MCM capture/admission latency — Worker observation prerequisite remains

The retained original refusal is reproducible on a small local production Worker route. No safe tooling latency correction was established. Changing Runtime.evaluate from awaitPromise:true to false failed the full-path comparison: each mode admitted one sample and refused one stale sample. The production helper is unchanged; another original driving attempt still needs the Worker observation prerequisite resolved.

Evidence: scratch/runs/20261107-mcm-capture-admission-latency/. Read retained-timing-analysis.json, probe/probe.json, comparison2/comparison.json, and the corresponding evaluation logs. These are focused controls on dx_boids through production index.html, guest Worker and owning legacy D3D WebGL Worker, using local Chrome 151 / SwiftShader. They are not MCM performance results or an accepted fixture matrix rerun.

Where the window goes

The retained original driver first performs a diagnostic state() read, including a serial owning-renderer census. That happens before the capture timestamp. chain.capture() then performs a complete live identity read, timestamps the screenshot request, captures PNG, and performs another complete live read. The latter consumes the existing scene age. Review/handoff follows; admit() checks age, performs another complete live read, and checks age again before dispatch. The pre-capture diagnostic can add acquisition time but cannot explain post-capture age by itself.

readLive() already runs candidate guest/renderer evaluations concurrently, recovers/hashes guest and renderer module bytes concurrently, caches only immutable CDP script bytes keyed by actual Worker/session/target/script set, and concurrently rechecks page/guest/renderer identities. It revalidates live membership, instances, running/failed state, and complete module inventories. There is no remaining serial guest-versus-renderer transfer to remove.

For the retained refreshed original race scenes, post-screenshot validation took 3,586–5,426 ms. In drive-fresh4, the first guest identity evaluation took 5,407 ms and renderer evaluation 4,560 ms concurrently; final guest/renderer rechecks took 17/7 ms. Both module hashes were already cached. Retained request bounds place three throttle proposals at 9,392–13,256, 9,806–13,011, and 10,866–14,905 ms of pixel age. No input was dispatched.

Measured rejected correction

The run-only probe schedules a bounded 1,800-ms JavaScript busy task in each actual production Worker, then reads the same identity expression in both CDP modes. Renderer replies took 1,703/1,715 ms with promise waiting and 1,703/1,704 ms without it. Thus disabling promise waiting cannot make a queued observation bypass a busy Worker task. Guest probe samples varied from 4,094–10,780 ms; these include the production workload and deliberately added task and are not isolated MCM timings.

A separate alternating comparison used the full unchanged chain with only its injected evaluator selecting the CDP promise flag. It added no busy tasks, input, fake clock, identity cache, or age adjustment:

Mode Full capture ms Post-screenshot live read ms Admission
Existing await 11,244 2,461 accepted at age 4,941 ms
Candidate synchronous 12,808 5,204 stale, zero dispatch
Existing await 18,499 4,585 stale, zero dispatch
Candidate synchronous 16,949 4,662 accepted at age 9,428 ms

The first and last admissions were observation callbacks only. No native game input was sent in this comparison. All samples used complete live checks; no passing-after latency fix is claimed. The actual reviewed reference is comparison2/review.png: a dark Boids scene with blue wire grid and small green/white shapes. review-approved.json corrects the overbroad terrain/sky template wording in the comparison callback's review string.

Concrete next step

The established blocker is multi-second service latency for fresh CDP Worker identity evaluations, particularly the first guest read after a screenshot. The exact internal wait is not established by these receipts. Neither the original natural renderer sync-poll probe nor this host-side timing identifies which guest slice, render fence, or debugger scheduling interval owns that delay.

The next bounded task should correlate requested/replied CDP reads with read-only start/end observations around the production guest slice handler (lib/guest-worker.js, uninterrupted ex.run and local-wait loop) and its render-fence waits, plus owning-renderer task/fence completion. Use the same small production route first. A resulting task-yield change would need to preserve guest order, existing completion fences and input semantics, with its own failing-before/passing-after controls. Do not substitute cached live identities or debugger pausing for that investigation. Another original or a held-risk lane is not authorized by this note.

Preservation and custody

tools/mcm-bound-input.cjs remains SHA-256 9df1455f51f91d76e17890ff4d90afe9264759d3e30e224a9a7569e29b1f8062. The 12,000-ms gate, every existing identity field and recheck, sticky UNKNOWN, numbered exact ACK transport and independent release implementation are unchanged. No original game, boat, remote runtime, source/media archive, emulator/configuration/clock change, synthetic ACK, or accepted fixture matrix was used.

The run records exact Git source identity and relevant binary patches, including the pre-existing untracked lib/wininit-boot.js. CDP recovered the executing module with SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. The locked flock /tmp/wine-build.lock bash tools/build.sh passed. Run-only syntax/startup failures are retained separately; they dispatched no native input. One invalid 60,000-ms requester timeout was rejected before queuing; that sticky lane was closed without replay. The final checked harness uses the existing 55,000-ms maximum.

focused4/controls.json passes actual native ArrowUp admission at age 11,302 ms, exact request/consumer ACK, trusted DOWN/UP and independent release at 100 ms despite a 350-ms sampler delay. Physical keys are all released. The same capture naturally ages to 12,014 ms and refuses with zero effects. Replacing the actual page owner and removing the host's recovered module inventory each refuse before dispatch and close their chains. Missing ACK closes its separate transport; no dependent command is created and no ACK is fabricated. These controls establish unchanged safety behavior, not a latency repair.

All seven local control Chromes are closed, their server/debug ports are closed, profiles and writer locks absent. Six exit receipts record code 0; the requester-prequeue failure used an external ordinary Browser.close followed by termination of the idle driver, with process/port absence verified separately. Duplicate identical recovered modules were reduced to one verified unique module plus hash/length references. Sampled free space remained above 2 GB; continuous telemetry is not claimed. No test box was assigned or manipulated: box ID/state/expiry are not applicable. Exact final file verification is in evidence-verification.json and cleanup in terminal-cleanup-verification.json.