docs: bound #477 command witness fix and resume #480

docs/research/2026-09-01-overhaul/477-fix-round1-command-witness.md

@@ -0,0 +1,55 @@
+# #477 fix round1 — execute the actual buffer recording path
+
+Lead return check2026-09-05. No production synchronization error found in the
+returned three fixes. This is one bounded test-evidence finding, before the
+two independent review lenses, not a new renderer scope.
+
+The frozen round0 scratch remains at29730506c with five changed files. The
+lead read all production/test diffs and the report, checked the final TRX
+counts and mutation failures, and independently rehashed the source. Release
+log says0W0E; focused11/11 and portableGPU312/312. Those results are real but
+do not satisfy the copy-recording part of the original contract.
+
+## Finding F1
+
+BufferCopyDependencySchedule_CoversSameDrainChainsAndEarlierDrainsOnlyForMigrations
+constructs arrays of copy-kind enums and calls a Boolean predicate plus a
+barrier factory. It never records an upload, migration or second drain.
+UploadQueue_RecordsMigrationBarrierImmediatelyBeforeItsDependentCopy checks
+substring order. Its mutation fails on missing text, not an emitted command.
+Therefore same-drain/A-to-B-to-C/across-drain execution is NOT proved. The
+original contract explicitly requires the actual recording body, not a
+separately reconstructed schedule. Do not claim these cases covered solely
+by the current tests' names.
+
+## Bounded correction
+
+Replace that synthetic schedule claim with a test exercising the real
+production native recording body. A test-local Silk native-call interception
+fixture is acceptable, or a narrowly extracted body directly invoked by
+VulkanUploadQueue.Record as allowed in the original contract. Do not add a
+generic recorder, planner, history map, second queue or driver dependency to
+the hermetic lane. Prefer no production change; if extraction is necessary,
+keep the loop/masks/ranges/order identical and mechanically bind its caller.
+
+Capture actual CmdCopyBuffer and CmdPipelineBarrier2 arguments/order for:
+staging-to-migration in one drain, A-to-B-to-C in one drain, then a producer
+in an earlier drain followed by migration. Verify exact source/destination
+handles and offsets/size, dependency immediately before the affected read,
+and the unchanged one trailing draw-consumer barrier. Include independent
+staging/no unnecessary pre-copy barrier and empty/drained-queue behavior.
+No screenshots or physical Vulkan run by the implementer.
+
+Remove the actual production pre-migration barrier again: the new command
+witness must fail at the missing/wrong command sequence, not a source-text
+assertion. Restore raw hashes and rerun the focused and affected portableGPU
+lane plus Release0W0E. Source guards may remain as secondary binding checks.
+No need to repeat the unchanged first three mutations unless their source or
+tests change; preserve the existing raw/TRX proof and clearly scope reuse.
+
+Only the existing #477 test file and, if indispensable, the exact recording
+body in VulkanUploadQueue may change. Other production source hashes remain
+frozen. Record the narrow result/commands and revised hashes; lead verifies
+before independent API/behavior and production/lifetime reviews. No commit,
+client launch, other scratch edits or hidden scope expansion. Both clients
+must be absent before EVERY build/test. The owner's review limit still binds.

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

@@ -4689,6 +4689,18 @@ proceed in an independent fresh scratch while GPU testing continues; no
 changes to GPU/lifetime policy, no fallback cell membership or new owner.
 Implementation/visual PASS remain owed. Lead owns commits/integration.
 
+Dispatch: contract9556c17480d042e2353de3dfddc70fb51a88c60e, fresh clean
+`.claude/worktrees/s5-480-openai-impl`, branch `codex/s5-480-openai-impl`.
+OpenAI Rawls implements and will not review his own repair. A usage-limit
+interruption stopped the initial dispatch before edits; the owner explicitly
+said Continue and both477/480 agents resumed. No new client run occurred.
+
+#477 round0 returned five files, Release0W0E, focused11/11, portableGPU312/312;
+lead read source/TRX and rehashed the change. One test-evidence finding F1:
+the copy-sequence test only inspected text/kind predicates, not actual native
+recording. `477-fix-round1-command-witness.md` binds that narrow correction,
+with no new synchronization behavior or broad framework. Not accepted yet.
+
 Primary ACE spell6150 resolves to3032001C and matches the owner's generation4;
 old route labels "Sawato"/"Rynthid" are unreliable identifiers. Evidence and
 candidate elevated2D31 complex are in479-rynthid-location-evidence.md. A