[Fix] (DirectVulkan): shrink the blended depth-write quirk to MIN/MAX extremum blends only - a fixture-wide trace sweep showed the additive ONE+ONE arm never fires on the 26.3 OIT chain (its accumulation passes disable depth writes themselves) and only hit unrelated additive glow content; also fail reflection toward the gl_FragDepth exemption and zero phantom default-FBO blend slots so stale indexed state cannot trigger the strip

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -131,25 +131,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     namespace {
-        // Order-independent accumulation blending: the write order of overlapping fragments
-        // does not change the result, which is what lets multi-pass chains re-rasterize the
-        // same geometry and combine per-pass contributions (MC 26.3 OIT: GL_MAX depth
-        // bounds, additive ONE+ONE transmittance/accumulate). Sorted-transparency "over"
-        // compositing (SRC_ALPHA-style factors) is order-dependent, drawn once per surface,
-        // and relies on its depth writes for occlusion - it must not be treated as hazardous.
-        // MIN/MAX ignore blend factors entirely per the Vulkan spec.
+        // MIN/MAX extremum blending: the signature of a depth-bounds accumulation pass
+        // (MC 26.3 OIT writes vec4(-linD, linD, deviceZ, 0) under GL_MAX while writing
+        // depth for its equality chain). MIN/MAX ignore blend factors per the Vulkan spec.
         //
-        // Deliberately color-channel only. A separate-alpha accumulation
-        // (glBlendEquationSeparate(GL_FUNC_ADD, GL_MAX)) whose color channel is an ordinary
-        // over-blend is not treated as hazardous: no known content pairs that shape with a
-        // depth-equality chain, and widening the test would re-capture sorted transparency.
+        // Deliberately the ONLY shape stripped. A quirk should touch as little unrelated
+        // content as possible, and a trace sweep of every fixture showed the wider
+        // alternatives all cost more than they fix:
+        //   - additive ONE+ONE with a depth write matched zero draws of the 26.3 chain
+        //     (its transmittance/accumulate passes disable depth writes themselves) - the
+        //     only real content it caught was harmless additive glow effects (Create);
+        //   - sorted-transparency "over" blends (SRC_ALPHA-style) are order-dependent,
+        //     drawn once per surface, and rely on their depth writes for occlusion;
+        //   - separate-alpha accumulation over an over-blending color channel has no
+        //     known pairing with a depth-equality chain (color channel only, see tests).
+        // If a future workload pairs another blend shape with an equality chain, widen
+        // this with that evidence in hand rather than pre-emptively.
         Bool IsAccumulationBlend(const VkPipelineColorBlendAttachmentState& attachment) {
-            if (attachment.colorBlendOp == VK_BLEND_OP_MIN || attachment.colorBlendOp == VK_BLEND_OP_MAX) {
-                return true;
-            }
-            return attachment.colorBlendOp == VK_BLEND_OP_ADD &&
-                   attachment.srcColorBlendFactor == VK_BLEND_FACTOR_ONE &&
-                   attachment.dstColorBlendFactor == VK_BLEND_FACTOR_ONE;
+            return attachment.colorBlendOp == VK_BLEND_OP_MIN ||
+                   attachment.colorBlendOp == VK_BLEND_OP_MAX;
         }
     } // namespace
 

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -70,11 +70,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // with an equality-inclusive compare on the re-rasterized geometry) requires
         // cross-pipeline position invariance that some mobile compilers do not provide, even
         // with the SPIR-V Invariant decoration; whole primitives then drop out of the later
-        // passes. Only order-independent accumulation blends (MIN/MAX, additive ONE+ONE) are
-        // stripped - that is the signature of such equality chains (MC 26.3 OIT) - while
-        // sorted-transparency "over" compositing (e.g. vanilla MC water, SRC_ALPHA factors),
-        // which draws each surface once and depends on its depth writes to occlude later
-        // passes, keeps them. Set at renderer initialization based on the active driver.
+        // passes. Only MIN/MAX extremum blends are stripped - the signature of such a
+        // chain's depth-bounds pass (MC 26.3 OIT), and per a fixture-wide trace sweep the
+        // only depth-writing shape the chain actually uses - so every other blend
+        // (sorted-transparency "over" like vanilla MC water, additive glows, ...) keeps
+        // its depth writes. Set at renderer initialization based on the active driver.
         static void SetSuppressBlendedDepthWrite(Bool enabled);
         static Bool IsSuppressBlendedDepthWriteEnabled() { return s_suppressBlendedDepthWrite; }
         // Device gate for the quirk: ForceOn/ForceOff bypass detection, Auto enables it on

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -1547,6 +1547,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                             "ProgramFactory::ReflectFragmentOutputs: failed to create reflection module (result=%d)",
                             static_cast<Int>(createResult));
             if (createResult != SPV_REFLECT_RESULT_SUCCESS) {
+                // Fail toward the exemption: stripping a genuine gl_FragDepth writer would
+                // corrupt its depth output outright, while wrongly exempting an accumulation
+                // pass merely reverts that one program to the pre-quirk behavior.
+                entry.fragmentReplacesDepth = true;
                 continue;
             }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -3552,6 +3552,17 @@ void main() {
                 (bufferMask.a() ? VK_COLOR_COMPONENT_A_BIT : 0u));
             Bool effectiveBlendEnabled = blendEnabled;
             MG_State::GLState::ITextureObject* colorAttachmentTexture = nullptr;
+            if (isDefaultDrawFbo && i < drawBuffers.size() &&
+                drawBuffers[i] == FramebufferAttachmentType::None) {
+                // The default framebuffer spans the same MAX_DRAW_BUFFERS slots as an FBO
+                // (slot 0 is the back buffer, or None after glDrawBuffer(GL_NONE); slots
+                // 1+ are always None). Discard writes and blend state for the None slots
+                // like the FBO path below does, so stale indexed blend state on phantom
+                // slots cannot leak into the pipeline - most notably into the blended
+                // depth-write quirk's accumulation scan.
+                attachmentColorWriteMask = 0;
+                effectiveBlendEnabled = false;
+            }
             if (!isDefaultDrawFbo && i < drawBuffers.size()) {
                 const auto drawBuffer = drawBuffers[i];
                 colorAttachmentTexture = resolveCompleteColorAttachmentTexture(i);