Original GLTron live gameplay interval — 2026-11-04

No gameplay FPS was measured. The one software launch reached personally viewed active gameplay, then failed the runner's capture-age guard before arming. WebGL was not launched and software was not retried.

Evidence: scratch/runs/20261104-gltron-live-gameplay-interval/. Input: 477c16e1870085bd0562554ea69eb3b74324ad26 and docs/re-notes/gltron-complete-boot-gameplay-fps-20261104.md.

Operator sequence designed before Chrome

Wait for the explicit MDK2 continuous-construction-onward local heavy release. Hold one external /tmp/wine-local-runtime.lock across one software launch followed by one WebGL launch. Stop the pair at the first fault; no backend retry or separate diagnostic launch. Reserve 360 combined browser seconds, 60 bootstrap seconds, 60 per command, and 180 cleanup seconds, with absolute Chrome/cleanup cutoffs of 23:30Z/23:33Z.

For each backend, capture and personally view startup and each original menu before the next dependent native key. Use balanced Return keys for Game, arena and unpause only as the actual reviewed scene permits. Capture fresh active gameplay with a named image; personally view it and immediately submit the measurement command. The browser records the capture-start monotonic time, callback table and owning generation. Actual arming requires that exact named active-gameplay capture to be at most five seconds old, the same generation, and the current translated callback table to still be 0x42a200. A missed review deadline is a first fault, not permission to retry arming.

The marker is the production nonzero-dt branch at 0x4051a9, pace zero. Sample the owning main-thread-1 lifetime counter every 250 ms for 5.5 seconds, checking scene, translated pointers, ownership, generation, tiers, visibility, frozen state and foreign counters. Preserve monotonic getter brackets and count deltas, including reset/wrap checks and a minimum five-second interval after read uncertainty. Preserve the exact first rejected callback translation at arming or in the interval; attempt an owning-context capture before normal closure. Personally view the resulting after capture before accepting a rate. Ordinary balanced steering may maintain a live round when the viewed scene requires it; no guest code/state/config/pacing edits, freezing or host steps.

Backend attribution requires the actual owning GL bridge/context evidence. The measured rate includes marker dispatch, guest tick reads, block/uop boundary enforcement, decoded-cache recycle and periodic observations; no uninstrumented overhead percentage is established. Callback-table sampling cannot detect a scene transition wholly inside one guest slice. Counts mean entered nonzero-dt gameplay updates, not completed updates, physics substeps or presentation.

Outcome

The required mdk2-continuous-onward LOCAL HEAVY RELEASE arrived at 19:54:54Z. One external local flock covered the attempted pair. The production boot graph passed with 654 pins, 92 original media files, all 365 prior HTTP records and 1,010 offline routes. The previous 81-file seal (2,453,581 bytes) is unchanged. The exact prelaunch Git commit, relevant binary diff, new-file patches, and existing media/module path/hash references are retained in the run. No source, media or module snapshots were made.

Personally viewed software/startup.png (original main menu), game-menu.png, arena.png (GAME IS PAUSED) and gameplay-before.png (yellow cycle and new trail, without the pause message). Each dependent native Return followed the preceding view; all three keydown/keyup pairs were trusted and targeted at screen. The active capture completed at 19:56:14.306Z. Software ownership was established by the matching main bridge and OpenGLHostBridge context with software: true, software enabled 1 and no WebGL context. The visible D3DIM dropdown does not identify this GL backend.

The actual owning module was canonical SHA-256 78a0b1d61b23428a1ae909d739440ae0de059bbbd84a69a2ad3e4c37471ad664, page-main generation 1, main thread 1, chain 0, uop 1, debug-tier guard 0. Frozen was false and host steps remained zero. All 234 served responses matched the production pins, including all 92 original media files.

At approximately 19:56:20Z the measurement request rejected with STALE_GAMEPLAY_REVIEW: the capture-start-to-arm age was exactly 6335.849999904633 ms, above the runner's 5000 ms limit. The exact live callback remained 0x42a200, translated from guest VA 0x42b230 to offset 250416. Marker and count were both zero; no setter or observation interval ran. This failure was the capture/view/command handoff exceeding the chosen freshness budget, not a callback mismatch or a production counter failure.

The retained rejection capture software/counter-rejected.png was personally viewed and shows PLAYER 1 CRASHED and debris. It is a failure capture, not an after-interval gameplay review. A table match alone therefore does not prove a live player. The exact collision time and cause were not measured; this does not establish the cause of the previous run's changed callback. There is no software rate, WebGL rate, paired comparison or post-interval acceptance. The pair stopped at this first fault.

Tested offline changes and prospective handoff repair

Before Chrome, the actual old arming producer failed the new stale-review test. The revised producer passed the five wrong-age/name/generation/scene negatives and retained the exact rejection. Both compiled modules executed the original dt branch under all four chaining/uop settings, plus actual translated pre-arm and interval rejection guards for zero, paused and adjacent foreign callback tables. Invalid game-pointer translation retains the callback read that preceded it. Reset, wrap, monotonic bracket, owner, generation, tier, visibility/frozen and bounded-storage tests passed. The existing production marker/uop tests passed 50 checks per module. The locked full build passed.

After the runtime first fault, the actual shot producer was extracted and changed to return the exact capture path, hash and capture epoch in the command ACK. A single operator tool invocation can now submit the capture command and VIEW its returned image, then submit the reviewed action on the next turn. This removes the avoidable ACK-only turn followed by a separate image-path handoff used in this attempt. The old producer failed this new test; the new producer passes capture-start -> pixels -> VIEW ordering and rejects overwrite and foreign paths. This is an offline repair of the handoff, not proof that future personal review will fit five seconds or keep a round alive. The five-second deadline and actual active-table guard remain unchanged. Ordinary balanced steering may still be needed after a reviewed live scene, before an interval. No second original tested the repair.

Final focused validation passes 55 checks including both modules; final syntax, diff checks and flock /tmp/wine-build.lock bash tools/build.sh pass. The final source patch records the offline repair separately from the prelaunch patch.

Closure and next dependency

Chrome PID 1031770, start ticks 186935344, actually exited 0 with signal null after the direct normal Browser.close request. Node PID 1031753, start ticks 186935310, and the wrapper exited; both owned PIDs are absent. HTTP port 37609 refused connections, the server reported not listening, and the correct local flock was available. Combined Chrome lifetime including closure was 36.596 seconds. Explicit local heavy release was posted at 19:57:15Z. The owned profile was removed only after closure verification.

Stop here. The next dependency is a separately authorized original pair with the direct capture-to-VIEW relay, an active live round maintained through personal review and at least five seconds of guarded counter observation, and personally reviewed active gameplay afterward. No retry is authorized here.

Box ID/state/expiry: N/A. Coordinator owns future boxes and shutdown. Normal closure of this worker's local Chrome is delegated; retain PID/start identity, actual zero exit with null signal, Node/server/port/flock closure evidence. Git SHA plus relevant binary diffs and owned new-file patches is the source archive. Media and modules remain at existing paths with SHA-256 references.