Q3 owner-capture causal investigation — 2026-11-08 22:03Z
STOP: causal operation not established; no production repair justified. The retained trace distinguishes two different stalls. The 11,152 ms held sample awaited the render Worker, whereas the final incomplete capture awaited a page evaluation. The page-owner expression contains no explicit asynchronous operation, and its actual production accessors do not enqueue guest work. Removing the final owner check, returning stale pixels, or treating the render delay as the cause of the later page delay would not be supported by this evidence.
Run evidence: scratch/runs/20261108-q3-owner-capture-latency-2203/. Base/current source at inspection: Git 7a38ceeeba9ad1380cf6ba4886389db5953ce795, shared main; no production/test files edited. Read the retained original report, CLAUDE.md, current AGENTS instructions and board ownership before work. This task was entirely offline: no Chrome/browser, original/game, remote runtime, input, box, backend/configuration/clock/seed changes, or other workers.
Exact retained chain
browser-run.cjs executes admission → bounded physical hold with an independent release timer → passive settle → diagnostic state → bound-input.cjs capture. Capture requires complete live binding before screenshot and again afterward, including independent CPU/render module byte hashes. worker-observe.js forwards original operations once and associates completed native GL draws/presentation with the actual bridge/adapter/owner/context; producer-ready.cjs selects those owning receipts. None of these readiness rules was changed or replaced.
For g-open-difficulty, trace request 403 (page.owner, span 2665) completed in 1 ms at 03:39:14.838Z. Request 404 (renderer.owner, span 2667) then waited 11,152 ms on the d3d-render-worker target, completing at 03:39:25.991Z. This was the last held sample. Independent hold.release completed at 03:39:14.991Z, taking 3 ms; the sample did not prolong the physical hold, but the command joined the sample before settling and capturing.
Post-action diagnostic state completed in 17 ms. The capture's first page-owner read, request 413/span 2695, completed in 1 ms. Guest-instance read and CPU module retrieval completed; renderer-owner and producer reads completed in 1 ms each; render module retrieval completed in 93 ms (96 ms surrounding span). The final page-reference recheck, request 419/span 2707, began at 03:39:27.776Z. At the command-work deadline snapshot, 03:39:35.431Z, it remained pending for 7,655 ms. Later passive page requests 420 and 421 also remained pending. There was no screenshot request for this capture, no complete capture, and no successful final ACK. Runtime close rejected request 419 at 03:39:36.721Z; this is a closure error, not a late successful observation. Existing sticky UNKNOWN/input closure/no replay remain correct.
analysis.json.gz machine-checks these spans, release timing and missing final completion, and records existing PATH/SHA-256/size references to the retained trace, helpers, patches, module, media and relevant current production dependencies. No source/media/module copies were made.
Production dataflow control and limits
PAGE_READ performs document-scoped WeakMap identity reads, running-app/port enumeration, cached slice/stat reads and four physical key-map peeks. Production GuestThreadHost.d3dStats returns link.d3dStats || null; WorkerLink refreshes that cache on completed slice replies. renderer-input.js:peekKeyDownState reads local physical/async maps. Neither accessor awaits a guest export, render batch, mutex or RPC. The page owner expression returns an ordinary object despite CDP's awaitPromise:true option.
probe.cjs runs the exact retained expression, SHA-256 f9e87085a25e34c2388f0abc7fff0a1a45aa3edb4ba645f16f5517ac1e3b1043, in a VM containing an actual current WineAssembly instance and production GuestThreadHost backed by an actual worker_threads.Worker running lib/guest-worker.js. While a genuine get_eip export request was pending, 100 synchronous owner reads added zero Worker requests and did not join that pending request. All returned ordinary values; the queued export subsequently settled, and production Worker retirement was awaited. The maximum local VM sample was about 1.48 ms. The page registry, canvas and key map were synthetic; no CDP controls, real rendering or real input were used. This is a negative dataflow control, not an original-latency reproduction or oldFAIL/newPASS repair test.
The control uses current module build/wine-assembly.wasm, SHA-256 4b45bda53a90b665896cf3b85d9a157b7c3703563c88b6666e4ee2e4b346bcc3, 2,184,500 bytes. The original used retained module f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121, 2,182,500 bytes, and source commit 6138f2546f6d8b7b1d2f2ebaf8ff274acbb67297 plus its retained patches. These are explicitly different paths/builds; the local timing cannot characterize the original Intel/Mesa provider.
The render dependency can synchronously execute GLRenderWorker.execute → OpenGLHostBridge.replay → native backend operations and flush pending draws, then transfer a presentation bitmap. Page-side callbacks/compositing run separately. The retained CDP await trace does not record which operation was running during either stall. Missing causal dataflow: page evaluation entry versus queue residence and the active page JS task/stack during 03:39:27.776–35.431Z; render batch/opcode begin/end and GPU synchronization timing during 03:39:14.839–25.991Z; corresponding frame-consumer/composite task timing. GC, scheduling starvation, synchronous native/GPU work and page callbacks cannot be distinguished from these receipts. NodeWorker access to the cached fields does not supply those missing browser/GPU observations.
No production change, shared host/census edit or persistent test was made; therefore no code build gate was required. The focused offline trace assertions and actual NodeWorker control passed. A future causal investigation needs those missing execution boundaries before selecting a consumer repair; this task authorizes no original retry. Existing 20 s work/2 s ACK/50 s request/cleanup reserves, exact freshness/binding and complete capture semantics remain. SoftwareGL current torch acceptance is preserved; WebGL gameplay, Full27, parity, audio and FPS remain unqualified/open.
Box NONE, state NOT_CREATED, expiry NULL, stop receipt NA. Evidence uses gzip from first byte for logs, independently reread original/compressed SHA-256/size and gzip CRC, Git SHA plus small owned-file patches, and existing PATH/HASH media/module references. The observable independent 3,600 s board-watch timeout was prearmed before the offline probe and is explicitly stopped at completion. Only sampled disk checks are claimed; /home samples exceeded 1,500,000,000 bytes. No old run deletion or peer WIP capture.