MCM independent passive sampling — 2026-11-07
Evidence: scratch/runs/20261107-mcm-independent-sampling/. Historical accepted
MCM driving remains separate; the immediately preceding fast-live run admitted
Up but had only one complete in-hold sample and proved neither driving nor
steering.
tools/mcm-bound-input.cjs separates page-reported guest progress and actual
owning-renderer progress into individually timed streams. Each stream has at
most one outstanding read, runs independently, and stops scheduling at physical
release. The renderer stream reads only the actual owner, verifies its epoch,
module inventory and live session, and checks page ownership again after its
reply. Page samples retain process, endpoint, Worker/module-object and document
checks plus live Worker/session/module-inventory validation. These counters are
passive execution observations, not DirectInput polls or simulation steps.
The hold helper accepts renderSample alongside sample; release remains an
independent timer. Every observation has its own start, finish and stream, and
is excluded unless wholly inside the physical hold. Slow reads finish after
release without extending it. Capture-start dating, the 12-second freshness
limit, admission validation and UNKNOWN/partial sticky closure are unchanged.
Actual Chromium qualification passed with production physical-state handlers:
- Normal trusted Up+Left: 701 ms release; seven whole page samples with guest sequence 15→47, six whole owning-renderer samples with completion 15→44.
- Deliberately busy actual renderer (450 ms evaluation): 261 ms release; three whole page samples, sequence 60→71; zero whole renderer samples. The late renderer observation is retained and explicitly excluded.
- Page renderer-owner replacement during an outstanding slow renderer read: rejected, balanced Up+Right release, closed admission chain.
- Nine pre-input owner/instance/module/document replacement cases: zero input effects. Stale, slow capture, slow sampler, partial, UNKNOWN and late-ACK regressions pass. Generated labeled fixture personally inspected.
Commands passed:
node scratch/runs/20261107-mcm-independent-sampling/generated-browser.cjs
flock /tmp/wine-build.lock bash tools/build.sh
One fresh original attempt follows publication of this tested sampler repair. No original driving or steering result is claimed by these generated controls. No FPS, audio or parity claim.
Original result: guest sampling repaired; at-speed prerequisite remains
Sampler repair 7715eab3fe3cbc4f429fb850ccc0ccd58e5796ed was published before
allocating the sole UUID-idempotent box bx_z3xstz2x at 08:47:52.767Z. All
930 staged runtime/original file pins matched before launch and at termination;
315 HTTP resolver pins matched. Git SHA, relevant uncommitted hunks and the
owned small run-input patches are retained. No source/media archive was made.
The 311 registered original files remain separately pinned, not copied into
run evidence. All runtime pins still match locally after retrieval.
Personally inspected the ordinary route: video-memory OK → RIDER → Single
Player Event → Stunt Quarry → Practice/two opponents → T-riffic → default
rider/yellow bike → Start. Chrome 151.0.7922.108 / V8 15.1.206.16 used actual
registered mcm, process 1000, guest endpoint 1 / slot 0 and guest target
2CB000C5A8049B998F44AF6550F7D9F2. Owning renderer target
2DC985FBF7D2C7426DBAD9C7DC839F33, endpoint 1, reports legacy WebGL, ANGLE
Intel UHD Graphics 620 / Mesa 23.2.1, no fallback. Both actual recovered live
modules hash to
e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467.
Four throttle requests had exact-ID, proven zero-effect stale refusals; each was followed only by a fresh capture. Local request handoff was shortened, without changing capture dating, admission checks or freshness. A premature fetch after the first refusal found no result capture, failed locally and remains recorded; it did not dispatch input or replay an uncertain request.
One ordinary Up hold was admitted, at 8,425 ms scene age after 1,954.239 ms live identity validation. Physical interval: 08:52:21.329–08:52:23.330Z, 2,001 ms; release was independent of diagnostics. The repaired sampler retained 20 whole page samples (1–3 ms each), all showing physical Up high and other arrows released. Two distinct whole observations establish guest sequence 7893→7901; it then stays 7901 through the remaining whole samples. These are actual page-reported guest execution observations, not logical frames, DirectInput polls or guest-step timestamps.
The owning renderer has exactly one whole observation: completion 296,
20 ms read. Its next read starts 123 ms into the hold and takes 2,707 ms,
finishing at 08:52:24.159Z after release with completion 298. It is excluded.
No renderer delta is inferred from that one whole observation. Crucially, this
late renderer reply no longer delays or erases the independent page evidence.
See original-throttle-summary.json and the full admission/hold/JSONL receipts.
Personally compared race-settled, quick-fresh, quick-response and
passive-after-throttle: RIDER's helmet/body rises transiently in the first
response and returns to the centered upright pose; near-ground textures and
ravine framing shift slightly, with the same distant plateau/ravine ahead and
near-zero gauge. That establishes a visible transient response but not a
sustained forward displacement or an at-speed terrain trajectory. FELINE,
INFERNO and the changing red HUD arrow are excluded. No Left/Right input was
sent; there was no held-risk retry or UNKNOWN replay, no forced state, injected
callback, clock or guest marker.
The first remaining blocker is the unestablished at-speed player trajectory: real renderer responsiveness still spans a bounded hold, despite independently observable guest progress. Driving and steering remain unproved for this run; accepted historical driving is preserved separately. Next work must address that concrete responsiveness/at-speed boundary before another original steering attempt. No FPS, audio, parity or historical-27 claim.
Terminal state
All 20 commands (0–19) have matching exact-ID replies. Physical and async
high-bit keys were released and input queue empty before ordinary quit. All
14 recorded processes are absent, both ports refuse, profile and writer lock
are absent. All 201 terminal artifacts / 9,797,865 bytes were retrieved and
SHA-256 verified before the explicitly delegated normal nonforced stop.
bx_z3xstz2x is archived, stopped 08:53:48.148Z, stop completed with
forced:false, active expiry null; lifecycle 5 minutes 55 seconds, within
the 30-minute bound. No assigned runtime remains running. Generated local
Chrome PID/ports/profile are clear. Free /home remained above 2 GB.
result.json is written last after this note's publication and artifact checks.