[Fix, Test] (MG_Backend/DirectGLES): hand every compiled shader's lifetime to its program - glDeleteShader was never called, so each program build leaked its driver shader objects

MobileGL/MG_Backend/DirectGLES/Managers.cpp

@@ -4386,11 +4386,28 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     log.back() = '\0';
                     MGLOG_E("Shader compilation failed for backend ID %u: %s", backendShaderId, log.data());
                     m_backendProgramUsable = false;
+                    // Nothing will ever attach this one, so nothing else can free it.
+                    g_GLESFuncs.glDeleteShader(backendShaderId);
                     continue;
                 }
 
                 MGLOG_D("Attaching shader ID: %u to program %u", backendShaderId, m_backendProgramId);
                 g_GLESFuncs.glAttachShader(m_backendProgramId, backendShaderId);
+                // Hand the shader's lifetime to the program, immediately and unconditionally.
+                //
+                // glDeleteShader only FLAGS a shader; the driver frees it when it is attached to
+                // nothing. Flagging it here is what makes the program own it, so deleting the
+                // program (or the detach loop above, on a relink) is what actually frees it.
+                // Without this call every program build leaked its shader objects for the process
+                // lifetime, and a relink leaked them twice - the detach loop above dropped the
+                // program's reference to shaders nothing had flagged, so they became unreachable
+                // AND undeletable. The GL swizzle conformance test builds 1,296 programs per case,
+                // so a handful of cases left tens of thousands of live driver shaders behind and
+                // the driver started mis-serving them (KHR-GL33/GL40.texture_swizzle.smoke_*).
+                // Same class of defect as the missing framebuffer/renderbuffer/sampler destructors
+                // fixed in Wave 1, and the last of that family: this is the one backend GL object
+                // MobileGL creates without an owning wrapper to destroy it.
+                g_GLESFuncs.glDeleteShader(backendShaderId);
 
                 MGLOG_D("Processed shader source length: %zu", source.length());
             }

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -2578,6 +2578,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     }
                     entry.storageBlockNameByBinding[binding] = uniformName;
                     entry.storageBlockIndexByBinding[binding] = static_cast<Int>(blockIndex);
+
+                    // A block INSTANCE array is ONE Vulkan binding carrying `count`
+                    // descriptors, while GL assigns its elements consecutive binding points
+                    // starting at the declared one (GL 4.6 core 7.8). Recording only element 0 -
+                    // which is all this used to do - left the layout claiming descriptorCount 1,
+                    // so every element past the first read a descriptor nobody wrote and
+                    // `b[1].data.length()` answered from an unconstrained buffer instead of its
+                    // own bound range (KHR-GL43.shader_storage_buffer_object.-
+                    // advanced-unsizedArrayLength-*).
+                    entry.bindingDescriptorCounts[binding] = static_cast<Uint16>(std::max<Uint32>(1u, sampler->count));
                     continue;
                 }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -677,7 +677,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool UniformManager::ResolveStorageBufferDescriptor(const MG_State::GLState::ProgramObject& program,
                                                         const ProgramFactory::VkProgramObject& programObj,
-                                                        Uint32 binding,
+                                                        Uint32 binding, Uint32 element,
                                                         VkDescriptorBufferInfo& outBufferInfo) const {
         outBufferInfo = {};
         MOBILEGL_ASSERT(m_bufferManager != nullptr, "ResolveStorageBufferDescriptor: buffer manager is null");
@@ -688,8 +688,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const Int blockIndex = programObj.storageBlockIndexByBinding[binding];
         MOBILEGL_ASSERT(blockIndex >= 0, "ResolveStorageBufferDescriptor: no SSBO block mapped to binding %u",
                         binding);
+        // A block instance array declares one block whose elements take consecutive GL binding
+        // points from the declared one (GL 4.6 core 7.8), and the reflection collapses the whole
+        // array to that one block - so the element index IS the offset from its binding.
         const GLuint frontendBinding =
-            GetShaderStorageBlockBinding(program, static_cast<GLuint>(blockIndex));
+            GetShaderStorageBlockBinding(program, static_cast<GLuint>(blockIndex)) + element;
         const Uint32 bindingPointCount =
             static_cast<Uint32>(MG_State::pGLContext->GetBufferBindingPointCount(BufferTarget::ShaderStorage));
         MOBILEGL_ASSERT(frontendBinding < bindingPointCount,
@@ -1416,12 +1419,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         dynamicOffsets.clear();
         // Arrayed UBO bindings contribute extra buffer infos and dynamic offsets; reserve for
         // the worst case so the pBufferInfo pointers taken below never dangle on reallocation.
+        // Arrayed SSBO bindings contribute extra buffer infos too (but no dynamic offsets).
         Uint32 uboArrayExtra = 0;
         for (const auto& arrayEntry : programObj.arrayedUniformBlockIndicesByBinding) {
             uboArrayExtra += static_cast<Uint32>(arrayEntry.second.size()) - 1u;
         }
+        Uint32 ssboArrayExtra = 0;
+        for (const Uint16 count : programObj.bindingDescriptorCounts) {
+            if (count > 1) ssboArrayExtra += static_cast<Uint32>(count) - 1u;
+        }
         writes.reserve(m_maxBindings);
-        bufferInfos.reserve(m_maxBindings + uboArrayExtra);
+        bufferInfos.reserve(m_maxBindings + uboArrayExtra + ssboArrayExtra);
         imageInfos.reserve(m_maxBindings);
         texelBufferViews.reserve(m_maxBindings);
         dynamicOffsets.reserve(programObj.dynamicBindings.size() + uboArrayExtra);
@@ -1493,18 +1501,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 write.pTexelBufferView = &texelBufferViews.back();
                 writes.push_back(write);
             } else if (kind == ProgramFactory::DescriptorBindingKind::StorageBuffer) {
-                VkDescriptorBufferInfo bufferInfo{};
-                if (!ResolveStorageBufferDescriptor(program, programObj, binding, bufferInfo)) {
-                    MGLOG_E(
-                        "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage buffer binding %u has no valid descriptor",
-                        binding);
-                    return false;
+                // One write per binding, but `descriptorCount` buffer infos: a GLSL block
+                // instance array occupies a single binding whose elements each come from their
+                // own GL binding point.
+                const Uint32 descriptorCount =
+                    binding < programObj.bindingDescriptorCounts.size()
+                        ? std::max<Uint32>(1, programObj.bindingDescriptorCounts[binding])
+                        : 1u;
+                const SizeT firstBufferInfoIndex = bufferInfos.size();
+                for (Uint32 element = 0; element < descriptorCount; ++element) {
+                    VkDescriptorBufferInfo bufferInfo{};
+                    if (!ResolveStorageBufferDescriptor(program, programObj, binding, element, bufferInfo)) {
+                        MGLOG_E(
+                            "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage buffer binding %u "
+                            "element %u has no valid descriptor",
+                            binding, element);
+                        return false;
+                    }
+                    bufferInfos.push_back(bufferInfo);
                 }
 
-                bufferInfos.push_back(bufferInfo);
                 fastRebindKindsEligible = false;
                 write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
-                write.pBufferInfo = &bufferInfos.back();
+                write.descriptorCount = descriptorCount;
+                write.pBufferInfo = &bufferInfos[firstBufferInfoIndex];
                 writes.push_back(write);
             } else if (kind == ProgramFactory::DescriptorBindingKind::StorageImage) {
                 VkDescriptorImageInfo imageInfo{};

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -166,9 +166,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool ResolveTexelBufferDescriptor(const MG_State::GLState::ProgramObject& program,
                                           const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                           Uint32 frameIndex, VkBufferView& outBufferView);
+        // `element` indexes a block INSTANCE array's descriptors; it is 0 for every ordinary
+        // block. Each element resolves through its own GL storage block, and so its own GL
+        // binding point, buffer and glBindBufferRange window.
         Bool ResolveStorageBufferDescriptor(const MG_State::GLState::ProgramObject& program,
                                             const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
-                                            VkDescriptorBufferInfo& outBufferInfo) const;
+                                            Uint32 element, VkDescriptorBufferInfo& outBufferInfo) const;
         Bool ResolveStorageImageDescriptor(VkCommandBuffer commandBuffer,
                                            const MG_State::GLState::ProgramObject& program,
                                            const ProgramFactory::VkProgramObject& programObj, Uint32 binding,

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -3016,6 +3016,7 @@ void main() {
         DestroySubmitFencePool();
 
         DestroyDeferredDepthMipmapCleanup();
+        DestroyMultisampleResolveScratchImage();
         DestroyComputePipelines();
 
         // No sweep runs during teardown, but the observers point at this renderer
@@ -7194,6 +7195,158 @@ void main() {
         return true;
     }
 
+    void VulkanRenderer::DestroyMultisampleResolveScratchImage() {
+        if (m_msResolveScratch.image != VK_NULL_HANDLE) {
+            vmaDestroyImage(m_allocator, m_msResolveScratch.image, m_msResolveScratch.allocation);
+        }
+        m_msResolveScratch = {};
+    }
+
+    Bool VulkanRenderer::AcquireMultisampleResolveScratchImage(VkCommandBuffer commandBuffer, VkFormat format,
+                                                               VkExtent2D extent) {
+        if (extent.width == 0 || extent.height == 0 || format == VK_FORMAT_UNDEFINED) {
+            return false;
+        }
+        // Grow-only, and never shrink: these blits repeat at one or two sizes, so the steady state
+        // is one allocation for the whole process.
+        if (m_msResolveScratch.image == VK_NULL_HANDLE || m_msResolveScratch.format != format ||
+            m_msResolveScratch.extent.width < extent.width || m_msResolveScratch.extent.height < extent.height) {
+            const VkExtent2D grown = {std::max(extent.width, m_msResolveScratch.extent.width),
+                                      std::max(extent.height, m_msResolveScratch.extent.height)};
+            DestroyMultisampleResolveScratchImage();
+
+            VkImageCreateInfo imageInfo{};
+            imageInfo.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
+            imageInfo.imageType = VK_IMAGE_TYPE_2D;
+            imageInfo.format = format;
+            imageInfo.extent = {grown.width, grown.height, 1};
+            imageInfo.mipLevels = 1;
+            imageInfo.arrayLayers = 1;
+            imageInfo.samples = VK_SAMPLE_COUNT_1_BIT;
+            imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
+            imageInfo.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+            imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+            imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+
+            VmaAllocationCreateInfo allocationInfo{};
+            allocationInfo.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
+            allocationInfo.requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+            if (vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &m_msResolveScratch.image,
+                               &m_msResolveScratch.allocation, nullptr) != VK_SUCCESS) {
+                // Soft failure: the caller keeps the direct resolve, which is what shipped before.
+                MGLOG_E("AcquireMultisampleResolveScratchImage: vmaCreateImage failed (format=%d %ux%u)",
+                        static_cast<Int>(format), grown.width, grown.height);
+                m_msResolveScratch = {};
+                return false;
+            }
+            m_msResolveScratch.format = format;
+            m_msResolveScratch.extent = grown;
+            m_msResolveScratch.layout = VK_IMAGE_LAYOUT_UNDEFINED;
+        }
+
+        VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+        VkAccessFlags srcAccessMask = 0;
+        GetImageTransitionSourceState(m_msResolveScratch.layout, srcStageMask, srcAccessMask);
+        if (!VkTextureManager::TransitionImageLayout(commandBuffer, m_msResolveScratch.image,
+                                                     m_msResolveScratch.layout,
+                                                     VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, srcStageMask,
+                                                     VK_PIPELINE_STAGE_TRANSFER_BIT, srcAccessMask,
+                                                     VK_ACCESS_TRANSFER_WRITE_BIT, VK_IMAGE_ASPECT_COLOR_BIT)) {
+            return false;
+        }
+        return true;
+    }
+
+    // The aspects a depth/stencil format actually carries. VkTextureManager keeps its own copy of
+    // this private, and the swapchain's depth/stencil image has no TextureResource to ask.
+    static VkImageAspectFlags GetDepthStencilAspectMaskForFormat(VkFormat format) {
+        switch (format) {
+        case VK_FORMAT_D16_UNORM:
+        case VK_FORMAT_X8_D24_UNORM_PACK32:
+        case VK_FORMAT_D32_SFLOAT:
+            return VK_IMAGE_ASPECT_DEPTH_BIT;
+        case VK_FORMAT_D16_UNORM_S8_UINT:
+        case VK_FORMAT_D24_UNORM_S8_UINT:
+        case VK_FORMAT_D32_SFLOAT_S8_UINT:
+            return VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;
+        case VK_FORMAT_S8_UINT:
+            return VK_IMAGE_ASPECT_STENCIL_BIT;
+        default:
+            return VK_IMAGE_ASPECT_NONE;
+        }
+    }
+
+    // The depth/stencil half of MaterializePendingClearForDefaultFramebuffer. Separate only
+    // because the image, the aspects and the clear value are all different from the colour one;
+    // the reason it exists is the same - a readback with no intervening draw has no render pass
+    // to fold the parked clear into.
+    Bool VulkanRenderer::MaterializePendingDepthStencilClearForDefaultFramebuffer(
+        VkCommandBuffer commandBuffer, const MG_State::GLState::FramebufferAttachmentObject& attachment,
+        const ClearAttachmentPayload& payload) {
+        const VkImage depthStencilImage = m_swapchainObject.GetDepthStencilImage(m_imageIndexAcquired);
+        if (depthStencilImage == VK_NULL_HANDLE) {
+            return false;
+        }
+        const VkImageAspectFlags imageAspects =
+            GetDepthStencilAspectMaskForFormat(m_swapchainObject.GetDepthStencilFormat());
+        VkImageAspectFlags clearAspects = 0;
+        if ((payload.mask & GL_DEPTH_BUFFER_BIT) != 0) clearAspects |= (imageAspects & VK_IMAGE_ASPECT_DEPTH_BIT);
+        if ((payload.mask & GL_STENCIL_BUFFER_BIT) != 0) clearAspects |= (imageAspects & VK_IMAGE_ASPECT_STENCIL_BIT);
+        if (clearAspects == 0) {
+            // Nothing this image can express; drop the pending clear rather than leave it to a
+            // later render pass that would load it against an aspect that does not exist.
+            m_clearManager->PopPendingClear(attachment);
+            return true;
+        }
+
+        VkImageLayout currentLayout = m_swapchainObject.GetDepthStencilImageLayout(m_imageIndexAcquired);
+        VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+        VkAccessFlags srcAccessMask = 0;
+        GetImageTransitionSourceState(currentLayout, srcStageMask, srcAccessMask);
+        VkImageLayout clearLayout = currentLayout;
+        if (!VkTextureManager::TransitionImageLayout(commandBuffer, depthStencilImage, clearLayout,
+                                                     VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, srcStageMask,
+                                                     VK_PIPELINE_STAGE_TRANSFER_BIT, srcAccessMask,
+                                                     VK_ACCESS_TRANSFER_WRITE_BIT, imageAspects)) {
+            return false;
+        }
+
+        VkClearDepthStencilValue clearValue{};
+        clearValue.depth = payload.depth;
+        clearValue.stencil = payload.stencil;
+        VkImageSubresourceRange range{};
+        range.aspectMask = clearAspects;
+        range.baseMipLevel = 0;
+        range.levelCount = 1;
+        range.baseArrayLayer = 0;
+        range.layerCount = 1;
+        vkCmdClearDepthStencilImage(commandBuffer, depthStencilImage, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, &clearValue,
+                                    1, &range);
+
+        VkImageLayout settledLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+        VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+        VkAccessFlags dstAccessMask = 0;
+        GetImageTransitionDestinationState(VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL, dstStageMask,
+                                           dstAccessMask);
+        if (!VkTextureManager::TransitionImageLayout(commandBuffer, depthStencilImage, settledLayout,
+                                                     VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
+                                                     VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask,
+                                                     VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, imageAspects)) {
+            return false;
+        }
+        m_swapchainObject.SetDepthStencilImageLayout(m_imageIndexAcquired,
+                                                     VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL);
+        // The image now holds real values, so the next render pass must LOAD them rather than
+        // treat the attachment as undefined and discard the clear that just executed.
+        m_swapchainObject.SetDepthStencilContentDefined(m_imageIndexAcquired, true);
+
+        m_clearManager->PopPendingClear(attachment);
+        MGLOG_D("MaterializePendingClearForDefaultFramebuffer: swapchain depth/stencil image %u pending clear "
+                "materialized (aspects=0x%x)",
+                m_imageIndexAcquired, static_cast<Uint32>(clearAspects));
+        return true;
+    }
+
     // A glClear on the DEFAULT framebuffer is parked as a pending clear and folded into the next
     // render pass's loadOp. With no draw in between there is no render pass, so a readback that
     // followed such a clear blitted the untouched swapchain image and returned the PREVIOUS
@@ -7217,15 +7370,14 @@ void main() {
         if (!m_clearManager->GetPendingClear(attachment, payload)) {
             return true;
         }
-        if ((payload.mask & GL_COLOR_BUFFER_BIT) == 0) {
-            // Depth/stencil on the default framebuffer keeps the loadOp route; the readback
-            // path for it declines default framebuffers outright (ReadDepthStencilPixels).
-            return true;
-        }
         MOBILEGL_ASSERT(VkRenderPassManager::GetActiveRenderPass() == nullptr ||
                             commandBuffer != m_frameContext.GetCurrent().commandBuffer,
                         "MaterializePendingClearForDefaultFramebuffer requires no active render pass");
 
+        if ((payload.mask & GL_COLOR_BUFFER_BIT) == 0) {
+            return MaterializePendingDepthStencilClearForDefaultFramebuffer(commandBuffer, attachment, payload);
+        }
+
         const VkImage swapchainImage = m_swapchainObject.GetImage(m_imageIndexAcquired);
         if (swapchainImage == VK_NULL_HANDLE) {
             return false;
@@ -7570,12 +7722,21 @@ void main() {
                 }
             }
 
-            if (!drawIsDefaultFbo) {
+            const auto destAttachmentType =
+                ResolveFramebufferCopyAttachmentType(*drawFbo, false, dstBinding.aspectMask);
+            if (drawIsDefaultFbo) {
+                // Same ordering rule for the default framebuffer's depth/stencil - see the
+                // colour twin below.
+                const Bool dstClearReady = MaterializePendingClearForDefaultFramebuffer(
+                    frame.commandBuffer, *drawFbo, destAttachmentType);
+                MOBILEGL_ASSERT(dstClearReady,
+                                "BlitFramebuffer: failed to materialize the default framebuffer's pending "
+                                "depth/stencil clear");
+            } else {
                 // A clear queued for the destination predates this blit in API order;
                 // execute it now, or its deferred materialization would later stomp the
                 // copied contents (MC 26.3 OIT clears cloud_depth, then blits the main
                 // depth into it - the stale loadOp=CLEAR erased the copy).
-                const auto destAttachmentType = ResolveFramebufferCopyAttachmentType(*drawFbo, false, dstBinding.aspectMask);
                 const auto& destAttachment = drawFbo->GetAttachment(destAttachmentType);
                 if (auto destTexture = destAttachment.GetTexture(); destTexture != nullptr) {
                     const Bool dstClearReady = MaterializePendingClearForTexture(frame.commandBuffer, *destTexture);
@@ -7668,7 +7829,15 @@ void main() {
                 MOBILEGL_ASSERT(ok, "%s: failed to transition depth destination image", __func__);
             }
 
-            if (depthBlitScales) {
+            // The default framebuffer is stored display-side-up, so a rect aimed at it (or read
+            // from it) has to be converted out of GL's bottom-origin space - the same conversion
+            // the colour blit below applies. vkCmdCopyImage cannot express it (it has no second
+            // offset to invert), so a default-framebuffer side forces the vkCmdBlitImage form even
+            // at equal size. Without this a scissored depth blit into the default framebuffer
+            // wrote the MIRRORED band: KHR-GL*.framebuffer_blit.scissor_blit clips to the lower
+            // left quadrant, and the depth landed in the upper one.
+            const Bool depthBlitNeedsOrientation = readIsDefaultFbo || drawIsDefaultFbo;
+            if (depthBlitScales || depthBlitNeedsOrientation) {
                 // vkCmdCopyImage cannot resize; NEAREST is the only filter Vulkan allows for a
                 // depth/stencil blit anyway, and the GL front end already rejects the others.
                 VkImageBlit blitRegion{};
@@ -7684,6 +7853,14 @@ void main() {
                 blitRegion.dstSubresource.layerCount = dstBinding.layerCount;
                 blitRegion.dstOffsets[0] = {dstX0, dstY0, 0};
                 blitRegion.dstOffsets[1] = {dstX1, dstY1, 1};
+                if (readIsDefaultFbo) {
+                    ApplyNativeBlitDefaultFramebufferSourceTransform(m_swapchainObject.GetPreTransform(), srcBinding,
+                                                                     blitRegion);
+                }
+                if (drawIsDefaultFbo) {
+                    ApplyNativeBlitDefaultFramebufferTransform(m_swapchainObject.GetPreTransform(), dstBinding,
+                                                               blitRegion);
+                }
                 vkCmdBlitImage(frame.commandBuffer,
                                srcBinding.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                                dstBinding.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
@@ -7780,7 +7957,19 @@ void main() {
             }
         }
 
-        if (!drawIsDefaultFbo) {
+        if (drawIsDefaultFbo) {
+            // The default framebuffer needs the same ordering, and needed it before anything
+            // consumed its parked clear: Minecraft clears the default framebuffer, renders the
+            // world into its own framebuffer and BLITS the result out, so nothing between the
+            // clear and the blit ever opens a render pass on the default framebuffer to fold the
+            // clear in as a loadOp. The clear therefore stayed pending across the whole frame,
+            // and the first path that did materialize it - the readback - executed it AFTER the
+            // blit and handed back a blank frame (every DirectVulkan retrace, ssim 0.000005).
+            const Bool dstClearReady = MaterializePendingClearForDefaultFramebuffer(
+                frame.commandBuffer, *drawFbo, drawFbo->GetDrawBuffers()[0]);
+            MOBILEGL_ASSERT(dstClearReady,
+                            "BlitFramebuffer: failed to materialize the default framebuffer's pending clear");
+        } else {
             // A clear queued for the destination predates this blit in API order; execute
             // it now, or its deferred materialization would later stomp the blitted color.
             const auto& destAttachment = drawFbo->GetAttachment(drawFbo->GetDrawBuffers()[0]);
@@ -7878,23 +8067,78 @@ void main() {
         }
 
         if (srcBinding.sampleCount != VK_SAMPLE_COUNT_1_BIT && dstBinding.sampleCount == VK_SAMPLE_COUNT_1_BIT) {
-            // NOTE: vkCmdResolveImage cannot flip, and this region is still built from the raw GL
-            // offsets. A multisample-resolve blit whose source or destination is the default
-            // framebuffer therefore keeps the pre-fix behaviour; it needs a resolve-then-blit
-            // (or blit-then-resolve) split, which is its own change.
             // GL multisample resolve blits are 1:1 by spec; vkCmdBlitImage cannot read a
-            // multisampled source.
+            // multisampled source, so the samples have to come down through vkCmdResolveImage.
+            const Uint32 resolveWidth = static_cast<Uint32>(std::abs(srcX1 - srcX0));
+            const Uint32 resolveHeight = static_cast<Uint32>(std::abs(srcY1 - srcY0));
+
+            // vkCmdResolveImage takes ONE offset per side, so it cannot express the axis inversion
+            // that a default-framebuffer rect needs - it would land the mirrored band. When the
+            // transforms above actually moved the region, split the operation: resolve into a
+            // single-sample scratch image at raw offsets, then blit THAT into the destination with
+            // the (already transformed) region, which vkCmdBlitImage can invert.
+            const Bool regionWasTransformed =
+                (readIsDefaultFbo || drawIsDefaultFbo) &&
+                (blitRegion.srcOffsets[0].x != srcX0 || blitRegion.srcOffsets[0].y != srcY0 ||
+                 blitRegion.srcOffsets[1].x != srcX1 || blitRegion.srcOffsets[1].y != srcY1 ||
+                 blitRegion.dstOffsets[0].x != dstX0 || blitRegion.dstOffsets[0].y != dstY0 ||
+                 blitRegion.dstOffsets[1].x != dstX1 || blitRegion.dstOffsets[1].y != dstY1);
+            const Bool useScratchResolve =
+                regionWasTransformed && resolveWidth > 0 && resolveHeight > 0 &&
+                AcquireMultisampleResolveScratchImage(frame.commandBuffer, srcBinding.format,
+                                                      {resolveWidth, resolveHeight});
+
             VkImageResolve resolveRegion{};
             resolveRegion.srcSubresource = blitRegion.srcSubresource;
-            resolveRegion.srcOffset = {std::min(srcX0, srcX1), std::min(srcY0, srcY1), 0};
             resolveRegion.dstSubresource = blitRegion.dstSubresource;
-            resolveRegion.dstOffset = {std::min(dstX0, dstX1), std::min(dstY0, dstY1), 0};
-            resolveRegion.extent = {static_cast<Uint32>(std::abs(srcX1 - srcX0)),
-                                    static_cast<Uint32>(std::abs(srcY1 - srcY0)), 1};
-            vkCmdResolveImage(frame.commandBuffer,
-                              srcBinding.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
-                              dstBinding.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
-                              1, &resolveRegion);
+            resolveRegion.extent = {resolveWidth, resolveHeight, 1};
+            if (useScratchResolve) {
+                // The scratch copy is a plain single-layer colour image, and the resolve reads the
+                // SOURCE band the (possibly inverted) transformed region names - taking its min so
+                // an inverted pair still describes the same band.
+                resolveRegion.srcOffset = {std::min(blitRegion.srcOffsets[0].x, blitRegion.srcOffsets[1].x),
+                                           std::min(blitRegion.srcOffsets[0].y, blitRegion.srcOffsets[1].y), 0};
+                resolveRegion.dstSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+                resolveRegion.dstSubresource.mipLevel = 0;
+                resolveRegion.dstSubresource.baseArrayLayer = 0;
+                resolveRegion.dstSubresource.layerCount = 1;
+                resolveRegion.dstOffset = {0, 0, 0};
+                vkCmdResolveImage(frame.commandBuffer,
+                                  srcBinding.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                                  m_msResolveScratch.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                                  1, &resolveRegion);
+
+                VkImageLayout scratchLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+                const Bool scratchReady = VkTextureManager::TransitionImageLayout(
+                    frame.commandBuffer, m_msResolveScratch.image, scratchLayout,
+                    VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                    VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_TRANSFER_READ_BIT,
+                    VK_IMAGE_ASPECT_COLOR_BIT);
+                MOBILEGL_ASSERT(scratchReady, "%s: failed to transition the resolve scratch image", __func__);
+                m_msResolveScratch.layout = scratchLayout;
+
+                // Second leg: the scratch image holds the resolved band at its own origin, so the
+                // source side of the region becomes the whole scratch rect and only the
+                // destination keeps the transform.
+                VkImageBlit scratchBlit = blitRegion;
+                scratchBlit.srcSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+                scratchBlit.srcSubresource.mipLevel = 0;
+                scratchBlit.srcSubresource.baseArrayLayer = 0;
+                scratchBlit.srcSubresource.layerCount = 1;
+                scratchBlit.srcOffsets[0] = {0, 0, 0};
+                scratchBlit.srcOffsets[1] = {static_cast<Int32>(resolveWidth), static_cast<Int32>(resolveHeight), 1};
+                vkCmdBlitImage(frame.commandBuffer,
+                               m_msResolveScratch.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                               dstBinding.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                               1, &scratchBlit, filter == GL_LINEAR ? VK_FILTER_LINEAR : VK_FILTER_NEAREST);
+            } else {
+                resolveRegion.srcOffset = {std::min(srcX0, srcX1), std::min(srcY0, srcY1), 0};
+                resolveRegion.dstOffset = {std::min(dstX0, dstX1), std::min(dstY0, dstY1), 0};
+                vkCmdResolveImage(frame.commandBuffer,
+                                  srcBinding.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                                  dstBinding.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                                  1, &resolveRegion);
+            }
         } else {
             vkCmdBlitImage(frame.commandBuffer,
                            srcBinding.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
@@ -8729,10 +8973,6 @@ void main() {
     void VulkanRenderer::ReadDepthStencilPixels(MG_State::GLState::FramebufferObject& readFbo, GLint x, GLint y,
                                                 GLsizei width, GLsizei height, GLenum format, GLenum type,
                                                 void* pixels) {
-        if (readFbo.IsDefaultFramebuffer()) {
-            MGLOG_E("DirectVulkan::ReadDepthStencilPixels skipped: default framebuffer readback is unsupported");
-            return;
-        }
         if (width <= 0 || height <= 0) {
             return;
         }
@@ -8743,10 +8983,13 @@ void main() {
         // framebuffers lacking either, so resolving via the depth attachment is enough.
         const auto attachmentType = wantDepth ? MobileGL::FramebufferAttachmentType::Depth
                                               : MobileGL::FramebufferAttachmentType::Stencil;
-        const auto& attachment = readFbo.GetAttachment(attachmentType);
-        if (!attachment.IsValid() || attachment.IsEmpty()) {
-            MGLOG_E("DirectVulkan::ReadDepthStencilPixels skipped: no depth/stencil attachment image");
-            return;
+        const Bool readIsDefaultFbo = readFbo.IsDefaultFramebuffer();
+        if (!readIsDefaultFbo) {
+            const auto& attachment = readFbo.GetAttachment(attachmentType);
+            if (!attachment.IsValid() || attachment.IsEmpty()) {
+                MGLOG_E("DirectVulkan::ReadDepthStencilPixels skipped: no depth/stencil attachment image");
+                return;
+            }
         }
 
         auto& frame = m_frameContext.GetCurrent();
@@ -8757,6 +9000,47 @@ void main() {
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
         }
 
+        // The default framebuffer's depth/stencil lives in the swapchain, not in an
+        // attachment object: its placeholder ITextureObject describes the format but backs no
+        // image, so the branches below would have synced (and read back) an unrelated one.
+        // Declining outright is what made every glReadPixels(GL_DEPTH_COMPONENT/
+        // GL_STENCIL_INDEX) of the default framebuffer leave the caller's buffer untouched -
+        // the whole KHR-GL*.framebuffer_blit family checks exactly that before it blits.
+        if (readIsDefaultFbo) {
+            const VkImage swapchainDepthImage = m_swapchainObject.GetDepthStencilImage(m_imageIndexAcquired);
+            if (swapchainDepthImage == VK_NULL_HANDLE) {
+                MGLOG_E("DirectVulkan::ReadDepthStencilPixels skipped: the default framebuffer has no "
+                        "depth/stencil image");
+                return;
+            }
+            // Per aspect, because the default framebuffer carries a SEPARATE placeholder
+            // attachment for depth and for stencil (MG_Impl/Init.cpp) and each parks its own
+            // pending clear; materializing only one would read the other back un-cleared.
+            if (wantDepth) {
+                const Bool clearReady = MaterializePendingClearForDefaultFramebuffer(
+                    frame.commandBuffer, readFbo, MobileGL::FramebufferAttachmentType::Depth);
+                MOBILEGL_ASSERT(clearReady,
+                                "ReadDepthStencilPixels: failed to materialize the default framebuffer's pending "
+                                "depth clear");
+            }
+            if (wantStencil) {
+                const Bool clearReady = MaterializePendingClearForDefaultFramebuffer(
+                    frame.commandBuffer, readFbo, MobileGL::FramebufferAttachmentType::Stencil);
+                MOBILEGL_ASSERT(clearReady,
+                                "ReadDepthStencilPixels: failed to materialize the default framebuffer's pending "
+                                "stencil clear");
+            }
+            const VkFormat swapchainDepthFormat = m_swapchainObject.GetDepthStencilFormat();
+            VkImageLayout trackedLayout = m_swapchainObject.GetDepthStencilImageLayout(m_imageIndexAcquired);
+            ReadDepthStencilImageToClient(swapchainDepthImage, swapchainDepthFormat, &trackedLayout,
+                                          GetDepthStencilAspectMaskForFormat(swapchainDepthFormat), 0, 0, x, y,
+                                          width, height, format, type, pixels,
+                                          /*defaultFramebufferOrientation=*/true);
+            m_swapchainObject.SetDepthStencilImageLayout(m_imageIndexAcquired, trackedLayout);
+            return;
+        }
+
+        const auto& attachment = readFbo.GetAttachment(attachmentType);
         VkImage image = VK_NULL_HANDLE;
         VkFormat vkFormat = VK_FORMAT_UNDEFINED;
         VkImageLayout* trackedLayout = nullptr;
@@ -8807,7 +9091,8 @@ void main() {
     void VulkanRenderer::ReadDepthStencilImageToClient(VkImage image, VkFormat vkFormat, VkImageLayout* trackedLayout,
                                                        VkImageAspectFlags imageAspect, Uint32 mipLevel,
                                                        Uint32 baseArrayLayer, GLint x, GLint y, GLsizei width,
-                                                       GLsizei height, GLenum format, GLenum type, void* pixels) {
+                                                       GLsizei height, GLenum format, GLenum type, void* pixels,
+                                                       Bool defaultFramebufferOrientation) {
         const Bool wantDepth = format != GL_STENCIL_INDEX;
         const Bool wantStencil = format != GL_DEPTH_COMPONENT;
         auto& frame = m_frameContext.GetCurrent();
@@ -8872,6 +9157,21 @@ void main() {
             VK_PIPELINE_STAGE_TRANSFER_BIT, srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, imageAspect, mipLevel, 1);
         MOBILEGL_ASSERT(ok, "%s: failed to transition depth-stencil source image", __func__);
 
+        // The swapchain's depth/stencil image is stored display-side-up like its colour twin, so
+        // the GL rect has to be mapped into that space before the copy and the copied rows
+        // re-oriented afterwards - the same two halves the colour ReadPixels path applies.
+        Int32 copyOffsetX = x;
+        Int32 copyOffsetY = y;
+        if (defaultFramebufferOrientation) {
+            const VkExtent2D defaultFboExtent = m_swapchainObject.GetExtent();
+            const DefaultFramebufferRectMapping mapping =
+                GetDefaultFramebufferRectMapping(m_swapchainObject.GetPreTransform());
+            copyOffsetX = MapDefaultFramebufferRectAxis(x, width, static_cast<Int>(defaultFboExtent.width),
+                                                        mapping.mirrorX);
+            copyOffsetY = MapDefaultFramebufferRectAxis(y, height, static_cast<Int>(defaultFboExtent.height),
+                                                        mapping.flipY);
+        }
+
         VkBufferImageCopy regions[2]{};
         Uint32 regionCount = 0;
         if (wantDepth) {
@@ -8881,7 +9181,7 @@ void main() {
             region.imageSubresource.mipLevel = mipLevel;
             region.imageSubresource.baseArrayLayer = baseArrayLayer;
             region.imageSubresource.layerCount = 1;
-            region.imageOffset = {x, y, 0};
+            region.imageOffset = {copyOffsetX, copyOffsetY, 0};
             region.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
         }
         if (wantStencil) {
@@ -8891,7 +9191,7 @@ void main() {
             region.imageSubresource.mipLevel = mipLevel;
             region.imageSubresource.baseArrayLayer = baseArrayLayer;
             region.imageSubresource.layerCount = 1;
-            region.imageOffset = {x, y, 0};
+            region.imageOffset = {copyOffsetX, copyOffsetY, 0};
             region.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
         }
         vkCmdCopyImageToBuffer(frame.commandBuffer, image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, readback.GetHandle(),
@@ -8916,6 +9216,38 @@ void main() {
         const Uint8* depthSrc = mapped;
         const Uint8* stencilSrc = mapped + stencilOffset;
 
+        // Re-orient the copied band per aspect, before any repacking reads it: the depth and
+        // stencil aspects were copied into their own tightly packed sub-buffers, so each is a
+        // plain width x height image of its own texel size.
+        Vector<Uint8> remappedDepth;
+        Vector<Uint8> remappedStencil;
+        if (defaultFramebufferOrientation) {
+            const VkSurfaceTransformFlagBitsKHR preTransform = m_swapchainObject.GetPreTransform();
+            Bool remapped = true;
+            if (wantDepth && depthCopyBytes > 0) {
+                remappedDepth.resize(pixelCount * depthCopyBytes);
+                remapped = RemapDefaultFboReadbackToGLOrientation(depthSrc, static_cast<Uint32>(width),
+                                                                  static_cast<Uint32>(height), preTransform,
+                                                                  depthCopyBytes, remappedDepth.data());
+            }
+            if (remapped && wantStencil) {
+                remappedStencil.resize(pixelCount);
+                remapped = RemapDefaultFboReadbackToGLOrientation(stencilSrc, static_cast<Uint32>(width),
+                                                                  static_cast<Uint32>(height), preTransform, 1,
+                                                                  remappedStencil.data());
+            }
+            if (remapped) {
+                if (!remappedDepth.empty()) depthSrc = remappedDepth.data();
+                if (!remappedStencil.empty()) stencilSrc = remappedStencil.data();
+            } else {
+                // Only a quarter-turn pre-transform reaches this, and nothing in this renderer
+                // models one. MGLOG_I because the INFO builds are the ones that run conformance.
+                MGLOG_I("DirectVulkan::ReadDepthStencilPixels: default-FBO remap declined (w=%d h=%d "
+                        "preTransform=%d); falling back to raw readback",
+                        width, height, static_cast<Int>(preTransform));
+            }
+        }
+
         const auto depthValueAt = [&](SizeT i) -> Float {
             switch (vkFormat) {
             case VK_FORMAT_D16_UNORM: {

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -211,10 +211,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                     GLsizei height, GLenum format, GLenum type, void* pixels);
         // Copy-and-repack core shared by depth-stencil ReadPixels and GetTexImage;
         // expects command recording to be active and any render pass already ended.
+        //
+        // `defaultFramebufferOrientation` is set only when the source is the swapchain's
+        // depth/stencil image, which this renderer stores display-side-up: the copy rect then
+        // has to be mapped out of GL's bottom-origin space and the copied rows re-oriented on
+        // the way back, exactly as the colour ReadPixels path does.
         void ReadDepthStencilImageToClient(VkImage image, VkFormat vkFormat, VkImageLayout* trackedLayout,
                                            VkImageAspectFlags imageAspect, Uint32 mipLevel, Uint32 baseArrayLayer,
                                            GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type,
-                                           void* pixels);
+                                           void* pixels, Bool defaultFramebufferOrientation = false);
         // Same-extent depth blit between images of different depth formats: host
         // round-trip with a per-texel re-encode (see BlitNamedFramebuffer).
         Bool BlitDepthAcrossFormats(FrameContext::FrameData& frame, VkImage srcImage, VkFormat srcFormat,
@@ -363,6 +368,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint32 samplerBinding = 0;
         };
 
+        // A single-sample staging image for multisample-resolve blits that also have to change
+        // orientation. vkCmdResolveImage cannot flip (it takes one offset per side, not the
+        // invertible pair vkCmdBlitImage takes), so a resolve into or out of the default
+        // framebuffer used to land the mirrored band. Resolving here first and then blitting from
+        // here separates the two operations, and each one then does only what it can express.
+        //
+        // Pooled rather than created per blit: the CTS runs hundreds of these back to back, and
+        // create-destroy per call would both cost allocations and, worse, need per-call deferred
+        // destruction to outlive the recording. It grows to the largest extent asked for and is
+        // reused; format changes recreate it.
+        struct MultisampleResolveScratchImage {
+            VkImage image = VK_NULL_HANDLE;
+            VmaAllocation allocation = VK_NULL_HANDLE;
+            VkFormat format = VK_FORMAT_UNDEFINED;
+            VkExtent2D extent = {0, 0};
+            VkImageLayout layout = VK_IMAGE_LAYOUT_UNDEFINED;
+        };
+        MultisampleResolveScratchImage m_msResolveScratch;
+        // Returns a scratch image at least `extent` in size with exactly `format`, transitioned to
+        // TRANSFER_DST and ready to be resolved into. Null image on failure (the caller then falls
+        // back to the direct resolve).
+        Bool AcquireMultisampleResolveScratchImage(VkCommandBuffer commandBuffer, VkFormat format,
+                                                   VkExtent2D extent);
+        void DestroyMultisampleResolveScratchImage();
+
         struct DeferredDepthMipmapCleanup {
             Vector<VkImageView> imageViews;
             Vector<VkFramebuffer> framebuffers;
@@ -1125,6 +1155,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool MaterializePendingClearForDefaultFramebuffer(VkCommandBuffer commandBuffer,
                                                           MG_State::GLState::FramebufferObject& fbo,
                                                           FramebufferAttachmentType attachmentType);
+        // Its depth/stencil half: a different image (the swapchain's depth/stencil twin), a
+        // different clear command and per-aspect masking.
+        Bool MaterializePendingDepthStencilClearForDefaultFramebuffer(
+            VkCommandBuffer commandBuffer, const MG_State::GLState::FramebufferAttachmentObject& attachment,
+            const ClearAttachmentPayload& payload);
         VkPipeline GetOrCreateBlitPipeline(const RenderPassEntry& renderPassEntry);
         Bool GenerateDepthMipmapWithShader(FrameContext::FrameData& frame,
                                            MG_State::GLState::ITextureObject& texture,

MobileGL/MG_IntegrationTest/CMakeLists.txt

@@ -58,6 +58,9 @@ add_executable(MobileGLIntegrationTest
     Scenarios/PixelStoreSweepScenario.cpp
     Scenarios/FragCoordOriginScenario.cpp
     Scenarios/ClearThenReadPixelsScenario.cpp
+    Scenarios/DepthStencilReadbackScenario.cpp
+    Scenarios/SsboArrayLengthScenario.cpp
+    Scenarios/SwizzleAccessRoutineScenario.cpp
 )
 
 target_include_directories(MobileGLIntegrationTest PRIVATE

MobileGL/MG_IntegrationTest/Scenarios/ClearThenReadPixelsScenario.cpp

@@ -180,4 +180,156 @@ void main() { o_color = vec4(0.1, 0.2, 0.3, 1.0); }
         gl.EndFrame();
         glDeleteProgram(program);
     }
+
+    // The other half of the same rule, and the one the first version of this fix got wrong: a
+    // parked clear must be executed BEFORE whatever writes the framebuffer next, not whenever the
+    // readback happens to notice it. Minecraft clears the default framebuffer, renders the world
+    // into its own framebuffer and blits the result out; nothing in between opens a render pass on
+    // the default framebuffer, so the clear stays parked across the whole frame. Materializing it
+    // at readback time therefore ran it AFTER the blit and returned a blank frame - which is what
+    // took every DirectVulkan retrace to ssim 0.000005.
+    TEST_F(ClearThenReadPixelsScenario, ABlitIntoTheDefaultFramebufferSurvivesAnEarlierClear) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        // Paint a source framebuffer, exactly as a game renders its world off-screen.
+        ColorFbo source = MakeColorFbo(width, height);
+        ASSERT_NE(source.fbo, 0u);
+        BindFbo(source);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        DrawFullViewportQuad(program);
+
+        // Clear the DEFAULT framebuffer, then blit the source over it. The clear is white so a
+        // frame that lost the blit is unmistakable, and the blit's colour is fshSimple's.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        ClearTo(1.0f, 1.0f, 1.0f, 1.0f);
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, source.fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const Image blitted = ReadPixels(width, height);
+        EXPECT_EQ(FirstGLError(), 0u);
+        const Rgba8 centre = blitted.At(width / 2, height / 2);
+        EXPECT_NEAR(centre.r, 26, 2) << "the blit into the default framebuffer did not survive the clear that "
+                                        "preceded it; read back rgba(" << static_cast<int>(centre.r) << ", "
+                                     << static_cast<int>(centre.g) << ", " << static_cast<int>(centre.b) << ", "
+                                     << static_cast<int>(centre.a) << ")";
+        EXPECT_NEAR(centre.g, 51, 2);
+        EXPECT_NEAR(centre.b, 77, 2);
+
+        DestroyColorFbo(source);
+        gl.EndFrame();
+        glDeleteProgram(program);
+    }
+
+    // A MULTISAMPLE-RESOLVE blit into the default framebuffer has to change orientation like any
+    // other, but vkCmdResolveImage takes one offset per side and cannot invert an axis, so it used
+    // to land the mirrored band. The renderer now resolves into a single-sample scratch image and
+    // blits from there. The source is painted in two horizontal bands so the mirror is visible;
+    // a full-extent uniform blit is a fixed point of the flip and would prove nothing.
+    TEST_F(ClearThenReadPixelsScenario, AMultisampleResolveBlitIntoTheDefaultFramebufferKeepsItsOrientation) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(height, 8);
+
+        GLint maxSamples = 0;
+        glGetIntegerv(GL_MAX_SAMPLES, &maxSamples);
+        if (maxSamples < 2) {
+            GTEST_SKIP() << "GL_MAX_SAMPLES is " << maxSamples << "; this needs a multisample renderbuffer";
+        }
+
+        GLuint fbo = 0, rbo = 0;
+        glGenFramebuffers(1, &fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glGenRenderbuffers(1, &rbo);
+        glBindRenderbuffer(GL_RENDERBUFFER, rbo);
+        glRenderbufferStorageMultisample(GL_RENDERBUFFER, 2, GL_RGBA8, width, height);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, rbo);
+        if (glCheckFramebufferStatus(GL_FRAMEBUFFER) != GL_FRAMEBUFFER_COMPLETE) {
+            glDeleteRenderbuffers(1, &rbo);
+            glDeleteFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            GTEST_SKIP() << "no complete 2x multisample RGBA8 renderbuffer on this driver";
+        }
+        glViewport(0, 0, width, height);
+
+        // Bottom half red, top half blue - via scissored clears, so no shader is involved.
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(0, 0, width, height / 2);
+        ClearTo(1.0f, 0.0f, 0.0f, 1.0f);
+        glScissor(0, height / 2, width, height - height / 2);
+        ClearTo(0.0f, 0.0f, 1.0f, 1.0f);
+        glDisable(GL_SCISSOR_TEST);
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const Image resolved = ReadPixels(width, height);
+        EXPECT_EQ(FirstGLError(), 0u);
+        const Rgba8 bottom = resolved.At(width / 2, height / 4);
+        const Rgba8 top = resolved.At(width / 2, height - 1 - height / 4);
+        EXPECT_GT(bottom.r, 200) << "the bottom band should be red after the resolve, got rgba("
+                                 << static_cast<int>(bottom.r) << ", " << static_cast<int>(bottom.g) << ", "
+                                 << static_cast<int>(bottom.b) << ") - blue there means the resolve landed "
+                                 << "in the mirrored band";
+        EXPECT_LT(bottom.b, 60);
+        EXPECT_GT(top.b, 200) << "the top band should be blue after the resolve, got rgba("
+                              << static_cast<int>(top.r) << ", " << static_cast<int>(top.g) << ", "
+                              << static_cast<int>(top.b) << ")";
+        EXPECT_LT(top.r, 60);
+
+        glDeleteRenderbuffers(1, &rbo);
+        glDeleteFramebuffers(1, &fbo);
+        gl.EndFrame();
+    }
+
+    // The same ordering claim for the path that DOES open a render pass. It passes today (the
+    // render pass folds the clear into its loadOp and pops it), and it is here so a future change
+    // to the pending-clear lifecycle cannot quietly reverse clear and draw.
+    TEST_F(ClearThenReadPixelsScenario, ADrawIntoTheDefaultFramebufferSurvivesAnEarlierClear) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(1.0f, 1.0f, 1.0f, 1.0f);
+        DrawFullViewportQuad(program);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const Image painted = ReadPixels(width, height);
+        const Rgba8 centre = painted.At(width / 2, height / 2);
+        EXPECT_NEAR(centre.r, 26, 2) << "the draw did not survive the clear that preceded it";
+        EXPECT_NEAR(centre.g, 51, 2);
+        EXPECT_NEAR(centre.b, 77, 2);
+
+        gl.EndFrame();
+        glDeleteProgram(program);
+    }
 } // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackScenario.cpp

@@ -0,0 +1,297 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - glReadPixels OF DEPTH AND STENCIL FROM THE DEFAULT FRAMEBUFFER.
+//
+// DirectVulkan's depth/stencil readback used to decline the default framebuffer outright
+// (`ReadDepthStencilPixels` returned at its first line) because that framebuffer's depth and
+// stencil "attachments" are placeholder texture objects backing no image - the real one is the
+// swapchain's depth/stencil twin. Declining meant the call raised no GL error and wrote NOTHING,
+// so the caller kept whatever its buffer already held.
+//
+// That silence is what the framebuffer_blit family trips over. Every one of its cases begins by
+// clearing the default framebuffer's depth and stencil and reading them straight back as a
+// sanity check, into a local pre-initialised to 0.2 (depth) and 50 (stencil); an untouched
+// buffer therefore reports "expected DEPTH[0.25] but got DEPTH[0.2]" and "expected STENCIL[1] but
+// got STENCIL[50]" - the exact strings in the 15 Magma failures - long before any blit happens.
+// A test that only checked "no GL error" would pass against the broken path, so every case here
+// poisons its destination with a value the correct answer cannot be.
+//
+// The orientation case is the second half. This renderer stores the default framebuffer
+// display-side-up and converts GL rects on their way in, so the depth copy needs the same rect
+// mapping and row re-ordering the colour readback got in the M-1 fix; without them a
+// vertically-varying depth buffer reads back mirrored, which no full-extent uniform-value test
+// can see.
+//
+// Depth/stencil readback through a USER framebuffer already worked and is asserted here too, as
+// the built-in control: it shares ReadDepthStencilImageToClient with the default-framebuffer
+// path, so it is what says a failure is about the default framebuffer specifically.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Values no correct read can produce, so "the backend wrote nothing" fails loudly instead
+        // of passing on whatever happened to be in the variable. These are the CTS's own poison
+        // values, which is why its logs report exactly them.
+        constexpr float kDepthPoison = 0.2f;
+        constexpr int kStencilPoison = 50;
+
+        class DepthStencilReadbackScenario : public ScenarioTest {
+        protected:
+            // DirectGLES reads depth and stencil back through the ES driver, which has no
+            // guaranteed path for either (GL_NV_read_depth / GL_NV_read_stencil are optional and
+            // absent on both the Adreno device and Mesa's ES). That gap is tracked separately as
+            // the packed_depth_stencil cluster and needs a shader-sampling emulation, not this
+            // change; asserting it here would only pin a known-missing feature.
+            bool BackendReadsDepthStencil() const { return Gl().BackendName() == "DirectVulkan"; }
+
+            float ReadDepthAt(int x, int y) const {
+                float depth = kDepthPoison;
+                glReadPixels(x, y, 1, 1, GL_DEPTH_COMPONENT, GL_FLOAT, &depth);
+                return depth;
+            }
+
+            int ReadStencilAt(int x, int y) const {
+                int stencil = kStencilPoison;
+                glReadPixels(x, y, 1, 1, GL_STENCIL_INDEX, GL_INT, &stencil);
+                return stencil;
+            }
+        };
+
+        // A depth buffer whose value depends on the row: bottom half `bottom`, top half `top`.
+        // Built with a scissored clear rather than a draw so the test stays independent of
+        // depth-test and shader behaviour.
+        void ClearDepthInBands(int width, int height, float bottom, float top) {
+            glEnable(GL_SCISSOR_TEST);
+            glScissor(0, 0, width, height / 2);
+            glClearDepth(bottom);
+            glClear(GL_DEPTH_BUFFER_BIT);
+            glScissor(0, height / 2, width, height - height / 2);
+            glClearDepth(top);
+            glClear(GL_DEPTH_BUFFER_BIT);
+            glDisable(GL_SCISSOR_TEST);
+        }
+
+    } // namespace
+
+    TEST_F(DepthStencilReadbackScenario, DefaultFramebufferDepthClearIsVisibleToReadPixels) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName()
+                         << " has no depth readback path (ES lacks GL_NV_read_depth); see the packed_depth_stencil "
+                            "cluster";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDepthMask(GL_TRUE);
+        glClearDepth(0.25);
+        glClear(GL_DEPTH_BUFFER_BIT);
+
+        const float centre = ReadDepthAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(centre, 0.25f, 1.0f / 4096.0f)
+            << "glReadPixels(GL_DEPTH_COMPONENT) of the default framebuffer returned " << centre
+            << (std::fabs(centre - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+
+        gl.EndFrame();
+    }
+
+    TEST_F(DepthStencilReadbackScenario, DefaultFramebufferStencilClearIsVisibleToReadPixels) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName()
+                         << " has no stencil readback path (ES lacks GL_NV_read_stencil); see the "
+                            "packed_depth_stencil cluster";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glStencilMask(0xFFu);
+        glClearStencil(3);
+        glClear(GL_STENCIL_BUFFER_BIT);
+
+        const int centre = ReadStencilAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(centre, 3) << "glReadPixels(GL_STENCIL_INDEX) of the default framebuffer returned " << centre
+                             << (centre == kStencilPoison ? " - the destination was never written at all" : "");
+
+        gl.EndFrame();
+    }
+
+    // The orientation half: a depth buffer that varies with the row must read back in GL's
+    // bottom-up order. A full-extent uniform clear is a fixed point of the flip, so only a banded
+    // buffer can tell the two apart.
+    TEST_F(DepthStencilReadbackScenario, DefaultFramebufferDepthReadbackKeepsTheGLRowOrder) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName() << " has no depth readback path";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(height, 8);
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDepthMask(GL_TRUE);
+        ClearDepthInBands(width, height, /*bottom=*/0.25f, /*top=*/0.75f);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const float bottom = ReadDepthAt(width / 2, height / 4);
+        const float top = ReadDepthAt(width / 2, height - 1 - height / 4);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(bottom, 0.25f, 1.0f / 4096.0f)
+            << "GL row " << (height / 4) << " is in the bottom band and was cleared to 0.25, but read back " << bottom
+            << " (0.75 there means the readback is upside down)";
+        EXPECT_NEAR(top, 0.75f, 1.0f / 4096.0f)
+            << "GL row " << (height - 1 - height / 4) << " is in the top band and was cleared to 0.75, but read back "
+            << top << " (0.25 there means the readback is upside down)";
+
+        gl.EndFrame();
+    }
+
+    // A depth blit INTO the default framebuffer has to convert its rect out of GL's bottom-origin
+    // space, exactly as the colour blit does. The colour path had that conversion and the
+    // depth path did not, so a scissored depth blit landed in the mirrored band - which is the
+    // whole of KHR-GL*.framebuffer_blit.scissor_blit once the readback above works well enough to
+    // see it (before that the test died on the poison values and never reached the blit).
+    TEST_F(DepthStencilReadbackScenario, AScissoredDepthBlitIntoTheDefaultFramebufferLandsInTheScissorBox) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName() << " has no depth readback path";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        // Source: a user framebuffer whose depth is uniformly 0.75.
+        GLuint fbo = 0, colorTex = 0, depthTex = 0;
+        glGenFramebuffers(1, &fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glGenTextures(1, &colorTex);
+        glBindTexture(GL_TEXTURE_2D, colorTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, colorTex, 0);
+        glGenTextures(1, &depthTex);
+        glBindTexture(GL_TEXTURE_2D, depthTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, width, height, 0, GL_DEPTH_STENCIL,
+                     GL_UNSIGNED_INT_24_8, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, depthTex, 0);
+        ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), GLenum(GL_FRAMEBUFFER_COMPLETE));
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDepthMask(GL_TRUE);
+        glClearDepth(0.75);
+        glClear(GL_DEPTH_BUFFER_BIT);
+
+        // Destination: the default framebuffer, depth 0 everywhere.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glClearDepth(0.0);
+        glClear(GL_DEPTH_BUFFER_BIT);
+
+        // Blit the whole rect, but scissored to the BOTTOM-LEFT quadrant in GL coordinates.
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(0, 0, width / 2, height / 2);
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
+        glDisable(GL_SCISSOR_TEST);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const float inside = ReadDepthAt(width / 4, height / 4);
+        const float above = ReadDepthAt(width / 4, height - 1 - height / 4);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(inside, 0.75f, 1.0f / 4096.0f)
+            << "GL (" << (width / 4) << ", " << (height / 4) << ") is inside the scissor box and should hold the "
+            << "blitted 0.75, but read back " << inside;
+        EXPECT_NEAR(above, 0.0f, 1.0f / 4096.0f)
+            << "GL (" << (width / 4) << ", " << (height - 1 - height / 4)
+            << ") is ABOVE the scissor box and must still hold the cleared 0.0, but read back " << above
+            << " (0.75 there means the depth blit landed in the mirrored band)";
+
+        glDeleteTextures(1, &depthTex);
+        glDeleteTextures(1, &colorTex);
+        glDeleteFramebuffers(1, &fbo);
+        gl.EndFrame();
+    }
+
+    // The control: the same read against a user framebuffer, which never went through the
+    // declined path. It is what makes a failure above specific to the default framebuffer.
+    TEST_F(DepthStencilReadbackScenario, UserFramebufferDepthClearIsVisibleToReadPixels) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName() << " has no depth readback path";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = 64;
+        const int height = 48;
+
+        GLuint fbo = 0, colorTex = 0, depthTex = 0;
+        glGenFramebuffers(1, &fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glGenTextures(1, &colorTex);
+        glBindTexture(GL_TEXTURE_2D, colorTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, colorTex, 0);
+        glGenTextures(1, &depthTex);
+        glBindTexture(GL_TEXTURE_2D, depthTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, width, height, 0, GL_DEPTH_STENCIL,
+                     GL_UNSIGNED_INT_24_8, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, depthTex, 0);
+        ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), GLenum(GL_FRAMEBUFFER_COMPLETE));
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDepthMask(GL_TRUE);
+        glStencilMask(0xFFu);
+        glClearDepth(0.5);
+        glClearStencil(7);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+
+        const float depth = ReadDepthAt(width / 2, height / 2);
+        const int stencil = ReadStencilAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, 0.5f, 1.0f / 4096.0f) << "user-framebuffer depth readback returned " << depth;
+        EXPECT_EQ(stencil, 7) << "user-framebuffer stencil readback returned " << stencil;
+
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        glDeleteTextures(1, &depthTex);
+        glDeleteTextures(1, &colorTex);
+        glDeleteFramebuffers(1, &fbo);
+        gl.EndFrame();
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/SsboArrayLengthScenario.cpp

@@ -0,0 +1,215 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/SsboArrayLengthScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - length() ON AN SSBO's UNSIZED ARRAY.
+//
+// GLSL's `arr.length()` on the trailing runtime array of a shader storage block is not a compile
+// time constant: it is (bound range - the array's byte offset inside the block) / array stride,
+// evaluated against whatever the descriptor actually covers. Three separate pieces of MobileGL
+// have to agree for that to come out right - the byte offsets the block layout was compiled with,
+// the buffer the frontend binding resolves to, and the offset/size a glBindBufferRange asked for -
+// and a defect in any one of them shows up only as a wrong integer, never as an error.
+//
+// KHR-GL43.shader_storage_buffer_object.advanced-unsizedArrayLength-* (28 Magma failures, all 28
+// passing on Espryt) reports exactly that: lengths too large by roughly the size of the members
+// preceding the array. The cases here are the same shape, reduced to what can be asserted in one
+// dispatch: a block with no preamble, a block with one, a two-element ARRAY OF BLOCKS (which
+// consumes two consecutive bindings and is where the conformance failures concentrate), and the
+// two glBindBufferRange forms.
+//
+// Every length is written into one output SSBO and read back, so a failure names the block and
+// prints the number the shader saw.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Bindings 0..3 are inputs (2 and 3 are the block array), 4 is the output.
+        constexpr const char* kComputeSource = R"(#version 430 core
+layout(local_size_x = 1) in;
+layout(std430, binding = 0) readonly buffer Input0 {
+    ivec4 g_input0[];
+};
+layout(std430, binding = 1) readonly buffer Input1 {
+    ivec4 pad1;
+    ivec4 data[];
+} g_input1;
+layout(std430, binding = 2) readonly buffer Input23 {
+    ivec4 data[];
+} g_input23[2];
+layout(std430, binding = 4) buffer Output {
+    int g_length[];
+};
+void main() {
+    g_length[0] = g_input0.length();
+    g_length[1] = g_input1.data.length();
+    g_length[2] = g_input23[0].data.length();
+    g_length[3] = g_input23[1].data.length();
+}
+)";
+
+        constexpr int kElementBytes = 16; // ivec4, std430
+
+        class SsboArrayLengthScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                GLint blocks = 0;
+                glGetIntegerv(GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS, &blocks);
+                if (blocks < 5) {
+                    GTEST_SKIP() << "GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS is " << blocks << "; this needs 5";
+                }
+                m_program = CompileComputeProgram(kComputeSource);
+                ASSERT_NE(m_program, 0u) << m_buildLog;
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                if (!m_buffers.empty()) glDeleteBuffers(static_cast<GLsizei>(m_buffers.size()), m_buffers.data());
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            unsigned int CompileComputeProgram(const char* source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute shader did not compile: ") + log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute program did not link: ") + log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            // A buffer of `elements` ivec4s, filled with a recognisable pattern.
+            GLuint MakeStorageBuffer(int elements) {
+                std::vector<int> contents(static_cast<std::size_t>(elements) * 4, 41);
+                GLuint buffer = 0;
+                glGenBuffers(1, &buffer);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+                glBufferData(GL_SHADER_STORAGE_BUFFER,
+                             static_cast<GLsizeiptr>(elements) * kElementBytes, contents.data(), GL_DYNAMIC_COPY);
+                m_buffers.push_back(buffer);
+                return buffer;
+            }
+
+            // Dispatches once and returns the four lengths the shader observed.
+            std::vector<int> RunAndReadLengths(GLuint outputBuffer) {
+                glUseProgram(m_program);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+                std::vector<int> lengths(4, -1);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, outputBuffer);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0,
+                                   static_cast<GLsizeiptr>(lengths.size() * sizeof(int)), lengths.data());
+                return lengths;
+            }
+
+            unsigned int m_program = 0;
+            std::string m_buildLog;
+            std::vector<GLuint> m_buffers;
+        };
+
+    } // namespace
+
+    // glBindBufferBase everywhere: the plain case, and the one that pins the block array.
+    TEST_F(SsboArrayLengthScenario, WholeBufferBindingsReportTheElementCount) {
+        if (!Ready() || IsSkipped()) return;
+
+        // input1 carries one ivec4 of preamble before its runtime array, so a length that ignores
+        // the member offset comes back one too large there and only there.
+        const GLuint input0 = MakeStorageBuffer(7);
+        const GLuint input1 = MakeStorageBuffer(1 + 5);
+        const GLuint input2 = MakeStorageBuffer(3);
+        const GLuint input3 = MakeStorageBuffer(4);
+        const GLuint output = MakeStorageBuffer(4);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, input0);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, input1);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 2, input2);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, input3);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 4, output);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<int> lengths = RunAndReadLengths(output);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(lengths[0], 7) << "Input0 (no preamble, 7 elements) reported length " << lengths[0];
+        EXPECT_EQ(lengths[1], 5) << "Input1 (1 ivec4 of preamble, 6 elements of storage) reported length "
+                                 << lengths[1] << "; 6 means the array's byte offset inside the block was ignored";
+        EXPECT_EQ(lengths[2], 3) << "Input23[0] (binding 2, 3 elements) reported length " << lengths[2];
+        EXPECT_EQ(lengths[3], 4) << "Input23[1] (binding 3, 4 elements) reported length " << lengths[3]
+                                 << "; a block array's second element must resolve to the NEXT binding";
+    }
+
+    // glBindBufferRange with a non-zero offset: length() must see only the bound window.
+    TEST_F(SsboArrayLengthScenario, RangeBindingsReportTheBoundWindow) {
+        if (!Ready() || IsSkipped()) return;
+
+        GLint alignment = 1;
+        glGetIntegerv(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT, &alignment);
+        if (alignment > 2 * kElementBytes) {
+            GTEST_SKIP() << "GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT is " << alignment
+                         << "; a two-element offset cannot be expressed";
+        }
+
+        const GLuint input0 = MakeStorageBuffer(7);
+        const GLuint input1 = MakeStorageBuffer(1 + 5);
+        const GLuint input2 = MakeStorageBuffer(3);
+        const GLuint input3 = MakeStorageBuffer(4);
+        const GLuint output = MakeStorageBuffer(4);
+        // Input0: window starts two elements in, so 5 remain.
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 0, input0, 2 * kElementBytes, 5 * kElementBytes);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, input1);
+        // Both elements of the block array get a window, so a failure says whether the array's
+        // FIRST element is handled and only the later ones are lost, or neither is.
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 2, input2, 0, 2 * kElementBytes);
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 3, input3, 0, 2 * kElementBytes);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 4, output);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<int> lengths = RunAndReadLengths(output);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(lengths[0], 5) << "Input0 bound as [2 elements, 5 elements) reported length " << lengths[0]
+                                 << "; 7 means glBindBufferRange's offset/size never reached the descriptor";
+        EXPECT_EQ(lengths[2], 2) << "Input23[0] bound as [0, 2 elements) reported length " << lengths[2];
+        EXPECT_EQ(lengths[3], 2) << "Input23[1] bound as [0, 2 elements) reported length " << lengths[3];
+
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, input0);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, input3);
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/SwizzleAccessRoutineScenario.cpp

@@ -0,0 +1,310 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/SwizzleAccessRoutineScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - EVERY TEXTURE ACCESS ROUTINE READS THE SAME TEXEL OUT OF A usampler2DArray.
+//
+// KHR-GL33/GL40.texture_swizzle.smoke_access_idx_* sweeps the fourteen GLSL texture access
+// routines against a 1x1x1 GL_RGBA32UI GL_TEXTURE_2D_ARRAY and asserts the fetched channel. On
+// Espryt, `texture` and `textureGrad` pass while `textureLod`, `textureOffset`, `texelFetch`,
+// `texelFetchOffset` and `textureLodOffset` fail - 21 cases per version, 42 across GL33 and GL40.
+// The discriminator is the important part: the swizzle state is IDENTICAL across all of them, so
+// swizzle delivery is not the defect; what differs is only how the routine is spelled, i.e. what
+// SPIRV-Cross has to emit into ESSL for it.
+//
+// This scenario is that discriminator, reduced to something that fails in milliseconds: one draw
+// per access routine against the same texture and the same swizzle, all reading the same texel.
+// A routine that disagrees with the others is the defect, and the failure message names it.
+//
+// The shader shape is copied from the conformance test rather than idealised - including its
+// `int(0)` level-of-detail argument, which is a desktop-GLSL implicit int->float conversion that
+// ESSL does not have, and its zero offsets. Both are exactly the things a GLSL -> SPIR-V -> ESSL
+// round trip can lose.
+//
+// DirectVulkan is the built-in control: it consumes the SPIR-V directly and never runs the ESSL
+// emission, so a failure there would mean the scenario, not the backend.
+
+#include <array>
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // The conformance test's own source texel, one recognisable value per channel.
+        constexpr std::uint32_t kSourceTexel[4] = {0x3FFFFFFFu, 0x7FFFFFFFu, 0xBFFFFFFFu, 0xFFFFFFFFu};
+
+        constexpr int kOutputWidth = 8;
+        constexpr int kOutputHeight = 8;
+
+        // The blank vertex shader the smoke test uses: a full-viewport strip with no attributes.
+        constexpr const char* kVertexSource = R"(#version 330 core
+void main()
+{
+    switch (gl_VertexID)
+    {
+      case 0: gl_Position = vec4(-1.0, 1.0, 0.0, 1.0); break;
+      case 1: gl_Position = vec4( 1.0, 1.0, 0.0, 1.0); break;
+      case 2: gl_Position = vec4(-1.0,-1.0, 0.0, 1.0); break;
+      case 3: gl_Position = vec4( 1.0,-1.0, 0.0, 1.0); break;
+    }
+}
+)";
+
+        struct AccessRoutine {
+            const char* name;     // as it appears in the conformance case name
+            const char* callText; // the whole TEXTURE_ACCESS(sampler, ARGUMENTS) expression
+        };
+
+        // Spelled exactly as gl3cTextureSwizzleTests.cpp's prepareArguments builds them for
+        // GL_TEXTURE_2D_ARRAY: three coordinates, `int(0)` for the level, ivec2 offsets.
+        constexpr AccessRoutine kRoutines[] = {
+            {"texture", "texture(smp, vec3(0, 0, 0))"},
+            {"textureLod", "textureLod(smp, vec3(0, 0, 0), int(0))"},
+            {"textureOffset", "textureOffset(smp, vec3(0, 0, 0), ivec2(0, 0))"},
+            {"texelFetch", "texelFetch(smp, ivec3(0, 0, 0), int(0))"},
+            {"texelFetchOffset", "texelFetchOffset(smp, ivec3(0, 0, 0), int(0), ivec2(0, 0))"},
+            {"textureLodOffset", "textureLodOffset(smp, vec3(0, 0, 0), int(0), ivec2(0, 0))"},
+            {"textureGrad", "textureGrad(smp, vec3(0, 0, 0), vec2(0, 0), vec2(0, 0))"},
+            {"textureGradOffset", "textureGradOffset(smp, vec3(0, 0, 0), vec2(0, 0), vec2(0, 0), ivec2(0, 0))"},
+        };
+
+        constexpr const char* kChannels[4] = {"x", "y", "z", "w"};
+
+        std::string FragmentSource(const AccessRoutine& routine, int channel) {
+            return std::string("#version 330 core\n\nuniform usampler2DArray smp;\n\nout uint out_color;\n\n"
+                               "void main()\n{\n    uint result = ") +
+                   routine.callText + "." + kChannels[channel] + ";\n\n    out_color = result;\n}\n";
+        }
+
+        class SwizzleAccessRoutineScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                // 1x1x1 RGBA32UI 2D array. Integer textures are not filterable, so NEAREST is
+                // mandatory, and a single level means every LOD argument must resolve to 0.
+                glGenTextures(1, &m_sourceTexture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, m_sourceTexture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_RGBA32UI, 1, 1, 1);
+                glTexSubImage3D(GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, 1, 1, 1, GL_RGBA_INTEGER, GL_UNSIGNED_INT,
+                                kSourceTexel);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
+                ASSERT_EQ(FirstGLError(), 0u) << "source texture setup left a GL error behind";
+
+                // 8x8 R32UI render target, read back with glReadPixels.
+                glGenTextures(1, &m_outputTexture);
+                glBindTexture(GL_TEXTURE_2D, m_outputTexture);
+                glTexStorage2D(GL_TEXTURE_2D, 1, GL_R32UI, kOutputWidth, kOutputHeight);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glGenFramebuffers(1, &m_fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, m_outputTexture, 0);
+                ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), GLenum(GL_FRAMEBUFFER_COMPLETE));
+                glGenVertexArrays(1, &m_vao);
+                ASSERT_EQ(FirstGLError(), 0u) << "output framebuffer setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_fbo != 0) glDeleteFramebuffers(1, &m_fbo);
+                if (m_outputTexture != 0) glDeleteTextures(1, &m_outputTexture);
+                if (m_sourceTexture != 0) glDeleteTextures(1, &m_sourceTexture);
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            }
+
+            void SetSwizzle(GLenum r, GLenum g, GLenum b, GLenum a) {
+                glBindTexture(GL_TEXTURE_2D_ARRAY, m_sourceTexture);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_R, static_cast<GLint>(r));
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_G, static_cast<GLint>(g));
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_B, static_cast<GLint>(b));
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_A, static_cast<GLint>(a));
+            }
+
+            // Renders one access routine into the 8x8 target and returns every texel it wrote.
+            // Returns an empty vector (with a gtest failure already recorded) if the program did
+            // not build.
+            std::vector<std::uint32_t> Render(const AccessRoutine& routine, int channel) {
+                const std::string fragment = FragmentSource(routine, channel);
+                std::string error;
+                const unsigned int program = CompileProgram(kVertexSource, fragment.c_str(), &error);
+                if (program == 0) {
+                    ADD_FAILURE() << routine.name << " channel " << kChannels[channel]
+                                  << ": program did not build: " << error << "\n--- source ---\n"
+                                  << fragment;
+                    return {};
+                }
+
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glViewport(0, 0, kOutputWidth, kOutputHeight);
+                glDisable(GL_SCISSOR_TEST);
+                glDisable(GL_DEPTH_TEST);
+                const GLuint clearValue[4] = {0xDEADBEEFu, 0u, 0u, 0u};
+                glClearBufferuiv(GL_COLOR, 0, clearValue);
+
+                glUseProgram(program);
+                glActiveTexture(GL_TEXTURE0);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, m_sourceTexture);
+                const GLint location = glGetUniformLocation(program, "smp");
+                glUniform1i(location, 0);
+                glBindVertexArray(m_vao);
+                glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+                glBindVertexArray(0);
+
+                std::vector<std::uint32_t> texels(static_cast<std::size_t>(kOutputWidth) * kOutputHeight, 0);
+                glReadPixels(0, 0, kOutputWidth, kOutputHeight, GL_RED_INTEGER, GL_UNSIGNED_INT, texels.data());
+                glUseProgram(0);
+                glDeleteProgram(program);
+                return texels;
+            }
+
+            // Asserts every texel equals `expected`, naming the routine and the first offender.
+            void ExpectAllTexels(const AccessRoutine& routine, int channel, std::uint32_t expected,
+                                 const std::vector<std::uint32_t>& texels) {
+                if (texels.empty()) return;
+                std::size_t offenders = 0;
+                std::uint32_t firstBad = 0;
+                std::size_t firstIndex = 0;
+                for (std::size_t i = 0; i < texels.size(); ++i) {
+                    if (texels[i] == expected) continue;
+                    if (offenders == 0) {
+                        firstBad = texels[i];
+                        firstIndex = i;
+                    }
+                    ++offenders;
+                }
+                EXPECT_EQ(offenders, 0u)
+                    << routine.name << "(...)." << kChannels[channel] << " returned 0x" << std::hex << firstBad
+                    << " instead of 0x" << expected << std::dec << " at texel " << firstIndex << " (" << offenders
+                    << " of " << texels.size() << " wrong)";
+            }
+
+            GLuint m_sourceTexture = 0;
+            GLuint m_outputTexture = 0;
+            GLuint m_fbo = 0;
+            GLuint m_vao = 0;
+        };
+
+    } // namespace
+
+    // Identity swizzle: every routine must fetch the channel it was asked for. This is the
+    // scenario's floor - it does not involve swizzling at all, so a failure here is purely about
+    // how the access routine itself survives the trip to the backend.
+    TEST_F(SwizzleAccessRoutineScenario, EveryAccessRoutineFetchesTheSameTexelUnderTheIdentitySwizzle) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_RED, GL_GREEN, GL_BLUE, GL_ALPHA);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        for (const AccessRoutine& routine : kRoutines) {
+            for (int channel = 0; channel < 4; ++channel) {
+                const std::vector<std::uint32_t> texels = Render(routine, channel);
+                EXPECT_EQ(FirstGLError(), 0u) << routine.name << " left a GL error behind";
+                ExpectAllTexels(routine, channel, kSourceTexel[channel], texels);
+            }
+        }
+        Gl().EndFrame();
+    }
+
+    // A real swizzle, applied to every routine. Reversing the channels means a routine that
+    // silently drops the swizzle returns the UNSWIZZLED texel rather than nothing, so the
+    // failure distinguishes "swizzle lost" from "fetch broken".
+    TEST_F(SwizzleAccessRoutineScenario, EveryAccessRoutineSeesAReversedSwizzle) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_ALPHA, GL_BLUE, GL_GREEN, GL_RED);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::uint32_t expected[4] = {kSourceTexel[3], kSourceTexel[2], kSourceTexel[1], kSourceTexel[0]};
+        for (const AccessRoutine& routine : kRoutines) {
+            for (int channel = 0; channel < 4; ++channel) {
+                const std::vector<std::uint32_t> texels = Render(routine, channel);
+                EXPECT_EQ(FirstGLError(), 0u) << routine.name << " left a GL error behind";
+                ExpectAllTexels(routine, channel, expected[channel], texels);
+            }
+        }
+        Gl().EndFrame();
+    }
+
+    // Program churn: the shape that made the conformance suite fail, reduced.
+    //
+    // The swizzle smoke test builds one program per swizzle combination - 1,296 per case - and
+    // DirectGLES created a driver shader object per attached shader without ever calling
+    // glDeleteShader. glDeleteShader only FLAGS a shader for deletion (the driver frees it once
+    // nothing has it attached), so without that call the program's own deletion could not free
+    // them either: eight cases left ~20,000 live driver shaders behind, the Adreno ES driver
+    // passed its ceiling, and it began mis-serving shaders - first the sampling variants with the
+    // most image operands (textureLod/texelFetch/*Offset), while plain texture/textureGrad still
+    // worked. On device this loop plus a value check is the whole defect.
+    //
+    // HONEST LIMIT OF THIS TEST: llvmpipe has no such ceiling, so this passes here whether or not
+    // the leak is present - it cannot fail on the CI lane. It is a standing guard for the SHAPE
+    // (build many programs, keep reading the right texel) and the place to raise the iteration
+    // count if a driver ceiling ever needs reproducing; the leak itself is pinned by device
+    // measurement (VmRSS flat at ~137 MB across the 32-case family, against 132 -> 154 MB and
+    // still climbing before the fix).
+    TEST_F(SwizzleAccessRoutineScenario, RepeatedProgramBuildsKeepFetchingTheSameTexel) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_RED, GL_GREEN, GL_BLUE, GL_ALPHA);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // One routine from each side of the device's failure order, so a ceiling that takes the
+        // vulnerable one down first is still caught.
+        const AccessRoutine& plain = kRoutines[0];      // texture
+        const AccessRoutine& explicitLod = kRoutines[1]; // textureLod
+        constexpr int kIterations = 200;
+
+        for (int i = 0; i < kIterations; ++i) {
+            const AccessRoutine& routine = (i % 2 == 0) ? plain : explicitLod;
+            const int channel = i % 4;
+            const std::vector<std::uint32_t> texels = Render(routine, channel);
+            if (::testing::Test::HasFailure()) return; // a build failure repeats 200 times; say it once
+            ExpectAllTexels(routine, channel, kSourceTexel[channel], texels);
+            if (::testing::Test::HasFailure()) {
+                ADD_FAILURE() << "diverged at iteration " << i << " of " << kIterations;
+                return;
+            }
+        }
+        EXPECT_EQ(FirstGLError(), 0u) << "the churn loop left a GL error behind";
+        Gl().EndFrame();
+    }
+
+    // GL_ONE and GL_ZERO, which the conformance table spells as the literal values 1 and 0 and
+    // which the backend has to synthesise rather than fetch.
+    TEST_F(SwizzleAccessRoutineScenario, EveryAccessRoutineSeesConstantSwizzleSources) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_ONE, GL_ZERO, GL_ONE, GL_ZERO);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::uint32_t expected[4] = {1u, 0u, 1u, 0u};
+        for (const AccessRoutine& routine : kRoutines) {
+            for (int channel = 0; channel < 4; ++channel) {
+                const std::vector<std::uint32_t> texels = Render(routine, channel);
+                EXPECT_EQ(FirstGLError(), 0u) << routine.name << " left a GL error behind";
+                ExpectAllTexels(routine, channel, expected[channel], texels);
+            }
+        }
+        Gl().EndFrame();
+    }
+} // namespace MGITest