(piglit_root.resolve())
+    out = []
+    refs = []
+    for i, arg in enumerate(argv):
+        a = arg
+        if i == 0:
+            # program: "build-android/bin/foo" or absolute
+            prog = a if os.path.isabs(a) else str((piglit_root / a).resolve())
+            refs.append(prog)
+            out.append(f'{device_dir}/bin/{os.path.basename(prog)}')
+            continue
+        p = a if os.path.isabs(a) else None
+        if p and p.startswith(build_abs + '/generated_tests/'):
+            refs.append(p)
+            out.append(p.replace(build_abs + '/generated_tests', device_dir + '/generated_tests', 1))
+        elif p and p.startswith(build_abs + '/tests/'):
+            # Serialized profiles record data paths under the build dir, but
+            # only *generated* data lives there in an out-of-tree build;
+            # plain test files stay in the source tests/ dir.
+            if not os.path.exists(p):
+                p = p.replace(build_abs + '/tests', src_abs + '/tests', 1)
+                refs.append(p)
+                out.append(p.replace(src_abs + '/tests', device_dir + '/tests', 1))
+            else:
+                refs.append(p)
+                out.append(p.replace(build_abs + '/tests', device_dir + '/build-tests', 1))
+        elif p and p.startswith(src_abs + '/tests/'):
+            refs.append(p)
+            out.append(p.replace(src_abs + '/tests', device_dir + '/tests', 1))
+        else:
+            out.append(a)
+    return out, refs
+
+
+def backend_env(backend, device_dir, use_angle=False):
+    env = {
+        'LD_LIBRARY_PATH': f'{device_dir}/lib',
+        'WAFFLE_EGL_LIBRARY': 'libMobileGL.so',
+        'WAFFLE_GL_LIBRARY': 'libMobileGL.so',
+        # Upgrade low compat context requests to what MobileGL implements;
+        # without this, piglit's supports_gl_compat_version=10 tests get the
+        # raw backend version string and skip themselves.
+        'WAFFLE_FORCE_GL_CONTEXT_VERSION': '33core',
+        'PIGLIT_PLATFORM': 'surfaceless_egl',
+        'PIGLIT_SOURCE_DIR': device_dir,
+        'MOBILEGL_BACKEND_TYPE': backend,
+        'MOBILEGL_LOG_FILE_PATH': f'{device_dir}/mobilegl.log',
+    }
+    if backend == 'DirectGLES':
+        env['MOBILEGL_USE_ANGLE'] = '1' if use_angle else '0'
+    if backend == 'DirectVulkan':
+        # The device ICD usually lacks VK_EXT_headless_surface, which the
+        # MobileGL pbuffer path needs; use a real ANativeWindow from
+        # AImageReader instead (patched waffle surfaceless_egl platform).
+        env['WAFFLE_ANDROID_WINDOW'] = 'imagereader'
+    return env
+
+
+def push_tree(host, dev, serial, marker_name, force=False):
+    """Push a directory once, tracked by a marker file on the device."""
+    marker = f'{dev}.pushed-{marker_name}'
+    if not force:
+        r = adb(['shell', f'test -e {shlex.quote(marker)} && echo yes'],
+                serial=serial, capture_output=True)
+        if r.stdout.strip() == b'yes':
+            return False
+    print(f'  pushing {host} -> {dev}')
+    adb_check(['shell', f'rm -rf {shlex.quote(dev)}'], serial=serial)
+    adb_check(['push', str(host), dev], serial=serial)
+    adb_check(['shell', f'touch {shlex.quote(marker)}'], serial=serial)
+    return True
+
+
+def push_files(files, strip_prefix, dev_prefix, serial):
+    """Tar up referenced data files (relative to strip_prefix) and unpack on device."""
+    files = sorted(set(files))
+    if not files:
+        return
+    with tempfile.NamedTemporaryFile(suffix='.tar', delete=False) as tf:
+        tar_path = tf.name
+    with tarfile.open(tar_path, 'w') as tar:
+        for f in files:
+            rel = os.path.relpath(f, strip_prefix)
+            if rel.startswith('..'):
+                raise RuntimeError(f'file {f} not under {strip_prefix}')
+            tar.add(f, arcname=rel)
+    dev_tar = dev_prefix + '/.data.tar'
+    adb_check(['shell', f'mkdir -p {shlex.quote(dev_prefix)}'], serial=serial)
+    adb_check(['push', tar_path, dev_tar], serial=serial)
+    adb_check(['shell', f'cd {shlex.quote(dev_prefix)} && tar xf .data.tar && rm .data.tar'],
+              serial=serial)
+    os.unlink(tar_path)
+
+
+def classify(exit_code, piglit_results, timed_out):
+    if timed_out:
+        return 'timeout'
+    result = None
+    for r in piglit_results:
+        if 'result' in r:
+            result = r['result']
+    if result is None:
+        return 'crash' if exit_code != 0 else 'notrun'
+    if exit_code not in (0, 1):
+        return 'crash'
+    return result
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument('--piglit-root', required=True, type=Path,
+                    help='piglit source checkout (with the Android build dir inside)')
+    ap.add_argument('--build-dir', default='build-android',
+                    help='Android build dir name inside piglit root')
+    ap.add_argument('--list', required=True,
+                    help='test list file from `piglit print-cmd` (name ::: cmd)')
+    ap.add_argument('--backend', required=True, choices=['DirectGLES', 'DirectVulkan'])
+    ap.add_argument('--use-angle', action='store_true',
+                    help='DirectGLES only: let MobileGL load ANGLE instead of the system driver')
+    ap.add_argument('--mobilegl-lib', required=True, type=Path,
+                    help='stripped libMobileGL.so for the device')
+    ap.add_argument('--waffle-lib', required=True, type=Path,
+                    help='cross-built libwaffle-1.so')
+    ap.add_argument('--device-dir', default='/data/local/tmp/piglit-mgl')
+    ap.add_argument('--out', required=True, type=Path, help='host results directory')
+    ap.add_argument('--serial', default=None, help='adb device serial')
+    ap.add_argument('--timeout', type=int, default=60, help='per-test timeout (seconds)')
+    ap.add_argument('--chunk', type=int, default=200, help='tests per device-side script')
+    ap.add_argument('--repush', action='store_true', help='force re-push of bin/lib/tests trees')
+    args = ap.parse_args()
+
+    piglit_root = args.piglit_root.resolve()
+    build = piglit_root / args.build_dir
+    dev = args.device_dir.rstrip('/')
+    args.out.mkdir(parents=True, exist_ok=True)
+
+    tests = parse_list(args.list)
+    if not tests:
+        print('no tests in list', file=sys.stderr)
+        return 2
+
+    print(f'[1/4] {len(tests)} tests; rewriting paths')
+    dev_cmds = []
+    data_refs = []
+    for name, argv in tests:
+        dcmd, refs = rewrite_cmd(argv, piglit_root, args.build_dir, dev)
+        dev_cmds.append((name, dcmd))
+        data_refs.extend(r for r in refs if not r.split('/')[-2:][0] == 'bin')
+
+    print('[2/4] pushing artifacts')
+    adb_check(['shell', f'mkdir -p {dev} {dev}/lib {dev}/chunks {dev}/logs'], serial=args.serial)
+    push_tree(build / 'bin', f'{dev}/bin', args.serial, 'bin', args.repush)
+    adb_check(['shell', f'chmod -R 755 {dev}/bin'], serial=args.serial)
+    # libs: piglit utils + waffle + MobileGL in one LD_LIBRARY_PATH dir
+    for lib in sorted((build / 'lib').glob('*.so')):
+        adb_check(['push', str(lib), f'{dev}/lib/'], serial=args.serial)
+    adb_check(['push', str(args.waffle_lib), f'{dev}/lib/libwaffle-1.so'], serial=args.serial)
+    adb_check(['push', str(args.mobilegl_lib), f'{dev}/lib/libMobileGL.so'], serial=args.serial)
+    # data files referenced by this run (tests/, build-tests/, generated_tests/)
+    src_abs = str(piglit_root)
+    build_abs = str(build)
+    groups = {
+        (src_abs + '/tests', f'{dev}/tests'): [],
+        (build_abs + '/tests', f'{dev}/build-tests'): [],
+        (build_abs + '/generated_tests', f'{dev}/generated_tests'): [],
+    }
+    for r in set(data_refs):
+        for (host_prefix, dev_prefix), bucket in groups.items():
+            if r.startswith(host_prefix + '/'):
+                bucket.append(r)
+                break
+    for (host_prefix, dev_prefix), bucket in groups.items():
+        push_files(bucket, host_prefix, dev_prefix, args.serial)
+
+    env = backend_env(args.backend, dev, args.use_angle)
+    env_lines = '\n'.join(f'export {k}={shlex.quote(v)}' for k, v in env.items())
+
+    print(f'[3/4] running {len(tests)} tests on {args.backend} '
+          f'(timeout {args.timeout}s/test, chunks of {args.chunk})')
+    raw_log = (args.out / 'raw.log').open('wb')
+    t0 = time.time()
+    outcomes = {}
+    chunks = [dev_cmds[i:i + args.chunk] for i in range(0, len(dev_cmds), args.chunk)]
+    idx_base = 0
+    for ci, chunk in enumerate(chunks):
+        lines = ['#!/system/bin/sh', env_lines, f'cd {dev}']
+        for j, (name, dcmd) in enumerate(chunk):
+            idx = idx_base + j
+            quoted = ' '.join(shlex.quote(a) for a in dcmd)
+            lines.append(f'echo "@@@T {idx} START"')
+            lines.append(f'timeout -k 5 {args.timeout} {quoted} </dev/null 2>&1')
+            lines.append(f'echo "@@@T {idx} EXIT $?"')
+        script = '\n'.join(lines) + '\n'
+        with tempfile.NamedTemporaryFile('w', suffix='.sh', delete=False) as tf:
+            tf.write(script)
+            host_script = tf.name
+        dev_script = f'{dev}/chunks/chunk{ci}.sh'
+        adb_check(['push', host_script, dev_script], serial=args.serial)
+        os.unlink(host_script)
+        r = adb(['shell', f'sh {dev_script}'], serial=args.serial, capture_output=True)
+        raw_log.write(r.stdout)
+        raw_log.flush()
+
+        # parse this chunk
+        cur = None
+        cur_results = []
+        cur_out = []
+        for line in r.stdout.splitlines():
+            m = MARK_START.search(line)
+            if m:
+                cur = int(m.group(1))
+                cur_results = []
+                cur_out = []
+                continue
+            m = MARK_EXIT.search(line)
+            if m and cur is not None and int(m.group(1)) == cur:
+                code = int(m.group(2))
+                name = dev_cmds[cur][0]
+                status = classify(code, cur_results, code in TIMEOUT_EXITS)
+                outcomes[name] = {
+                    'status': status,
+                    'exit': code,
+                    'subtests': {k: v for r_ in cur_results if 'subtest' in r_
+                                 for k, v in r_['subtest'].items()},
+                    'tail': b'\n'.join(cur_out[-8:]).decode(errors='replace')
+                            if status in ('crash', 'timeout', 'fail', 'notrun') else '',
+                }
+                cur = None
+                continue
+            if cur is not None:
+                cur_out.append(line)
+                m = RESULT_LINE.search(line)
+                if m:
+                    try:
+                        cur_results.append(json.loads(m.group(1)))
+                    except json.JSONDecodeError:
+                        pass
+        # tests whose markers never appeared (adb drop, device reboot)
+        for j, (name, _) in enumerate(chunk):
+            outcomes.setdefault(dev_cmds[idx_base + j][0], {
+                'status': 'missing', 'exit': -1, 'subtests': {}, 'tail': ''})
+        idx_base += len(chunk)
+        done = idx_base
+        print(f'  chunk {ci + 1}/{len(chunks)} done ({done}/{len(dev_cmds)}, '
+              f'{time.time() - t0:.0f}s elapsed)')
+    raw_log.close()
+
+    print('[4/4] writing results')
+    counts = {}
+    for name, o in outcomes.items():
+        counts[o['status']] = counts.get(o['status'], 0) + 1
+    result_doc = {
+        'backend': args.backend,
+        'use_angle': args.use_angle,
+        'device_dir': dev,
+        'timeout': args.timeout,
+        'elapsed_sec': round(time.time() - t0, 1),
+        'totals': counts,
+        'tests': outcomes,
+    }
+    (args.out / 'results.json').write_text(json.dumps(result_doc, indent=1, sort_keys=True))
+    lines = [f"backend: {args.backend}  tests: {len(outcomes)}  elapsed: {result_doc['elapsed_sec']}s"]
+    lines.append('totals: ' + ', '.join(f'{k}={v}' for k, v in sorted(counts.items())))
+    for status in ('crash', 'timeout', 'fail', 'missing', 'notrun', 'warn'):
+        bad = sorted(n for n, o in outcomes.items() if o['status'] == status)
+        if bad:
+            lines.append(f'\n== {status} ({len(bad)}):')
+            lines.extend(f'  {n}' for n in bad)
+    (args.out / 'summary.txt').write_text('\n'.join(lines) + '\n')
+    print('\n'.join(lines[:2]))
+    print(f"results: {args.out / 'results.json'}")
+    return 0
+
+
+if __name__ == '__main__':
+    sys.exit(main())

tools/piglit-android/skills/piglit-on-android/SKILL.md

@@ -0,0 +1,112 @@
+---
+name: piglit-on-android
+description: Run piglit desktop-GL tests on a connected Android device with MobileGL as the OpenGL implementation - cross-build waffle (patched) and piglit for aarch64, push to /data/local/tmp, execute over adb against the DirectGLES (system driver or ANGLE) and DirectVulkan backends, and produce pass/fail/crash summaries. Use when validating MobileGL's OpenGL 3.3 core conformance on-device or bisecting a piglit regression.
+---
+
+# piglit on Android against MobileGL
+
+## Variables
+
+```sh
+export REPO="$PWD"                        # MobileGL checkout
+export WORK="$HOME/piglit-android"        # piglit/waffle workdir (outside the repo)
+export NDK="$HOME/Library/Android/sdk/ndk/27.3.13750724"
+export TOOLS="$REPO/tools/piglit-android"
+export DEVICE_DIR="/data/local/tmp/piglit-mgl"
+```
+
+## Architecture (read first)
+
+- MobileGL exports the full `egl*`/`gl*` API with real symbol names from
+  `libMobileGL.so`; waffle's `surfaceless_egl` platform is pointed at it via
+  `WAFFLE_EGL_LIBRARY=libMobileGL.so` and `WAFFLE_GL_LIBRARY=libMobileGL.so`.
+- **Never ship MobileGL under the name `libEGL.so`**: the DirectGLES backend
+  dlopens the system driver by the bare soname `libEGL.so` and would load
+  itself recursively.
+- `WAFFLE_FORCE_GL_CONTEXT_VERSION=33core` (patched waffle) upgrades piglit's
+  low compat context requests (`supports_gl_compat_version=10`) to GL 3.3
+  core; without it those tests see a raw backend version string and skip.
+- DirectVulkan cannot use MobileGL's EGL-pbuffer path on real devices (Android
+  ICDs lack `VK_EXT_headless_surface`), so the patched waffle creates windows
+  from an `AImageReader` ANativeWindow: `WAFFLE_ANDROID_WINDOW=imagereader`.
+  The runner sets this automatically for `--backend DirectVulkan`.
+- The piglit patch also fixes `piglit_dispatch_default_init` to always install
+  the waffle resolvers; upstream silently bound gl* through the SYSTEM
+  libEGL's `eglGetProcAddress`, running every test on the raw GLES driver.
+- MobileGL's lifecycle is owned by the EGL layer (lazy init on first EGL call,
+  full teardown on the last `eglTerminate`), so piglit processes start and
+  exit cleanly with no static ctor/dtor involvement.
+
+## Steps
+
+1. **Clone + patch** piglit and waffle (shallow clones are fine; never pull
+   git-LFS):
+
+   ```sh
+   mkdir -p "$WORK" && cd "$WORK"
+   git clone --depth 1 https://gitlab.freedesktop.org/mesa/piglit.git
+   git clone --depth 1 https://gitlab.freedesktop.org/mesa/waffle.git
+   git -C waffle apply "$TOOLS/patches/waffle-mobilegl-android.patch"
+   git -C piglit apply "$TOOLS/patches/piglit-mobilegl-android.patch"
+   python3 -m venv venv && ./venv/bin/pip install mako numpy packaging
+   ```
+
+2. **Build libMobileGL.so** for arm64 (plain CMake + NDK; gradle not needed)
+   and strip it. See `$TOOLS/README.md` for exact invocations.
+
+3. **Cross-build waffle** with meson (`surfaceless_egl` enabled, everything
+   else disabled) using `$TOOLS/cross-example/android-arm64.ini` and the stub
+   `egl.pc` (Cflags → `$REPO/include` for EGL 1.5 headers). `meson install` to
+   `$WORK/prefix` so piglit's pkg-config finds `waffle-1.pc`.
+
+4. **Cross-build piglit** with the NDK toolchain file; the full CMake flag set
+   is in `$TOOLS/README.md`. All GL test binaries land in
+   `piglit/build-android/bin` (~1600), shared utils in `lib/`.
+
+5. **Smoke test** with wflinfo on both backends before any long run:
+
+   ```sh
+   adb shell "cd $DEVICE_DIR && env LD_LIBRARY_PATH=$DEVICE_DIR/lib \
+     WAFFLE_EGL_LIBRARY=libMobileGL.so WAFFLE_GL_LIBRARY=libMobileGL.so \
+     MOBILEGL_BACKEND_TYPE=DirectVulkan WAFFLE_ANDROID_WINDOW=imagereader \
+     ./wflinfo --platform surfaceless_egl --api gl --version 3.3 --profile core"
+   ```
+
+   Expect `3.3.0 MobileGL … Direct (Vulkan) Backend` and exit code 0.
+
+6. **Enumerate tests on the host** with `piglit print-cmd` (profiles: `opengl`,
+   `shader`, `glslparser`; group separator is `@`). For the GL 3.3 core suite
+   use the version groups + GLSL groups + the ARB extension groups folded into
+   3.1–3.3 core (~15k tests).
+
+7. **Run per backend** with the driver script (serial, chunked, per-test
+   timeout; DirectGLES first, then DirectVulkan):
+
+   ```sh
+   python3 "$TOOLS/run_piglit_android.py" \
+     --piglit-root "$WORK/piglit" --list gl33-core.list \
+     --backend DirectGLES \
+     --mobilegl-lib "$WORK/libMobileGL-stripped.so" \
+     --waffle-lib "$WORK/waffle/build-android/src/waffle/libwaffle-1.so" \
+     --out results-gles
+   ```
+
+   `--use-angle` switches DirectGLES to packaged ANGLE; default is the system
+   GLES driver. `--repush` forces re-pushing bin/lib/tests trees (needed after
+   rebuilds or if the device wiped `/data/local/tmp`).
+
+8. **Compare backends**: `results.json` has per-test status
+   (`pass/fail/skip/crash/timeout/missing`) and output tails for non-passes;
+   `summary.txt` lists the bad tests. Diff the two runs' failing sets to
+   separate frontend issues (fail on both) from backend-specific ones.
+
+## Gotchas
+
+- A test crash after `PIGLIT: {"result": ...}` was printed is classified by
+  the result line, not the exit code, except that unknown nonzero exits count
+  as `crash`.
+- The device may clear `/data/local/tmp` (vendor cleaners); `--repush` recovers.
+- Keep chunks ≤ ~250 tests: one adb connection per chunk bounds the damage of
+  USB hiccups, and progress prints per chunk.
+- MobileGL writes `$DEVICE_DIR/mobilegl.log` (set by the runner); check it and
+  `adb logcat -b crash` when triaging crashes.

tools/piglit-android/skills/piglit-on-android/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "MobileGL piglit on Android"
+  short_description: "Run piglit GL tests on-device against MobileGL backends"
+  default_prompt: "Use $piglit-on-android to run the piglit OpenGL 3.3 core suite on the connected Android device against MobileGL's DirectGLES and DirectVulkan backends and summarize the results."

tools/wgl-smoke/README.md

@@ -0,0 +1,25 @@
+# WGL smoke tests
+
+Manual bring-up tests for the Windows host layer (drop-in `opengl32.dll`).
+Both replicate the GLFW/LWJGL bootstrap Minecraft uses and verify a rendered
+triangle via `glReadPixels`; exit code 0 means pass.
+
+- `wgl_smoke.cpp` - hand-rolled Win32/WGL sequence (dummy context on a hidden
+  helper window, ARB pixel format + 3.2 core forward-compatible context,
+  gdi32 pixel-format forwarding, swap loop, readback).
+- `glfw_smoke.cpp` - the same flow driven by a real GLFW binary, which also
+  exercises GLFW's zero-area helper window (the case that needs the
+  DirectVulkan no-swapchain-at-init path).
+
+Build (x64 Native Tools prompt), with GLFW binaries for the second test:
+
+    cl /nologo /EHsc /W3 wgl_smoke.cpp user32.lib gdi32.lib
+    cl /nologo /EHsc /W3 glfw_smoke.cpp /I%GLFW%\include %GLFW%\lib-vc2022\glfw3dll.lib user32.lib gdi32.lib
+
+Run with the MobileGL `opengl32.dll` copy (and `glfw3.dll`, plus ANGLE's
+`libEGL.dll`/`libGLESv2.dll`/`d3dcompiler_47.dll` for DirectGLES) in the exe
+directory:
+
+    set MOBILEGL_BACKEND_TYPE=DirectVulkan   (or DirectGLES)
+    set MOBILEGL_LOG_FILE_PATH=smoke.log
+    wgl_smoke.exe && glfw_smoke.exe

tools/wgl-smoke/glfw_smoke.cpp

@@ -0,0 +1,189 @@
+// GLFW-driven smoke test: exercises the exact GLFW WGL bootstrap Minecraft uses
+// (helper-window dummy context, ARB pixel format, 3.2 core forward-compatible
+// no-error context) against the MobileGL opengl32.dll drop-in.
+//
+// Build: cl /nologo /EHsc /W3 glfw_smoke.cpp /I<GLFW>/include <GLFW>/lib-vc2022/glfw3dll.lib user32.lib gdi32.lib
+// Run with opengl32.dll + glfw3.dll in the exe directory.
+
+#include <GLFW/glfw3.h>
+#include <windows.h>
+#include <cstdio>
+
+typedef const unsigned char*(WINAPI* PFN_glGetString)(unsigned int);
+typedef void(WINAPI* PFN_glClearColor)(float, float, float, float);
+typedef void(WINAPI* PFN_glClear)(unsigned int);
+typedef void(WINAPI* PFN_glReadPixels)(int, int, int, int, unsigned int, unsigned int, void*);
+typedef void(WINAPI* PFN_glFinish)(void);
+typedef unsigned int(WINAPI* PFN_glCreateShader)(unsigned int);
+typedef void(WINAPI* PFN_glShaderSource)(unsigned int, int, const char* const*, const int*);
+typedef void(WINAPI* PFN_glCompileShader)(unsigned int);
+typedef unsigned int(WINAPI* PFN_glCreateProgram)(void);
+typedef void(WINAPI* PFN_glAttachShader)(unsigned int, unsigned int);
+typedef void(WINAPI* PFN_glLinkProgram)(unsigned int);
+typedef void(WINAPI* PFN_glUseProgram)(unsigned int);
+typedef void(WINAPI* PFN_glGenVertexArrays)(int, unsigned int*);
+typedef void(WINAPI* PFN_glBindVertexArray)(unsigned int);
+typedef void(WINAPI* PFN_glGenBuffers)(int, unsigned int*);
+typedef void(WINAPI* PFN_glBindBuffer)(unsigned int, unsigned int);
+typedef void(WINAPI* PFN_glBufferData)(unsigned int, long long, const void*, unsigned int);
+typedef void(WINAPI* PFN_glVertexAttribPointer)(unsigned int, int, unsigned int, unsigned char, int, const void*);
+typedef void(WINAPI* PFN_glEnableVertexAttribArray)(unsigned int);
+typedef void(WINAPI* PFN_glDrawArrays)(unsigned int, int, int);
+#define GL_COLOR_BUFFER_BIT 0x00004000
+#define GL_VERSION 0x1F02
+#define GL_RENDERER 0x1F01
+#define GL_RGBA 0x1908
+#define GL_UNSIGNED_BYTE 0x1401
+#define GL_TRIANGLES 0x0004
+#define GL_FLOAT 0x1406
+#define GL_ARRAY_BUFFER 0x8892
+#define GL_STATIC_DRAW 0x88E4
+#define GL_VERTEX_SHADER 0x8B31
+#define GL_FRAGMENT_SHADER 0x8B30
+
+static void RawPrint(HANDLE f, const char* s) {
+    DWORD written = 0;
+    WriteFile(f, s, (DWORD)lstrlenA(s), &written, nullptr);
+}
+
+static void RawHex(HANDLE f, unsigned long long v) {
+    char buf[20];
+    for (int i = 0; i < 16; ++i) {
+        int nib = (int)((v >> ((15 - i) * 4)) & 0xF);
+        buf[i] = (char)(nib < 10 ? '0' + nib : 'a' + nib - 10);
+    }
+    buf[16] = 0;
+    RawPrint(f, buf);
+}
+
+static LONG WINAPI CrashDump(EXCEPTION_POINTERS* ep) {
+    static volatile LONG entered = 0;
+    if (ep->ExceptionRecord->ExceptionCode != 0xC0000005) return EXCEPTION_CONTINUE_SEARCH;
+    if (InterlockedExchange(&entered, 1)) return EXCEPTION_CONTINUE_SEARCH;
+    HANDLE f = CreateFileA("glfw-crash-stack.txt", GENERIC_WRITE, 0, nullptr, CREATE_ALWAYS,
+                           FILE_ATTRIBUTE_NORMAL, nullptr);
+    if (f == INVALID_HANDLE_VALUE) TerminateProcess(GetCurrentProcess(), 0x43);
+    RawPrint(f, "AV at ");
+    RawHex(f, (unsigned long long)ep->ExceptionRecord->ExceptionAddress);
+    RawPrint(f, " reading ");
+    RawHex(f, (unsigned long long)ep->ExceptionRecord->ExceptionInformation[1]);
+    RawPrint(f, "\n");
+    const unsigned long long* sp = (const unsigned long long*)ep->ContextRecord->Rsp;
+    for (int i = 0; i < 512; ++i) {
+        MEMORY_BASIC_INFORMATION mbi;
+        if (!VirtualQuery(sp + i, &mbi, sizeof(mbi)) || mbi.State != MEM_COMMIT ||
+            (mbi.Protect & (PAGE_READONLY | PAGE_READWRITE | PAGE_EXECUTE_READ | PAGE_EXECUTE_READWRITE)) == 0)
+            break;
+        unsigned long long v = sp[i];
+        HMODULE mod = nullptr;
+        if (v > 0x10000 &&
+            GetModuleHandleExA(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS | GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT,
+                               (LPCSTR)v, &mod) &&
+            mod) {
+            char name[MAX_PATH];
+            name[0] = 0;
+            GetModuleFileNameA(mod, name, MAX_PATH);
+            RawPrint(f, "stack ");
+            RawHex(f, v);
+            RawPrint(f, " ");
+            RawPrint(f, name);
+            RawPrint(f, "+");
+            RawHex(f, v - (unsigned long long)mod);
+            RawPrint(f, "\n");
+        }
+    }
+    CloseHandle(f);
+    TerminateProcess(GetCurrentProcess(), 0x42);
+    return EXCEPTION_EXECUTE_HANDLER;
+}
+
+int main() {
+    AddVectoredExceptionHandler(1, CrashDump);
+    if (!glfwInit()) {
+        std::fprintf(stderr, "GLFW-SMOKE FAIL: glfwInit\n");
+        return 1;
+    }
+    glfwWindowHint(GLFW_CLIENT_API, GLFW_OPENGL_API);
+    glfwWindowHint(GLFW_CONTEXT_VERSION_MAJOR, 3);
+    glfwWindowHint(GLFW_CONTEXT_VERSION_MINOR, 2);
+    glfwWindowHint(GLFW_OPENGL_PROFILE, GLFW_OPENGL_CORE_PROFILE);
+    glfwWindowHint(GLFW_OPENGL_FORWARD_COMPAT, GLFW_TRUE);
+    glfwWindowHint(GLFW_CONTEXT_NO_ERROR, GLFW_TRUE);
+    glfwWindowHint(GLFW_VISIBLE, GLFW_FALSE);
+
+    GLFWwindow* win = glfwCreateWindow(854, 480, "MobileGL GLFW smoke", nullptr, nullptr);
+    if (!win) {
+        const char* desc = nullptr;
+        int code = glfwGetError(&desc);
+        std::fprintf(stderr, "GLFW-SMOKE FAIL: glfwCreateWindow (0x%x %s)\n", code, desc ? desc : "?");
+        return 1;
+    }
+    glfwMakeContextCurrent(win);
+    glfwShowWindow(win);
+    glfwSwapInterval(0);
+
+    auto glGetString_ = (PFN_glGetString)glfwGetProcAddress("glGetString");
+    auto glClearColor_ = (PFN_glClearColor)glfwGetProcAddress("glClearColor");
+    auto glClear_ = (PFN_glClear)glfwGetProcAddress("glClear");
+    auto glReadPixels_ = (PFN_glReadPixels)glfwGetProcAddress("glReadPixels");
+    auto glFinish_ = (PFN_glFinish)glfwGetProcAddress("glFinish");
+    std::printf("GL_VERSION: %s\nGL_RENDERER: %s\n", glGetString_(GL_VERSION), glGetString_(GL_RENDERER));
+
+    auto glCreateShader_ = (PFN_glCreateShader)glfwGetProcAddress("glCreateShader");
+    auto glShaderSource_ = (PFN_glShaderSource)glfwGetProcAddress("glShaderSource");
+    auto glCompileShader_ = (PFN_glCompileShader)glfwGetProcAddress("glCompileShader");
+    auto glCreateProgram_ = (PFN_glCreateProgram)glfwGetProcAddress("glCreateProgram");
+    auto glAttachShader_ = (PFN_glAttachShader)glfwGetProcAddress("glAttachShader");
+    auto glLinkProgram_ = (PFN_glLinkProgram)glfwGetProcAddress("glLinkProgram");
+    auto glUseProgram_ = (PFN_glUseProgram)glfwGetProcAddress("glUseProgram");
+    auto glGenVertexArrays_ = (PFN_glGenVertexArrays)glfwGetProcAddress("glGenVertexArrays");
+    auto glBindVertexArray_ = (PFN_glBindVertexArray)glfwGetProcAddress("glBindVertexArray");
+    auto glGenBuffers_ = (PFN_glGenBuffers)glfwGetProcAddress("glGenBuffers");
+    auto glBindBuffer_ = (PFN_glBindBuffer)glfwGetProcAddress("glBindBuffer");
+    auto glBufferData_ = (PFN_glBufferData)glfwGetProcAddress("glBufferData");
+    auto glVertexAttribPointer_ = (PFN_glVertexAttribPointer)glfwGetProcAddress("glVertexAttribPointer");
+    auto glEnableVertexAttribArray_ = (PFN_glEnableVertexAttribArray)glfwGetProcAddress("glEnableVertexAttribArray");
+    auto glDrawArrays_ = (PFN_glDrawArrays)glfwGetProcAddress("glDrawArrays");
+
+    const char* vsSrc = "#version 150 core\nin vec2 aPos;\nvoid main(){ gl_Position = vec4(aPos, 0.0, 1.0); }\n";
+    const char* fsSrc = "#version 150 core\nout vec4 c;\nvoid main(){ c = vec4(1.0, 0.5, 0.0, 1.0); }\n";
+    unsigned int vs = glCreateShader_(GL_VERTEX_SHADER);
+    glShaderSource_(vs, 1, &vsSrc, nullptr);
+    glCompileShader_(vs);
+    unsigned int fs = glCreateShader_(GL_FRAGMENT_SHADER);
+    glShaderSource_(fs, 1, &fsSrc, nullptr);
+    glCompileShader_(fs);
+    unsigned int prog = glCreateProgram_();
+    glAttachShader_(prog, vs);
+    glAttachShader_(prog, fs);
+    glLinkProgram_(prog);
+    glUseProgram_(prog);
+    const float verts[] = {-0.9f, -0.9f, 0.9f, -0.9f, 0.0f, 0.9f};
+    unsigned int vao = 0, vbo = 0;
+    glGenVertexArrays_(1, &vao);
+    glBindVertexArray_(vao);
+    glGenBuffers_(1, &vbo);
+    glBindBuffer_(GL_ARRAY_BUFFER, vbo);
+    glBufferData_(GL_ARRAY_BUFFER, sizeof(verts), verts, GL_STATIC_DRAW);
+    glVertexAttribPointer_(0, 2, GL_FLOAT, 0, 0, nullptr);
+    glEnableVertexAttribArray_(0);
+
+    unsigned char px[4] = {};
+    for (int i = 0; i < 60 && !glfwWindowShouldClose(win); ++i) {
+        glClearColor_(0.1f, 0.2f, 0.4f, 1.0f);
+        glClear_(GL_COLOR_BUFFER_BIT);
+        glDrawArrays_(GL_TRIANGLES, 0, 3);
+        if (i == 58) {
+            glFinish_();
+            glReadPixels_(427, 240, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, px);
+        }
+        glfwSwapBuffers(win);
+        glfwPollEvents();
+    }
+    std::printf("pixel: %u,%u,%u,%u\n", px[0], px[1], px[2], px[3]);
+    glfwDestroyWindow(win);
+    glfwTerminate();
+    const bool ok = px[0] > 200 && px[1] > 90 && px[1] < 160 && px[2] < 50;
+    std::printf(ok ? "GLFW-SMOKE PASS\n" : "GLFW-SMOKE FAIL: wrong pixel\n");
+    return ok ? 0 : 1;
+}

tools/wgl-smoke/wgl_smoke.cpp

@@ -0,0 +1,472 @@
+// WGL smoke test for the MobileGL opengl32.dll drop-in.
+// Replicates GLFW 3.3.3's WGL bootstrap exactly as Minecraft 1.21.4 drives it:
+//   1. dummy window: gdi32 ChoosePixelFormat/SetPixelFormat -> wglCreateContext
+//      -> wglMakeCurrent -> probe wglGetProcAddress/extension strings
+//      -> wglMakeCurrent(NULL) -> wglDeleteContext
+//   2. real window: SetPixelFormat -> wglCreateContextAttribsARB (3.2 core,
+//      forward-compatible) -> wglMakeCurrent -> render loop with SwapBuffers
+// Verifies rendering via glReadPixels. Exit code 0 = pass.
+//
+// Build: cl /nologo /EHsc /W3 wgl_smoke.cpp user32.lib gdi32.lib
+// Run with opengl32.dll (MobileGL) in the exe directory.
+
+#include <windows.h>
+#include <cstdio>
+#include <cstring>
+#include <cstdlib>
+
+typedef unsigned int GLenum;
+typedef unsigned int GLuint;
+typedef int GLint;
+typedef int GLsizei;
+typedef float GLfloat;
+typedef unsigned char GLubyte;
+typedef char GLchar;
+typedef signed long long GLsizeiptr;
+typedef unsigned int GLbitfield;
+
+#define GL_COLOR_BUFFER_BIT 0x00004000
+#define GL_DEPTH_BUFFER_BIT 0x00000100
+#define GL_TRIANGLES 0x0004
+#define GL_FLOAT 0x1406
+#define GL_FALSE 0
+#define GL_ARRAY_BUFFER 0x8892
+#define GL_STATIC_DRAW 0x88E4
+#define GL_FRAGMENT_SHADER 0x8B30
+#define GL_VERTEX_SHADER 0x8B31
+#define GL_COMPILE_STATUS 0x8B81
+#define GL_LINK_STATUS 0x8B82
+#define GL_VENDOR 0x1F00
+#define GL_RENDERER 0x1F01
+#define GL_VERSION 0x1F02
+#define GL_RGBA 0x1908
+#define GL_UNSIGNED_BYTE 0x1401
+#define GL_MAJOR_VERSION 0x821B
+#define GL_MINOR_VERSION 0x821C
+#define GL_NO_ERROR 0
+
+#define WGL_CONTEXT_MAJOR_VERSION_ARB 0x2091
+#define WGL_CONTEXT_MINOR_VERSION_ARB 0x2092
+#define WGL_CONTEXT_FLAGS_ARB 0x2094
+#define WGL_CONTEXT_FORWARD_COMPATIBLE_BIT_ARB 0x0002
+#define WGL_CONTEXT_PROFILE_MASK_ARB 0x9126
+#define WGL_CONTEXT_CORE_PROFILE_BIT_ARB 0x00000001
+#define WGL_CONTEXT_OPENGL_NO_ERROR_ARB 0x31B3
+#define WGL_NUMBER_PIXEL_FORMATS_ARB 0x2000
+#define WGL_SUPPORT_OPENGL_ARB 0x2010
+#define WGL_DRAW_TO_WINDOW_ARB 0x2001
+#define WGL_PIXEL_TYPE_ARB 0x2013
+#define WGL_TYPE_RGBA_ARB 0x202B
+#define WGL_DEPTH_BITS_ARB 0x2022
+#define WGL_STENCIL_BITS_ARB 0x2023
+#define WGL_DOUBLE_BUFFER_ARB 0x2011
+
+static HMODULE g_opengl32;
+
+typedef HGLRC(WINAPI* PFN_wglCreateContext)(HDC);
+typedef BOOL(WINAPI* PFN_wglDeleteContext)(HGLRC);
+typedef PROC(WINAPI* PFN_wglGetProcAddress)(LPCSTR);
+typedef BOOL(WINAPI* PFN_wglMakeCurrent)(HDC, HGLRC);
+typedef HGLRC(WINAPI* PFN_wglGetCurrentContext)(void);
+typedef HDC(WINAPI* PFN_wglGetCurrentDC)(void);
+
+static PFN_wglCreateContext p_wglCreateContext;
+static PFN_wglDeleteContext p_wglDeleteContext;
+static PFN_wglGetProcAddress p_wglGetProcAddress;
+static PFN_wglMakeCurrent p_wglMakeCurrent;
+static PFN_wglGetCurrentContext p_wglGetCurrentContext;
+static PFN_wglGetCurrentDC p_wglGetCurrentDC;
+
+typedef HGLRC(WINAPI* PFN_wglCreateContextAttribsARB)(HDC, HGLRC, const int*);
+typedef const char*(WINAPI* PFN_wglGetExtensionsStringARB)(HDC);
+typedef BOOL(WINAPI* PFN_wglSwapIntervalEXT)(int);
+typedef BOOL(WINAPI* PFN_wglGetPixelFormatAttribivARB)(HDC, int, int, UINT, const int*, int*);
+
+// GL function pointers, loaded LWJGL-style: wglGetProcAddress first, then
+// GetProcAddress on the opengl32 module.
+static void* GetGLProc(const char* name) {
+    PROC proc = p_wglGetProcAddress(name);
+    if (!proc) {
+        proc = ::GetProcAddress(g_opengl32, name);
+    }
+    return reinterpret_cast<void*>(proc);
+}
+
+typedef const GLubyte*(WINAPI* PFN_glGetString)(GLenum);
+typedef void(WINAPI* PFN_glGetIntegerv)(GLenum, GLint*);
+typedef void(WINAPI* PFN_glClearColor)(GLfloat, GLfloat, GLfloat, GLfloat);
+typedef void(WINAPI* PFN_glClear)(GLbitfield);
+typedef void(WINAPI* PFN_glViewport)(GLint, GLint, GLsizei, GLsizei);
+typedef GLenum(WINAPI* PFN_glGetError)(void);
+typedef void(WINAPI* PFN_glReadPixels)(GLint, GLint, GLsizei, GLsizei, GLenum, GLenum, void*);
+typedef void(WINAPI* PFN_glFinish)(void);
+typedef GLuint(WINAPI* PFN_glCreateShader)(GLenum);
+typedef void(WINAPI* PFN_glShaderSource)(GLuint, GLsizei, const GLchar* const*, const GLint*);
+typedef void(WINAPI* PFN_glCompileShader)(GLuint);
+typedef void(WINAPI* PFN_glGetShaderiv)(GLuint, GLenum, GLint*);
+typedef void(WINAPI* PFN_glGetShaderInfoLog)(GLuint, GLsizei, GLsizei*, GLchar*);
+typedef GLuint(WINAPI* PFN_glCreateProgram)(void);
+typedef void(WINAPI* PFN_glAttachShader)(GLuint, GLuint);
+typedef void(WINAPI* PFN_glLinkProgram)(GLuint);
+typedef void(WINAPI* PFN_glGetProgramiv)(GLuint, GLenum, GLint*);
+typedef void(WINAPI* PFN_glGetProgramInfoLog)(GLuint, GLsizei, GLsizei*, GLchar*);
+typedef void(WINAPI* PFN_glUseProgram)(GLuint);
+typedef void(WINAPI* PFN_glGenVertexArrays)(GLsizei, GLuint*);
+typedef void(WINAPI* PFN_glBindVertexArray)(GLuint);
+typedef void(WINAPI* PFN_glGenBuffers)(GLsizei, GLuint*);
+typedef void(WINAPI* PFN_glBindBuffer)(GLenum, GLuint);
+typedef void(WINAPI* PFN_glBufferData)(GLenum, GLsizeiptr, const void*, GLenum);
+typedef void(WINAPI* PFN_glVertexAttribPointer)(GLuint, GLint, GLenum, unsigned char, GLsizei, const void*);
+typedef void(WINAPI* PFN_glEnableVertexAttribArray)(GLuint);
+typedef void(WINAPI* PFN_glDrawArrays)(GLenum, GLint, GLsizei);
+
+static PFN_glGetString glGetString_;
+static PFN_glGetIntegerv glGetIntegerv_;
+static PFN_glClearColor glClearColor_;
+static PFN_glClear glClear_;
+static PFN_glViewport glViewport_;
+static PFN_glGetError glGetError_;
+static PFN_glReadPixels glReadPixels_;
+static PFN_glFinish glFinish_;
+static PFN_glCreateShader glCreateShader_;
+static PFN_glShaderSource glShaderSource_;
+static PFN_glCompileShader glCompileShader_;
+static PFN_glGetShaderiv glGetShaderiv_;
+static PFN_glGetShaderInfoLog glGetShaderInfoLog_;
+static PFN_glCreateProgram glCreateProgram_;
+static PFN_glAttachShader glAttachShader_;
+static PFN_glLinkProgram glLinkProgram_;
+static PFN_glGetProgramiv glGetProgramiv_;
+static PFN_glGetProgramInfoLog glGetProgramInfoLog_;
+static PFN_glUseProgram glUseProgram_;
+static PFN_glGenVertexArrays glGenVertexArrays_;
+static PFN_glBindVertexArray glBindVertexArray_;
+static PFN_glGenBuffers glGenBuffers_;
+static PFN_glBindBuffer glBindBuffer_;
+static PFN_glBufferData glBufferData_;
+static PFN_glVertexAttribPointer glVertexAttribPointer_;
+static PFN_glEnableVertexAttribArray glEnableVertexAttribArray_;
+static PFN_glDrawArrays glDrawArrays_;
+
+#define FAIL(...)                                                                                                      \
+    do {                                                                                                               \
+        std::fprintf(stderr, "SMOKE FAIL: " __VA_ARGS__);                                                              \
+        std::fprintf(stderr, "\n");                                                                                    \
+        return 1;                                                                                                      \
+    } while (0)
+
+// CRT-free crash dumper: runs during process teardown when stdio is already
+// dead, so it may only touch raw Win32.
+static void RawPrint(HANDLE f, const char* s) {
+    DWORD written = 0;
+    WriteFile(f, s, (DWORD)lstrlenA(s), &written, nullptr);
+}
+
+static void RawHex(HANDLE f, DWORD64 v) {
+    char buf[20];
+    for (int i = 0; i < 16; ++i) {
+        int nib = (int)((v >> ((15 - i) * 4)) & 0xF);
+        buf[i] = (char)(nib < 10 ? '0' + nib : 'a' + nib - 10);
+    }
+    buf[16] = 0;
+    RawPrint(f, buf);
+}
+
+static LONG WINAPI CrashDump(EXCEPTION_POINTERS* ep) {
+    static volatile LONG entered = 0;
+    if (ep->ExceptionRecord->ExceptionCode != 0xC0000005) return EXCEPTION_CONTINUE_SEARCH;
+    if (InterlockedExchange(&entered, 1)) return EXCEPTION_CONTINUE_SEARCH;
+    HANDLE f = CreateFileA(
+        "C:\\Users\\yello\\AndroidStudioProjects\\FoldCraftLauncher\\MobileGL\\.claude\\worktrees\\wgl-host\\wgl-smoke\\crash-stack.txt",
+        GENERIC_WRITE, 0, nullptr, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, nullptr);
+    if (f == INVALID_HANDLE_VALUE) {
+        TerminateProcess(GetCurrentProcess(), 0x43);
+    }
+    RawPrint(f, "AV at ");
+    RawHex(f, (DWORD64)ep->ExceptionRecord->ExceptionAddress);
+    RawPrint(f, "\n");
+    const DWORD64* sp = reinterpret_cast<const DWORD64*>(ep->ContextRecord->Rsp);
+    for (int i = 0; i < 512; ++i) {
+        MEMORY_BASIC_INFORMATION mbi;
+        if (!VirtualQuery(sp + i, &mbi, sizeof(mbi)) || mbi.State != MEM_COMMIT ||
+            (mbi.Protect & (PAGE_READONLY | PAGE_READWRITE | PAGE_EXECUTE_READ | PAGE_EXECUTE_READWRITE)) == 0)
+            break;
+        DWORD64 v = sp[i];
+        HMODULE mod = nullptr;
+        if (v > 0x10000 &&
+            GetModuleHandleExA(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS | GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT,
+                               reinterpret_cast<LPCSTR>(v), &mod) &&
+            mod) {
+            char name[MAX_PATH];
+            name[0] = 0;
+            GetModuleFileNameA(mod, name, MAX_PATH);
+            RawPrint(f, "stack ");
+            RawHex(f, v);
+            RawPrint(f, " ");
+            RawPrint(f, name);
+            RawPrint(f, "+");
+            RawHex(f, v - (DWORD64)mod);
+            RawPrint(f, "\n");
+        }
+    }
+    CloseHandle(f);
+    TerminateProcess(GetCurrentProcess(), 0x42);
+    return EXCEPTION_EXECUTE_HANDLER;
+}
+
+static LRESULT CALLBACK WndProc(HWND hwnd, UINT msg, WPARAM wp, LPARAM lp) {
+    return DefWindowProcA(hwnd, msg, wp, lp);
+}
+
+static HWND MakeWindow(const char* title, int w, int h, bool visible) {
+    WNDCLASSA wc{};
+    wc.style = CS_OWNDC;
+    wc.lpfnWndProc = WndProc;
+    wc.hInstance = GetModuleHandleA(nullptr);
+    wc.lpszClassName = "MGWGLSmoke";
+    static bool registered = false;
+    if (!registered) {
+        RegisterClassA(&wc);
+        registered = true;
+    }
+    DWORD style = WS_OVERLAPPEDWINDOW | (visible ? WS_VISIBLE : 0);
+    RECT rect{0, 0, w, h};
+    AdjustWindowRect(&rect, WS_OVERLAPPEDWINDOW, FALSE);
+    return CreateWindowExA(0, "MGWGLSmoke", title, style, CW_USEDEFAULT, CW_USEDEFAULT, rect.right - rect.left,
+                           rect.bottom - rect.top, nullptr, nullptr, GetModuleHandleA(nullptr), nullptr);
+}
+
+int main() {
+    AddVectoredExceptionHandler(1, CrashDump);
+    // Load exactly like GLFW: the app-dir opengl32.dll must win the search.
+    g_opengl32 = LoadLibraryA("opengl32.dll");
+    if (!g_opengl32) FAIL("LoadLibraryA(opengl32.dll) failed");
+    char modulePath[MAX_PATH]{};
+    GetModuleFileNameA(g_opengl32, modulePath, MAX_PATH);
+    std::printf("opengl32.dll -> %s\n", modulePath);
+
+    p_wglCreateContext = (PFN_wglCreateContext)::GetProcAddress(g_opengl32, "wglCreateContext");
+    p_wglDeleteContext = (PFN_wglDeleteContext)::GetProcAddress(g_opengl32, "wglDeleteContext");
+    p_wglGetProcAddress = (PFN_wglGetProcAddress)::GetProcAddress(g_opengl32, "wglGetProcAddress");
+    p_wglMakeCurrent = (PFN_wglMakeCurrent)::GetProcAddress(g_opengl32, "wglMakeCurrent");
+    p_wglGetCurrentContext = (PFN_wglGetCurrentContext)::GetProcAddress(g_opengl32, "wglGetCurrentContext");
+    p_wglGetCurrentDC = (PFN_wglGetCurrentDC)::GetProcAddress(g_opengl32, "wglGetCurrentDC");
+    if (!p_wglCreateContext || !p_wglDeleteContext || !p_wglGetProcAddress || !p_wglMakeCurrent ||
+        !p_wglGetCurrentContext || !p_wglGetCurrentDC)
+        FAIL("missing classic wgl exports");
+
+    // ---- Phase 1: GLFW-style dummy context on a hidden helper window ----
+    HWND helper = MakeWindow("helper", 1, 1, false);
+    if (!helper) FAIL("helper window creation failed");
+    HDC helperDC = GetDC(helper);
+
+    PIXELFORMATDESCRIPTOR pfd{};
+    pfd.nSize = sizeof(pfd);
+    pfd.nVersion = 1;
+    pfd.dwFlags = PFD_DRAW_TO_WINDOW | PFD_SUPPORT_OPENGL | PFD_DOUBLEBUFFER;
+    pfd.iPixelType = PFD_TYPE_RGBA;
+    pfd.cColorBits = 24;
+
+    int dummyFormat = ChoosePixelFormat(helperDC, &pfd); // gdi32 -> wglChoosePixelFormat
+    if (dummyFormat <= 0) FAIL("gdi32 ChoosePixelFormat returned %d", dummyFormat);
+    if (!SetPixelFormat(helperDC, dummyFormat, &pfd)) FAIL("gdi32 SetPixelFormat failed");
+    std::printf("dummy pixel format: %d (DescribePixelFormat count=%d)\n", dummyFormat,
+                DescribePixelFormat(helperDC, 1, sizeof(pfd), nullptr));
+
+    HGLRC dummyCtx = p_wglCreateContext(helperDC);
+    if (!dummyCtx) FAIL("wglCreateContext (dummy) failed");
+    if (!p_wglMakeCurrent(helperDC, dummyCtx)) {
+        typedef int(WINAPI * PFN_eglGetError)(void);
+        auto eglGetError_ = (PFN_eglGetError)::GetProcAddress(g_opengl32, "eglGetError");
+        FAIL("wglMakeCurrent (dummy) failed (GetLastError=%lu eglError=0x%x)", GetLastError(),
+             eglGetError_ ? eglGetError_() : -1);
+    }
+
+    auto wglGetExtensionsStringARB =
+        (PFN_wglGetExtensionsStringARB)p_wglGetProcAddress("wglGetExtensionsStringARB");
+    auto wglCreateContextAttribsARB =
+        (PFN_wglCreateContextAttribsARB)p_wglGetProcAddress("wglCreateContextAttribsARB");
+    auto wglSwapIntervalEXT = (PFN_wglSwapIntervalEXT)p_wglGetProcAddress("wglSwapIntervalEXT");
+    auto wglGetPixelFormatAttribivARB =
+        (PFN_wglGetPixelFormatAttribivARB)p_wglGetProcAddress("wglGetPixelFormatAttribivARB");
+    if (!wglGetExtensionsStringARB) FAIL("wglGetExtensionsStringARB unavailable");
+    if (!wglCreateContextAttribsARB) FAIL("wglCreateContextAttribsARB unavailable");
+    if (!wglSwapIntervalEXT) FAIL("wglSwapIntervalEXT unavailable");
+    if (!wglGetPixelFormatAttribivARB) FAIL("wglGetPixelFormatAttribivARB unavailable");
+    std::printf("WGL extensions: %s\n", wglGetExtensionsStringARB(helperDC));
+
+    glGetString_ = (PFN_glGetString)GetGLProc("glGetString");
+    if (!glGetString_) FAIL("glGetString unavailable");
+    std::printf("dummy ctx GL_VERSION: %s\n", glGetString_(GL_VERSION));
+    std::printf("dummy ctx GL_RENDERER: %s\n", glGetString_(GL_RENDERER));
+
+    int nfmtAttrib = WGL_NUMBER_PIXEL_FORMATS_ARB;
+    int nativeCount = 0;
+    if (!wglGetPixelFormatAttribivARB(helperDC, 1, 0, 1, &nfmtAttrib, &nativeCount) || nativeCount < 1)
+        FAIL("WGL_NUMBER_PIXEL_FORMATS_ARB query failed (%d)", nativeCount);
+    std::printf("ARB pixel formats: %d\n", nativeCount);
+
+    if (!p_wglMakeCurrent(nullptr, nullptr)) FAIL("wglMakeCurrent(NULL) failed");
+    if (!p_wglDeleteContext(dummyCtx)) FAIL("wglDeleteContext (dummy) failed");
+
+    // ---- Phase 2: real window + 3.2 core forward-compatible context ----
+    HWND window = MakeWindow("MobileGL WGL smoke", 640, 480, true);
+    if (!window) FAIL("main window creation failed");
+    HDC dc = GetDC(window);
+
+    // Pick the depth24/stencil8 RGBA format via the ARB query, GLFW-style.
+    int chosenFormat = 0;
+    for (int i = 1; i <= nativeCount; ++i) {
+        const int attribs[] = {WGL_SUPPORT_OPENGL_ARB, WGL_DRAW_TO_WINDOW_ARB, WGL_PIXEL_TYPE_ARB,
+                               WGL_DOUBLE_BUFFER_ARB, WGL_DEPTH_BITS_ARB, WGL_STENCIL_BITS_ARB};
+        int values[6] = {};
+        if (!wglGetPixelFormatAttribivARB(dc, i, 0, 6, attribs, values)) continue;
+        if (values[0] && values[1] && values[2] == WGL_TYPE_RGBA_ARB && values[3] && values[4] >= 24 &&
+            values[5] >= 8) {
+            chosenFormat = i;
+            break;
+        }
+    }
+    if (!chosenFormat) FAIL("no ARB pixel format with depth24+stencil8");
+    PIXELFORMATDESCRIPTOR realPfd{};
+    if (!DescribePixelFormat(dc, chosenFormat, sizeof(realPfd), &realPfd))
+        FAIL("DescribePixelFormat(%d) failed", chosenFormat);
+    if (!SetPixelFormat(dc, chosenFormat, &realPfd)) FAIL("SetPixelFormat (real) failed");
+
+    const int ctxAttribs[] = {
+        WGL_CONTEXT_OPENGL_NO_ERROR_ARB, 1,
+        WGL_CONTEXT_MAJOR_VERSION_ARB, 3,
+        WGL_CONTEXT_MINOR_VERSION_ARB, 2,
+        WGL_CONTEXT_FLAGS_ARB, WGL_CONTEXT_FORWARD_COMPATIBLE_BIT_ARB,
+        WGL_CONTEXT_PROFILE_MASK_ARB, WGL_CONTEXT_CORE_PROFILE_BIT_ARB,
+        0,
+    };
+    HGLRC ctx = wglCreateContextAttribsARB(dc, nullptr, ctxAttribs);
+    if (!ctx) FAIL("wglCreateContextAttribsARB failed (err=0x%lx)", GetLastError());
+    if (!p_wglMakeCurrent(dc, ctx)) FAIL("wglMakeCurrent (real) failed");
+    if (p_wglGetCurrentContext() != ctx || p_wglGetCurrentDC() != dc) FAIL("current context/DC mismatch");
+    if (!wglSwapIntervalEXT(0)) FAIL("wglSwapIntervalEXT(0) failed");
+
+#define LOAD(fn)                                                                                                       \
+    fn##_ = (PFN_##fn)GetGLProc(#fn);                                                                                  \
+    if (!fn##_) FAIL("missing GL function " #fn);
+    LOAD(glGetIntegerv)
+    LOAD(glClearColor)
+    LOAD(glClear)
+    LOAD(glViewport)
+    LOAD(glGetError)
+    LOAD(glReadPixels)
+    LOAD(glFinish)
+    LOAD(glCreateShader)
+    LOAD(glShaderSource)
+    LOAD(glCompileShader)
+    LOAD(glGetShaderiv)
+    LOAD(glGetShaderInfoLog)
+    LOAD(glCreateProgram)
+    LOAD(glAttachShader)
+    LOAD(glLinkProgram)
+    LOAD(glGetProgramiv)
+    LOAD(glGetProgramInfoLog)
+    LOAD(glUseProgram)
+    LOAD(glGenVertexArrays)
+    LOAD(glBindVertexArray)
+    LOAD(glGenBuffers)
+    LOAD(glBindBuffer)
+    LOAD(glBufferData)
+    LOAD(glVertexAttribPointer)
+    LOAD(glEnableVertexAttribArray)
+    LOAD(glDrawArrays)
+#undef LOAD
+
+    GLint major = 0, minor = 0;
+    glGetIntegerv_(GL_MAJOR_VERSION, &major);
+    glGetIntegerv_(GL_MINOR_VERSION, &minor);
+    std::printf("real ctx: GL %d.%d | %s | %s\n", major, minor, glGetString_(GL_RENDERER),
+                glGetString_(GL_VERSION));
+    if (major < 3 || (major == 3 && minor < 2)) FAIL("context version below 3.2");
+
+    const char* vsSrc = "#version 150 core\n"
+                        "in vec2 aPos;\n"
+                        "void main() { gl_Position = vec4(aPos, 0.0, 1.0); }\n";
+    const char* fsSrc = "#version 150 core\n"
+                        "out vec4 fragColor;\n"
+                        "void main() { fragColor = vec4(1.0, 0.5, 0.0, 1.0); }\n";
+
+    GLuint vs = glCreateShader_(GL_VERTEX_SHADER);
+    glShaderSource_(vs, 1, &vsSrc, nullptr);
+    glCompileShader_(vs);
+    GLint ok = 0;
+    glGetShaderiv_(vs, GL_COMPILE_STATUS, &ok);
+    if (!ok) {
+        char log[2048];
+        glGetShaderInfoLog_(vs, sizeof(log), nullptr, log);
+        FAIL("vertex shader compile failed: %s", log);
+    }
+    GLuint fs = glCreateShader_(GL_FRAGMENT_SHADER);
+    glShaderSource_(fs, 1, &fsSrc, nullptr);
+    glCompileShader_(fs);
+    glGetShaderiv_(fs, GL_COMPILE_STATUS, &ok);
+    if (!ok) {
+        char log[2048];
+        glGetShaderInfoLog_(fs, sizeof(log), nullptr, log);
+        FAIL("fragment shader compile failed: %s", log);
+    }
+    GLuint prog = glCreateProgram_();
+    glAttachShader_(prog, vs);
+    glAttachShader_(prog, fs);
+    glLinkProgram_(prog);
+    glGetProgramiv_(prog, GL_LINK_STATUS, &ok);
+    if (!ok) {
+        char log[2048];
+        glGetProgramInfoLog_(prog, sizeof(log), nullptr, log);
+        FAIL("program link failed: %s", log);
+    }
+
+    const GLfloat verts[] = {-0.8f, -0.8f, 0.8f, -0.8f, 0.0f, 0.8f};
+    GLuint vao = 0, vbo = 0;
+    glGenVertexArrays_(1, &vao);
+    glBindVertexArray_(vao);
+    glGenBuffers_(1, &vbo);
+    glBindBuffer_(GL_ARRAY_BUFFER, vbo);
+    glBufferData_(GL_ARRAY_BUFFER, sizeof(verts), verts, GL_STATIC_DRAW);
+    glVertexAttribPointer_(0, 2, GL_FLOAT, GL_FALSE, 0, nullptr);
+    glEnableVertexAttribArray_(0);
+
+    glViewport_(0, 0, 640, 480);
+    glUseProgram_(prog);
+
+    unsigned char centerPixel[4] = {}, cornerPixel[4] = {};
+    for (int frame = 0; frame < 60; ++frame) {
+        MSG msg;
+        while (PeekMessageA(&msg, nullptr, 0, 0, PM_REMOVE)) {
+            TranslateMessage(&msg);
+            DispatchMessageA(&msg);
+        }
+        glClearColor_(0.1f, 0.2f, 0.4f, 1.0f);
+        glClear_(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
+        glDrawArrays_(GL_TRIANGLES, 0, 3);
+        if (frame == 58) {
+            glFinish_();
+            glReadPixels_(320, 240, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, centerPixel);
+            glReadPixels_(5, 5, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, cornerPixel);
+        }
+        if (!SwapBuffers(dc)) FAIL("gdi32 SwapBuffers failed at frame %d", frame); // gdi32 -> wglSwapBuffers
+    }
+
+    GLenum err = glGetError_();
+    std::printf("center pixel: %u,%u,%u,%u  corner pixel: %u,%u,%u,%u  glGetError=0x%x\n", centerPixel[0],
+                centerPixel[1], centerPixel[2], centerPixel[3], cornerPixel[0], cornerPixel[1], cornerPixel[2],
+                cornerPixel[3], err);
+
+    if (err != GL_NO_ERROR) FAIL("GL error 0x%x", err);
+    // Orange triangle at center (255,127±3,0), clear color at corner (~25,51,102).
+    if (!(centerPixel[0] > 200 && centerPixel[1] > 90 && centerPixel[1] < 160 && centerPixel[2] < 50))
+        FAIL("center pixel is not the triangle color");
+    if (!(cornerPixel[2] > 60 && cornerPixel[0] < 60)) FAIL("corner pixel is not the clear color");
+
+    if (!p_wglMakeCurrent(nullptr, nullptr)) FAIL("final wglMakeCurrent(NULL) failed");
+    if (!p_wglDeleteContext(ctx)) FAIL("final wglDeleteContext failed");
+
+    std::printf("SMOKE PASS\n");
+    return 0;
+}