Recoil passive probe timing repair, 2026-11-08

Run: scratch/runs/20261108-recoil-probe-timing-objective/.

Tested substantive integration repair: the owned native-CDP admission helper now overlaps independent guest and owning-renderer identity/complete-byte read lanes, drains both lanes on failure, and joins them before the final page-owner check. Native CDP tests through the actual compound HUD probe reduced mean probe time 364.87 → 271.56 ms without injected delays, and 643.65 → 413.22 ms with an 80 ms generated delay on each worker identity evaluation. Three trials per configuration alternate baseline/fixed order. These are generated fixture transport measurements, not native gameplay performance or an assurance that the original's deadline will now pass.

No fresh box or original game was allocated. Ordinary M, its visible consequence, and reviewed objective movement remain unobserved. This task finishes with the tested integration repair, rather than spending another original on an unchanged late-probe result. Full27, parity, audio and FPS remain open.

Retained evidence first

Read the handoff at 67aed1955 and its retained helpers, manifest, compound receipt, observer records and retrieval ledger before changing the helper. retained-costs.json records the exact receipt hash and measured intervals:

Probe Final pixel start through inspection Previous inspection to pixel start
MENU baseline 46 ms —
MENU 1 25 ms 443 ms
MENU 2 19 ms 343 ms
MENU 3 49 ms 338 ms
Late living HUD candidate 190 ms 3,569 ms

The last interval includes the 25 ms poll delay and both complete live reads before pixels. It does not measure decode time alone. The retained original has no CDP subphase timestamps: complete-byte transport, base64 decoding/hash, identity evaluation and page capture cannot be individually attributed within that historical gap. The final image starts 285 ms after the unchanged deadline; the retained independent HUD images at 11:04:58.132Z and 11:04:59.064Z establish visible Resume but do not substitute for admission. No historical causal claim is made that serialization alone explains the late candidate.

The concrete source inefficiency is independently demonstrable: each old readLive awaited the guest identity, full guest byte response/hash, renderer identity, and full renderer byte response/hash in sequence, although the two targets have independent CDP sessions. The repaired helper starts the two lanes together after reading the page owner. Promise.allSettled drains both before success or failure. It also explicitly rejects renderer detach during the byte read. There is no cached-byte shortcut or guest/runtime change.

Measured subphases and unchanged admission

native-results.json retains bounded per-read spans and native CDP request serialization, response/buffering, JSON parsing, full byte count, base64/hash, capture and real HUD decoder costs. Each generated module is 2,182,704 bytes, comparable to the retained original's 2,182,663-byte module, but instantiates only a generated worker fixture. Its custom section is generated in memory; no original module, guest assets or media are substituted into an original game.

Mean fixed-helper costs without injected delay are: page-before 3.22 ms, guest identity 4.00 ms, renderer identity 3.30 ms, guest complete-byte response 82.11 ms, renderer complete-byte response 89.44 ms, guest base64/hash 2.27 ms, renderer base64/hash 2.01 ms, first final-page check 5.74 ms, final co-located page/pixels 22.15 ms, HUD decode 20.74 ms. Module response JSON parsing averages 2.87 ms and is included in response latency, not additive to it. Concurrent spans overlap and must not be summed as probe wall time. Final pixel-to-decode ages are 55/45/30 ms without delay, 32/45/21 ms with delay. Initial direct readHud native measurements are separately retained in native-results-readHud.json; the second fixture run added the actual compound HUD integration and decode instrumentation.

Every tested probe retrieves four complete uncached module responses, two from the actual guest target and two from the actual owning renderer target. Exact page objects, worker instance, targets, module hashes, renderer epoch and final co-located page-owner references/pixels remain required. The unchanged 250 ms freshness check runs after decode and again after saving. The native timer-mask temporal pair remains required. No input deadline, freshness bound, or identity requirement was widened.

focused-results.json verifies full read counts, material overlap and seven failure cases: changed guest bytes, renderer bytes, guest instance, renderer epoch, final page owner, renderer detach and guest transport error. Both lanes are drained before refusal. The real retained MENU and advancing HUD images exercise the unchanged decoder: first living candidate, identical timer, stale candidate, stale pair seed and slow-save cases cannot admit gameplay. gate-results.json tests the unchanged compound runner with gate4500/M500: the generated positive dispatch releases M after 501 ms despite a slow sampler; a candidate completing after the immutable gate produces zero gameplay DOWNs and independent Escape only 5 ms after its due time. This is generated dispatch evidence, not trusted original-game consumption. No M/C/arrows or attack was sent to an original.

Integration, recovery and lifecycle

The owned browser-position.cjs loads the repaired bound-input.cjs and instrumented compound-v4.cjs, logging bounded per-read timing rows. The 822-provider composed manifest explicitly pins sixteen owned browser helpers; integration-preflight.json verifies the real hermetic preloader and decoder, accepted 500 ms Escape and visible-pointer helper with an intentionally wrong ambient PIN_FILE overridden. No browser is constructed by that preflight. closure.json rechecks all 806 inherited source/media/helper providers, all 295 retained artifact references, twelve accepted controls and six prior custody/ review records. Controls are reused by exact hash, without broad control reruns.

The accepted reap-owned.py, actual locked-launch-fixed.py wrapper and artifact-custody.py collector are retained unchanged with their prior focused tests. Their earlier real zombie and mutable-alias failures remain in the prior run and are explicitly referenced by the closure. They are not relabeled as clean original success. An initial local test mistakenly assumed a PNG existed at its remote relative name; the ENOENT failure is retained. The corrected test resolves the existing retrieval ledger's exact local path and checks its hash.

Two generated local Chrome fixtures used /tmp/wine-runtime.lock only while running/draining. The observable independent timeout watchdog was armed before Chrome launch with a 90 s cap and 5 s kill fallback. The subreaper supervisor records all observed descendant PID/start identities, reaps adopted children, and verifies their absence, closed local HTTP port and absent profile before releasing the lock. Final native fixture exit is 0; all 18 recorded identities are absent. The earlier fixture's 18 identities are separately rechecked absent. No local lock was held for closure checks, analysis or publication.

There is no new box ID, expiry or stop request. The inherited original is only referenced through retained evidence: bx_afvq3a5t, archived, expiry null, normal stop stop_e067c3dd2575e338, completed 11:07:14.264Z, forced false, error null. No old box was resumed/deleted, no deployment or subworker was used.

Source identity is recorded in source-head.txt; selected original source is 49874a81c5ae180c20a15c6bb169d9041f028cae. recovery.json references the exact prior relevant binary diff, owned untracked source patch and existing media path/hash recipe. owned-inputs.patch records only small owned helper/test inputs, with timing-repair.patch isolating the two behavioral changes and the browser timing hook. No source archive, media copy or peer WIP is included. Evidence results are retained by the accepted immutable content-addressed collector and independently reread. No production file changed or full-build claim is made; this publishes the tested run-helper integration and recovery note.

Reproduce cheaply with verify-closure.py, prepare-integration.py, node test-probe.cjs and node test-gate.cjs in this run. The bounded native fixture is python3 local-runtime.py from the repository root. A future original must recheck pins, identify its UUID-idempotent box and prearm the independent 25-minute watchdog, hold its own remote startup/runtime lock, visually review Resume again, then pass this unchanged gate and observe ordinary M's consequence before reviewed movement. Attack still requires a reviewed living target.