GLTron low-latency personal WebGL review — 2026-11-04
The sole original WebGL attempt stopped at 5,496 ms capture-to-reviewed action, before a measurement command was queued or a counter armed. No living five-second entered-update interval or personally reviewed after image was obtained. Automatic first-fault closure worked without another command or another actual attempt. The concrete blocker is live image/transport/model handoff latency exceeding the unchanged five-second bound.
Input: b590f8a1c55b3e1d5a04dd24527584628767dc9c and
gltron-atomic-ack-webgl-20261104.md. Evidence lives in
scratch/runs/20261104-gltron-low-latency-webgl-gameplay/.
The accepted software result and historical qualification are not repeated.
Offline latency work
The capture now asks Chrome for lossless PNG with optimizeForSpeed: true.
It preserves the complete visible canvas clip, pixels, hash, original host
capture epoch, page monotonic capture time and module generation. A separate
completion timestamp measures capture duration; neither start clock moves.
Runtime mailbox polling falls from 100 ms to 10 ms. Native ownership, input,
scene, liveness and five-second guards remain unchanged. The prior atomic ACK
and persistent first-fault/blocking/normal-close behavior remain in place.
The current complete CLI executes offline with an explicit plain mocked Puppeteer browser, real HTTP server, real pins, both compiled-module probes, real command mailbox, atomic ACK producer and actual CLI client. Public child spawn methods and the native asynchronous spawn entry are independently forbidden; only exact read-only Git inventory calls are allowed. Fixture permission and the separate private ancestor lock apply only inside that VM; they never provide permission to the actual browser entry.
I personally viewed the fresh 800×600 blue checkerboard returned by the exact
successful capture ACK, using view_image in the same tool invocation. The
next actual CLI measurement action reached the mock counter arm at 3,181 ms
in the first probe. A second complete-entry probe injected the previously
observed 1,875 ms capture delay. Its personally viewed capture reached the
next action at 4,859 ms from original capture start, within the unchanged
5,000 ms guard by 141 ms. The synthetic pixels and mock counter are not live
gameplay evidence, and this narrow margin does not guarantee live timing.
Both fixture runtimes closed normally. No timestamp was refreshed or personal
review claimed by the fixture automation.
A final fresh complete-entry mechanical preflight uses the final source and
fresh pins, with zero Chrome launches. It exercises capture, ACK, balanced
controls, mock measurement and normal cleanup. Release tests now require the
exact stronghold-repaired-entry actor and its latest preceding ownership of
the repaired-entry assignment's exact note path, rejecting the prior
Stronghold/MDK releases, quoted messages, stale/future times, wrong ownership
and reversed record order. Actual permission comes from the real messageboard.
The initial gate demanded the task token on the release itself; this redundant
condition was corrected offline before any Chrome. The real 21:27:19Z ownership
and 21:35:31Z release provide the binding without a repeated permission request.
Sole original WebGL attempt
The final actual entry consumed Stronghold's explicit assignment-bound release
at 21:38:40Z, following its verified owned exits, refused ports, removed
profile and free external lock. MDK had released before Stronghold started.
The original ran from main 5d2ac7e6d8907dcf3c65b16fbc5058907a0dda13 plus
the recorded relevant binary diff. One long-lived external
/tmp/wine-local-runtime.lock covered the runtime. Limits stayed at
720/60/40/180 seconds with full 23:50Z/23:53Z absolute reserves.
I personally viewed fresh main, Game, paused-arena-confirm and active-before captures returned by successful exact ACKs. An earlier capture named paused-confirm still showed the menu; no unpause input used it. The later paused-arena-confirm visibly showed the yellow cycle and "GAME IS PAUSED". Active-before visibly showed the living yellow cycle in the original arena. The next action was measurement, rejected before queuing. No further operator command was issued and no after image is claimed.
The exact owning bridge reported OpenGLHostBridge / WebGLBackend,
matchesOwner: true, with WebGL 1.0 and browser-reported unmasked renderer
ANGLE (Intel, Mesa Intel(R) UHD Graphics 620 (KBL GT2), OpenGL 4.6 (Core Profile) Mesa 23.2.1).
This does not establish physical GPU performance. Ordinary balanced native
S steering had both original acknowledgements direction 3 → 0 and
trail offset 0 → 1 before the fault. The native steering logic was unchanged.
Active-before capture started at 21:40:43.642Z, completed PNG acquisition at 21:40:44.679Z and returned its successful ACK at 21:40:44.697Z. Capture-to-ACK fell to 1,055 ms, compared with the input attempt's 1,875 ms. The remaining handoff consumed 4,441 ms, producing the 5,496 ms rejection. This live handoff exceeded the offline probe's margin. Fast capture helped but did not resolve the full live path; the five-second rule was not weakened, and neither timestamp was refreshed.
client-first-failure.json retains the exact prequeue cause with queued: false; the owner retained that cause, blocked further commands and requested
normal close at 21:40:49.218Z. Chrome PID 1209382, start 187560245, exited
0/signal null at 21:40:49.854Z. Node 1209344, start 187560206, and
flock wrapper 1209343 are absent; the wrapper returned 0 (its start identity
was not recorded). HTTP 33801 and DevTools 33707 refused connections;
the exact owned profile had no processes and was removed. The server closed
and the external flock was verified available. Explicit local heavy release
was posted at 21:41:19Z. No foreign process was signalled.
Validation: 81 focused checks, the final fresh full-CLI mock with independent
spawn prohibition, syntax/diff checks and the full locked build passed. Both
compiled module hashes remain 78a0b1d6… / 99b579cd…; existing media/modules
are referenced by path and SHA only. Git SHA, relevant binary diff and new-note
patch preserve source provenance, without source/media/module snapshots or
archives. Box ID/state/expiry N/A; coordinator retains lifecycle ownership.
There is no WebGL rate result. Any future result must describe ENTERED nonzero-dt gameplay updates per second, not display, completed-update or physical GPU FPS. A further original attempt needs separate authorization and a stronger demonstration that the full personal image handoff fits; this assignment stops after publishing the failure and verified cleanup.