MCM fast live identity validation — 2026-11-07

Evidence: scratch/runs/20261107-mcm-fast-live-driving/. Prior evidence in 20261107-mcm-bound-input is immutable. This is a continuation of its measured 7.286-second admission / 13.917-second stale refusal, not a driving claim from that run.

tools/mcm-bound-input.cjs now recovers and hashes each immutable CDP WASM script once per actual Worker/session/target and complete script-inventory signature. A changed signature forces retrieval; failed, incomplete or ambiguous inventories refuse. Live process, page object, endpoint, Worker, module object, instance, document and renderer identities are still read, never cached. Independent guest and renderer reads and initial module transfers run together. After them the page, guest instance and owning renderer instance are checked again, along with session membership, detachment and module inventory. Read receipts include operation durations, transferred bytes and cache hits.

The existing 12,000 ms freshness limit and screenshot-start timestamp remain. There are no guest callback, physics, clock, configuration or input-semantic changes. Native release remains independently timed; UNKNOWN and partial input remain sticky closures without replay.

Generated real-Chromium controls use the actual capture/admission/hold helper, real Workers and production renderer physical-state handlers. Trusted Up+Left advances guest and owning renderer counters wholly inside the balanced hold. Nine process/endpoint/Worker/module/document/guest-instance/renderer-object/ renderer-instance/renderer-endpoint replacements reject with zero events. Stale, partial, UNKNOWN, late-ACK and stale-dependent requests close as required; slow capture cannot re-date pixels and a slow sampler cannot prolong release. The generated labeled fixture image was personally inspected. Commands:

node scratch/runs/20261107-mcm-fast-live-driving/generated-browser.cjs
flock /tmp/wine-build.lock bash tools/build.sh

Both pass. A separate real-browser comparison uses generated WASM modules with 2,097,194 bytes apiece. The old chain makes eight module transfers including collector recovery, capture and admission; the new chain makes four and none on admission. New admission's complete live validation is 6.684 ms on this quiet local fixture. Capture-to-dispatch age is 498 ms before, 329 ms after. These are fixture timings, not an explanation of every millisecond on the previous heavy box. The initial large-fixture attempts failed before input because renderer instantiation was not ready; those failures are retained and the readiness condition was corrected. See latency-comparison.json and per-CDP read logs.

One original: freshness repaired, at-speed prerequisite unproved

Repair commit b2be783e344fea122af1a1c9021ea7a4e6663a8c is published on main. One UUID-idempotent box bx_jrgerftp was allocated at 08:35:22.692Z. All 931 staged file pins matched remotely before the sole original launch; 316 local and remote HTTP pins cover registered media and resolved DLLs. The registered runtime/fonts/module/helper closure and exact Git plus relevant uncommitted/new-file patches are retained, without a source or media archive. The complete inventory matched again at termination.

The ordinary route was video-memory OK → RIDER → Single Player Event → Stunt Quarry → Practice/two opponents → T-riffic → default rider/yellow bike → Start. The settled race was personally inspected in Chrome 151.0.7922.108, V8 15.1.206.16, actual guest endpoint 1 / owning renderer endpoint 1 legacy WebGL, ANGLE Intel UHD Graphics 620, Mesa 23.2.1. Actual loaded module: e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467.

The actual first guest module transfer took 72.191 ms and hashing 1.059 ms; the owning renderer's first transfer took 82.547 ms. Subsequent reads show cache hits for both 2,182,493-byte modules. Real post-screenshot validation can still wait seconds for Workers: the settled capture's simultaneous guest and renderer reads took 2,152.748 / 2,149.202 ms; the next capture took 2,701.916 / 2,684.126 ms. This identifies passive Worker responsiveness as a remaining cost, rather than attributing the predecessor's entire 7.286 seconds to byte transfer without evidence. Every read duration is in capture bindings and original-timing-summary.json.

Two Up requests were refused at the initial freshness check with exact-ID, zero-effect replies; neither began live validation or dispatched a key. A third request used a new personally inspected capture and succeeded:

Measurement Result
Screenshot-start timestamp 08:39:20.254Z
Screenshot validation completed 08:39:24.177Z
Admission live identity validation 824.163 ms
Scene age immediately before dispatch 11,068 ms, within 12,000 ms
Up hold / release initiation 08:39:31.323Z / 08:39:33.325Z
First complete in-hold sample 19 ms, slice 8971, flips 24, render completion 212
Second sample 3,694 ms, finishes 08:39:35.139Z, excluded from in-hold evidence

The second sample sees slice 8980, flips 25 and renderer completion 216, but only after its read completes outside the physical hold. One complete in-hold sample cannot establish an in-hold delta. Release remained independent and prompt at 2,002 ms. Console Up dequeue messages are not DirectInput polls or guest-step timestamps. No callback, instruction marker or clock was injected.

Personally compared race-settled, throttle-fresh, throttle-result and final-no-input: upright central RIDER, ravine contours, far flat ridge and near-zero speedometer. Small framing/terrain shifts do not establish an at-speed trajectory; INFERNO/ROCKY motion and the changing red arrow are excluded. No supported speed baseline or steering result is claimed. No Left/Right input was sent. The new measured boundary is a multi-second passive renderer read spanning the bounded hold, with no established at-speed player motion. After measuring it, only passive capture/diagnostic and ordinary quit were performed. No uncertain input was replayed.

One premature image read before its capture download finished failed locally without effects. The first terminal census ran before all processes had been reaped and refused; a later passive inspection confirmed all absent before final census. Those failures remain in evidence.

Terminal state

All 17 commands 0–16 have matching exact-ID replies. Before ordinary quit, physical and async high-bit keys were released and the queue was empty. All 14 recorded processes are absent, both ports refuse, profile and writer locks are absent. All 183 terminal artifacts (8,923,232 bytes) were retrieved and hash-verified before the delegated normal nonforced stop. The box is archived, stopped 08:41:28.893Z, active expiry null, a lifecycle of 6 minutes 6 seconds. Local generated browsers, ports and profiles are clear; free /home stayed above 2 GB. No assigned runtime remains running.

Next work is a generated-first passive sampler that can observe page execution and renderer progress independently without waiting on a blocked renderer, then a separately authorized original driving attempt. This run stops at the measured boundary. No FPS, audio, parity, full27 or public deployment claim. result.json is published last after artifact checks and Git publication.