Forsaken owning-guest observation latency, 8 November 2026

The owning-guest bottleneck remains unresolved. A run-local composition that overlaps fresh module-byte retrieval with renderer observations still reaches the unchanged command work deadline. No fresh original or box was started; current look and further corridor navigation remain unverified. Current movement/fire evidence is unchanged. No emulator mismatch was demonstrated.

Evidence: scratch/runs/20261108-forsaken-owner-observation-latency/. Read both the early-look and current-native-progress notes and their actual bound helpers, numbered transport, release helper, production fixture and raw original trace. The previous 4,800 ms / 17.030 s controlled result does not qualify the failed original latency.

Exact original requests

original-request-ledger.json records the original trace hash, request ancestry, completion/error and deadline records. All three requests below target guest 2DDE99F3197B2EC20E53699B8EA311E2, session 2E0BAB60F2F8C3FA6CA195B686E1A2E5, in number 5 open-singleplayer:

Phase Request Actual observation
Admission #533 Runtime.evaluate Fresh endpoint, slot, actual instance token and GPU-enable value; 8,085 ms, completed.
Admission #536 Debugger.getScriptSource Fresh actual owning module bytes, SHA-256 checked; 170 ms, completed.
Capture before #575 Runtime.evaluate The same expression, independently submitted for this boundary; 9,092 ms pending at work deadline.

Expression SHA-256 is 93a59ecbd66836cd6b985e4b1652276aa7c88e2f0c6c368e1b13efbe4b5741c4. There is no second hidden owning-guest state evaluation: the capture-before read supplies that capture's CPU state. The subsequent module-byte request, screenshot and separate after-image proof were never reached in this command. Earlier completed captures have independent before/after observations.

Admission, including the independently released Enter hold and sampler drain, lasted 10,870 ms. DOWN began 06:40:34.132 UTC, UP began 06:40:36.132; the sampler drained at 06:40:36.512. After the 1,001 ms settle, capture-before started 06:40:37.516. The work deadline fired 06:40:46.610. Browser closure rejected #575 after 10,989 ms; that is a cancellation/error duration, not a completed service response. The command drained after 22,862 ms without an ACK. It remains the first UNKNOWN, with no replay.

This distinguishes the measured CDP request wait from a proved internal cause. The trace has no stack or guest-run/RPC wait attribution for #575. Production guest-worker.js has synchronous ex.run and local wait paths, but their existence does not identify which path delayed this particular evaluation. Unrelated-thread selection was already refuted. Byte transfer is much smaller than the measured instance-evaluation waits.

Tested composition and limits

The run-local candidate starts the guest module-byte request, observes the renderer and its bytes, then joins the guest byte result before validating the final page references. It changes no owner comparisons, uses no module-hash or mutable identity cache, and does not combine admission with capture-before or capture-before with capture-after. Two stable readiness reads, 120-second freshness, reviewed image hash, exact ACK, sticky UNKNOWN and independent UP remain unchanged. The candidate is retained as tested investigation input, not installed as a production remedy.

local-serial/ and local-overlap/ run the actual browser driver, numbered transport, production guest/render Workers, public COM fixture, native trusted Enter, admission, hold, settle, state, screenshot and after-image chain. The ordinary baseline has an exact success ACK in each. For the diagnostic command, host receipt delays are 8,085 ms for admission and 9,092 ms for capture-before; the after read has no injected delay. A slow passive sample models the independent release followed by sampler drain. These are controlled receipt delays, not an assertion that the original Worker ran those durations.

This is deliberately an optimistic censored-prefix control: #575 had not returned at 9,092 ms in the original, and its real completion remains unknown. Returning at that lower bound cannot establish equivalence with the original. It can expose a composition that already fails under that favorable condition.

Composition Already-started command drain Numbered result
Serial current helper 21,839 ms Work deadline, UNKNOWN, no success ACK
Overlapping fresh byte reads 21,373 ms Work deadline, UNKNOWN, no success ACK

comparison.json retains phase/deadline records. The single 466 ms difference is not a benchmark or validated original-latency improvement. Both controls release Enter independently (1,999 ms serial, 1,995 ms overlap), dispatch no later input after UNKNOWN and drain their already-started work. A completed PNG after the deadline does not restore the missing ACK or authorize input. WORK=20 s plus declared settle, CDP=20 s, ACK minimum 50 s and cleanup reserve 30 s were unchanged. No scheduler, clock or guest implementation was patched.

The production module actually loaded locally is cc438f727a39e39d1016045d37ae95f60cf8f9793e6b117e14beabb80c27e07a, from the shared checkout's current build. It differs from the original's 7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf. This is an additional qualification limit; no original-module equivalence is claimed. Source SHA, relevant uncommitted diff, file pins and actual loaded modules are preserved. Git remains the source archive.

Identity tests and closure

The overlap helper's actual production Worker negative tests pass for missing renderer readiness, process replacement before binding, endpoint/render-owner/ process-owner replacement, all exact reviewed fields, replacement inside the observation and after screenshot, stale image with zero input, and changed returned module bytes. Additional controls replace the actual Worker instance between observations and during capture, and instantiate a different actual valid module in the generated Worker. Each is rejected; these controls alter only the generated fixture, never an original guest. The positive trusted native key route and release pass. No cached-byte equivalence claim is made.

Independent slow-sampler and injected-UNKNOWN release tests pass; all eight capture-deadline/ACK/freshness regressions pass. An initial nonexistent test filename was corrected and its failed invocation retained. Generated caption pixels were inspected only as helper evidence. Three local browsers exit zero; owned PIDs, ports, profiles, writer locks and watcher are drained. Evidence hash/size verification and served-byte checks are retained. This commit changes only this note, so no new locked build is claimed.

The remaining request is precise: a fresh actual instance/endpoint/slot witness for each required observation still blocks at the owning guest, especially original #575. A smaller byte-transfer path does not remove that wait. Dropping it or reusing a prior mutable witness would weaken admission or capture proof. Further work needs a defensible fresh identity observation mechanism and matched production qualification, or direct attribution of this request's wait. No unchanged original retry is justified by these results.

Box none, state not allocated, expiry not applicable, stop receipt not applicable. No remote runtime or watchdog was needed because the local qualification gate failed. Disk stayed above 5 GB. No old box was resumed or deleted; no deployment or archive of the source/media tree was made. Release this task's slot. Full 27, parity, audio, unmeasured FPS, current look and distinct further navigation remain open.