GLTron browser capture: idle-publication blocker, 2026-11-14
No fresh original was launched. The historical Chrome screenshot stall was not reproduced, and the replacement candidate fails on an idle unchanged visible scene. Two live intact-cycle corners, a natural round outcome and native menu Quit remain missing browser proof. This publishes the report only. The three adopted code files remain uncommitted, unapproved WIP with exact before and candidate patches.
Evidence: scratch/runs/20261114-gltron-browser-capture-live/result.json.
original-decision.json records the refusal. No old run was reopened,
rebound or replayed. No guest input or emulator patch was made.
Capture candidate and controls
The fixed deadline is in connection() in the browser controller: every CDP
request receives 3,000 ms, including Page.captureScreenshot. The previous
failure establishes expiration of that deadline, but has no trace resolving
which Chrome compositor/readback stage stalled. Its successful cleanup
capture does not establish a repair or reopen the attempt.
The released three-file WIP was claimed before adoption and preserved as binary Git before-patches. The candidate makes transport UNKNOWN, socket closure and response-session mismatch permanent. Late replies cannot reopen it. It adds window/output identities, keeps conservative 5-second precision and 30-second menu freshness, and preserves finite 1,500 ms pulses with independent 1,750 ms release fallback. A contemplated 600/620-second original also has a mechanical 180-second gameplay cutoff and irreversible Quit phase. These are candidate controls, not approval to launch an original.
Capture uses the production host's public owner-bound
requestImmutableCapture. It reads the complete compositor output at the
next ordinary publication. A passive public publication observer binds the
output canvas, publication and display timestamp. Freshness remains dated
from request start; later identity checks never make pixels younger.
Cancellation removes the ticket and restores the observer. There is no forced
repaint, publication-hold release, extra guest draw/swap, raw slot read,
timeout widening or cached-frame fallback.
Focused controls use actual loopback endpoints and release children for scheduled release, EOF fallback, late EOF, wrong PID start, changed owner, UNKNOWN UP, late capture reply, response-session replacement and socket closure. Chain controls cover replacement refusal, unchanged request age, 5/30-second expiry with zero dispatch, and terminal UNKNOWN/no replay. Loading the exact predecessor controller in memory gives a causal transport red: it sends another fixture request after capture UNKNOWN. The candidate green refuses it, including after the late reply. This never attaches to the old original endpoint and does not reproduce its Chrome screenshot stall.
Real Chrome results and missing proof
Controls waited for paused LTris and then Reversi actual native, watchdog,
outer-child and outer waits with both EOFs and heavy release. Real Chrome
controls ran locally under the heavy lock with independent deadline
watchdogs. Chrome was 151.0.7922.108 / V8 15.1.206.16. A synthetic WebGL 2
Worker drew an 800×600 gradient and transferred image bitmaps to the
production Win98Renderer compositor. This is a real browser capture
workload, not original guest, production GL-command bridge or native CLI
proof. Actual GPU device identity was not recorded by this fixture.
The final control has a fully visible 480×360 owner window within the 684×559 complete desktop output. The initial PNG is pixel-equal to the currently displayed output. A separately requested WebGL color change produces different pixels and a newer ordinary publication with the same owner/output. Both final PNGs were personally viewed: the complete gradient rectangle and surrounding desktop are visible.
| Final control | Actual observation |
|---|---|
| Initial capture | 373 ms total; 2 ms arm; publication 4 → 5 |
| Read/encode response | 365 ms from start; PNG encoding 90.995 ms |
| Changed scene | 320 ms to response; publication 11 → 12; encoding 120.5 ms |
Default Page.captureScreenshot |
135 ms, succeeded |
| Idle unchanged scene, no hold | 3,011 ms timeout; publication 15 → 15 |
| Held compositor cancellation | 3,009 ms timeout; no pending ticket/readback |
| Window replacement | Refused; ticket removed; no input |
Default screenshots also succeeded in 62 ms and 132 ms in earlier controls. None reproduces the old 3,000 ms Chrome stall. The loopback delayed reply is transport evidence, not a replacement for real-browser causal proof.
The idle negative is decisive: a stable visible scene can remain current without a new ordinary publication. The candidate ticket cannot complete until one arrives. This control used no artificial hold, and no rendering was forced to pass it. Paused-menu/arena capture therefore remains unqualified. The conditional original was withheld and the alternative is not published as a capture repair.
Earlier fixture failures remain preserved: desktop rendering is created only on launch, and the first synthetic Worker source had a newline escaping error. A later intermediate fixture passed capture checks but clipped its owner window; its two personally viewed PNGs are explicitly rejected for complete-window proof. The final fixture adds a mechanical containment assertion. Exact duplicate intermediate PNGs were hash-checked against the retained final files before removal.
Validation, provenance and shutdown
The focused suite and full build passed under the heavy lock; the build also
held /tmp/wine-build.lock. Final verification after the fixture corrections
also passed: closure-verified-wait.json and full-build-verified-wait.json
record actual zero waits and both EOFs. An initial final-check tool session ended 143 with
partial test artifacts and no durable child wait/EOF, so no PASS is inferred.
The next validation attempt refused the existing output directory. Those
failures are retained; the final suite uses a fresh directory and durable
supervision. Passing controls include expected idle refusal and do not mean
the requested original proof passed.
The two direct validation tool wrappers report their actual exit status but do not expose outer pipe EOF fields; those remain unknown separately from the durable supervised child waits. The final fixture/candidate pins are exact; earlier fixture revisions lack independent source pins at launch and remain diagnostic evidence. Selected PNG decoding/CRC/inflate/hash checks passed, and the evidence checker reports zero own errors, missing or invalid references. Existing foreign evidence errors were left untouched.
pins.json records main 9e18bcb17731c396a20e203fa00fd9970d874e68, 675
file pins, all 92 unchanged local media files, and Git binary patches for
relevant dirty/new files. The standard build remains
c94ccf1b179f18bd4b47867445afed8cef2f1393b19aabfd259c7f3b94a7be09;
no executing original module is claimed. Git is the source archive. No source
tree, WASM or media bundle was copied into this run.
All five fixture Chrome closes were zero with actual waits and both EOFs. Native-control, watchdog, supervisor and outer waits are retained separately; the first two native control exits are 1, the later three are 0. Watchdogs did not fire and heavy locks were released. No original lifecycle wait is fabricated. The final decision/release records permit the next queued worker to advance without a nonexistent original directory.
The predecessor's 289 evidence files are inventoried and hash-checked
unchanged. Its report-only commit
232a6ae30275c70d06ff4b7c3662843b0703dfac, screenshot observation UNKNOWN,
diagnostic intact paused yellow cycle and host Chrome close remain separate
from native Quit. Native CLI evidence at 85d1ddc922 remains a different
route. PLAYER 3 WINS is not relabelled as a human victory.
No box was created: ID null, state NOT_CREATED, expiry null, stop receipt
N/A. Free /home remains above 12 GB and the 1.5 GB floor. Shared helpers
and foreign WIP are untouched. No full-27, parity, FPS, audio or physical-GPU
claim is made. The next prerequisite is a causal real-Chrome reproduction
and a complete-scene route that handles idle visible scenes while preserving
owner/frame identity, cancellation and pixel freshness. This bounded task
stops without an original or another speculative capture workaround.