docs: record independent GPU mutation proof and gate repair dispatches

docs/research/2026-09-01-overhaul/477-lead-verification.md

@@ -0,0 +1,61 @@
+# #477 — independent lead verification, 2026-09-05
+
+Frozen scratch s5-477-openai-impl at29730506c57b44cfa3ba454a8594e170c95d7233;
+five changed files, including the untracked synchronization test. This is
+pre-integration proof, not graphical acceptance or AMD-reset attribution.
+
+Lead read every production/test change and the actual native Record witness.
+The test-local Silk context executes VulkanUploadQueue.Record and intercepts
+CmdCopyBuffer/PipelineBarrier2 arguments. Host-staging descriptors are inserted
+using the private production descriptor shape; this does not execute Stage's
+native allocation path. Migrations use the real EnqueueBufferCopy entry point.
+Same-drain, chained and earlier-drain cases therefore exercise real recording,
+not GPU execution or a synthetic schedule. Source guards bind passive acquire
+and depth factories to their production callers. No AC algorithm changes.
+
+## Lead execution
+
+Both AcDream.App and acclient absent before EVERY build/test. ACE UDP9000
+PID13340 up at the preflight. No graphical client launched. Commands in the
+scratch, Release, --no-restore and --nologo:
+
+- dotnet build AcDream.slnx -c Release: 0 warnings/0 errors, before and after
+  mutations (final3.25s). No full-solution test claim.
+- dotnet test tests/AcDream.App.Tests/AcDream.App.Tests.csproj -c Release
+  --no-build --filter 'FullyQualifiedName~AcDream.App.Tests.Rendering.Gpu&Lane!=Vulkan':
+  314/314, zero failures/skips both before and after mutations. Includes all
+  13 synchronization tests. Final TRX artifacts/477-root/restored/root-gpu-rhi-restored.trx,
+  SHA256 3E24078A6FAFEE2B921FFEB43D87270F20CC4F0BA8B95BC68F98C638BDF7490F.
+- git diff --check: clean (existing LF/CRLF notices only).
+
+Four separate actual-production mutations, each built by dotnet test without
+--no-build and individually restored by apply_patch, all failed exactly1/1:
+
+| Mutation | Filter suffix | Intended failure | TRX under artifacts/477-root |
+|---|---|---|---|
+| First acquire source reverted | AcquiredImage_FirstTransitionChainsToTheActualSemaphoreWait | expected ColorAttachmentOutput, actual TopOfPipe | m1/root-m1.trx |
+| Resolve entry masks reverted | DepthTransitions_SampleableMsaaResolveUsesColorWriterAtEntryAndExit | expected ColorAttachmentOutput, actual Early/LateFragmentTests | m2/root-m2.trx |
+| Resolve sampling-exit masks reverted | same depth test | same incorrect source stage | m3/root-m3.trx |
+| Real pre-migration barrier removed | UploadQueue_ActualNativeRecord_StagesThenMigratesWithExactDependencyAndOneConsumerBarrier | emitted sequence index1 expected BufferBarrier, actual CopyBuffer | m4/root-m4.trx |
+
+TRX SHA256 respectively:
+DEA861CED1EF1F9130E6166EC26D6C60C64D8F33FBEAE81A5648AD2FFD33852D;
+D4ABD02A37C63E658CA413D3BCB42F7AA13A563EE00A57F7DFC92E360B4D7905;
+69C818DB50DBA30B0B6A6857C7847A4524B3C7FD8CC1DBAF80055FC793C01C94;
+8EDF4C756D0963BB9E097A37BD1E89E33FB95685C287434D37F15F2F04D5BB32.
+Lead read XML ErrorInfo, not just test exit status. No setup/build failures.
+
+Raw source hashes after each applicable restoration and at final freeze:
+
+- Resources:57F4D23CAF3FAA6B51DB10F24193BDDE09C52401F228AD092CD35B17FB247250
+- Device:E545DF89768B8D9EA45DC34BD37F2FFEA9929422C78A1F118498426E355A6D03
+- UploadQueue:57E1B42FDDE4E26FD94DD94F1EB0AED91F0C41F55551612D72DE2A9AFEC5EB5B
+- Tests:60D106A46DDB61DB0F2EAC3C3B3FEFF4EE9C95C543B1CF12395C12DB8B65ADD4
+- Architecture:EE32F0EBCDAD39B0AF288F742045B44D66899432F601102B141EEE3D73C5B630
+
+F1 command-witness gap is closed by this proof; production unchanged from
+round0. Independent API/behavior lens1 by Wegener PASS, no actionable finding:
+actual native producer/consumer bindings, narrow masks, exact ranges, source
+hashes and four lead mutation outcomes checked. Rawls now performs sequential
+production/lifetime lens2. Integration and native validation remain owed.
+G4 remains FAIL. No intentional deviation or new register row.

docs/research/2026-09-01-overhaul/s5-consumers-material-closeout-packet.md

@@ -4714,3 +4714,23 @@ independently; root must integrate only accepted capture portions alongside
 waits for477 validation and an exclusive no-build/test window. The self-gate
 runner now explicitly starts hidden and refuses either graphical client;
 PowerShell parse/diff check pass. No rendering policy/quality change.
+
+## 51. Independent GPU checks; capture dispatch; hand-item return
+
+2026-09-05. #477 fix-round1 returned actual native-command recording tests.
+Lead independently repeated all four mutations, observed their intended FAILs,
+restored raw hashes and ran Release0W0E plus314/314 portableGPU tests. The
+lead's original F1 evidence gap is closed;477-lead-verification.md records
+exact results. Wegener API/behavior review1 PASS, no findings; Rawls performs
+production/lifetime review2. Native graphical validation owed. No integration yet.
+
+#476 dispatched to James in fresh codex/s5-476-openai-impl at6347e0f8269e201340b0c7a6b826ad4cfc21f43e,
+using476-completed-frame-capture-contract.md. #477 scratch is frozen; root
+will integrate accepted changes narrowly to preserve both Vulkan partial edits.
+#480 returned eight files from Rawls at9556c1748: agent-reported Release0W0E,
+92/92 affected tests and two expected mutation failures. Lead verification and
+independent reviews are not yet complete; no visual PASS is claimed.
+
+Neither graphical client is running; ACE9000 PID13340 up. No new client run.
+#478 actual lighting cause, #479 exact displaced part and #481 performance
+remain unproven. All outstanding§49 obligations and owner G4 FAIL remain.