[Fix] (DirectGLES/Managers, MG_Remote/Server): m-5 nulls a stale hostBytes only when its server shadow is actually gone (StagedShadowStore::HasShadow), not on a twins first-ensure generation sync - the unguarded null dropped a subdatas freshly staged base and made the reduced-path VBO read shifted; HasShadow unit control added

MobileGL/MG_Backend/DirectGLES/Managers.cpp

@@ -2958,13 +2958,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
 #if MOBILEGL_BUILD_DISAGGREGATED
                 // m-5 / codex 5: OnBackendContextDestroyed ran MGL_SERVER_STAGED_DROP_ALL(), which
                 // frees every server shadow but does NOT null the hostBytes that name them - so a
-                // twin that SURVIVES a context loss (this is the block that repairs it) still
+                // twin that SURVIVES a context loss still carries a base into the freed allocation.
-                // carries a base into the freed allocation. The two other drop sites pair the drop
+                // Null it here, where a surviving twin is re-armed - but ONLY when the shadow is
-                // with something that makes the base unreachable (Ops_H_Destroy retires the twin;
+                // really gone. This block also runs on a twin's FIRST ensure (contextGeneration
-                // the map-persistent site nulls hostBytes on the next line); DropAll did neither.
+                // starts mismatched), and there the shadow a preceding resource_subdata just staged
-                // Null it here, at the one place a surviving twin is re-armed, so no freed base
+                // is still live; nulling it then would drop the reduced path's own bytes (it did,
-                // reaches glBufferData/glBufferSubData before the next content record refills it.
+                // and TriangleScenario read the wrong VBO). HasShadow is the discriminator: false
-                resource->hostBytes = nullptr;
+                // after DropAll, true after an ordinary Adopt.
+                if (!ServerStaged().HasShadow(resource)) {
+                    resource->hostBytes = nullptr;
+                }
 #endif
             }
 
@@ -3159,12 +3162,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 resource->persistentMapped = false;
                 resource->persistentPtr = nullptr;
                 resource->immutableStorage = false;
-#if MOBILEGL_PIPE_PUSH
-                // m-5 / codex 5: mirror the handle arm - DropAll freed the shadow this base named
-                // on context loss, so a surviving twin repaired here must not carry it forward.
-                resource->hostBytes = nullptr;
-#endif
             }
+            // m-5: the LEGACY arm (EnsureBufferResource) is reached only under monolith/push, where
+            // liveHostBase reads the frontend object's MappedData rather than a server shadow, so
+            // there is no freed server base to null here - the split freed-base hazard lives in
+            // EnsureBufferResourceForHandle above, guarded by HasShadow.
 
             // An immutable store nothing maps any more: a respecification of a buffer that
             // had been persistently mapped, which Ops_Respecify could not retire because it

MobileGL/MG_Remote/Server/StagedShadow.h

@@ -133,6 +133,19 @@ namespace MobileGL::MG_Remote::Server {
             return m_shadows.size();
         }
 
+        // Is there STILL a live shadow for this key? The m-5 discriminator: after DropAll (context
+        // loss) the key is erased, so a twin's cached hostBytes names freed memory and must be
+        // nulled; after an ordinary Adopt the key is present and its base is live. The
+        // generation-reset block runs on a twin's FIRST ensure too (the generation starts
+        // mismatched), and there the shadow a preceding subdata just staged is present - so nulling
+        // must key on THIS answer, not on the generation change alone, or the reduced path drops
+        // its own bytes.
+        Bool HasShadow(const void* key) const {
+            if (!m_any.load(std::memory_order_acquire)) return false;
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            return m_shadows.find(key) != m_shadows.end();
+        }
+
         // Adjacent ranges merge - there is no gap between them, so the union really is one run.
         // Ranges with a gap do NOT merge, and that is the whole mechanism: it is what makes a
         // missing record detectable instead of papered over.

MobileGL/MG_Test/Wire/ServerLoopTest.cpp

@@ -617,6 +617,25 @@ TEST(StagedShadowTest, CoverageIsExactAndAGapIsNotCovered) {
     EXPECT_TRUE(store.IsCovered(&key, 0, 48));
 }
 
+// m-5's discriminator, at unit scope: HasShadow is TRUE for a key that was Adopted and FALSE once
+// DropAll has run - which is exactly what tells the generation-reset block whether a twin's
+// hostBytes names a live server shadow (keep it) or a freed one (null it). A version that answered
+// "always live" would let a freed base reach the driver; "always gone" would drop a base a subdata
+// just staged in the same generation (that regression really happened - TriangleScenario read a
+// shifted VBO). Both directions are asserted here.
+TEST(StagedShadowTest, HasShadowIsTrueAfterAdoptAndFalseAfterTheShadowIsDropped) {
+    Server::StagedShadowStore store(/*copies=*/true);
+    const int key = 0;
+    Vector<Uint8> bytes(16, Uint8{0x44});
+    EXPECT_FALSE(store.HasShadow(&key)) << "nothing staged yet";
+    store.Adopt(&key, 16, bytes.data(), 0, 16);
+    EXPECT_TRUE(store.HasShadow(&key)) << "a staged key must read live, or the reset block nulls a "
+                                          "base a subdata just filled";
+    store.DropAll();
+    EXPECT_FALSE(store.HasShadow(&key)) << "after DropAll the base is freed and must read gone, or "
+                                           "a surviving twin hands the driver a dangling pointer";
+}
+
 TEST(StagedShadowTest, DropForgetsOneResourceAndDropAllForgetsEveryOne) {
     Server::StagedShadowStore store(/*copies=*/true);
     const int a = 0;