GLTron original USER/render phases, 2026-11-17

Evidence: scratch/runs/20261117-gltron-original-render-user-phase-cause/result.json.

One fresh original per backend was used; neither may be repeated in this assignment. No production repair, reviewed live turn, round consequence or original native Quit is established. The WebGL attempt isolates an unresolved 444 ms interval before the owning Worker's next USER import. The recorded 623.46 ms render drain starts after cancellation and cannot explain it.

Later external persistence diagnostic

Evidence: scratch/runs/20261117-gltron-page-stall-passive-tail-cause/result.json. Test-only observer commit 4b141013d adds an independent exporter Worker. Direct console events from a blocked producer failed to reach Node in the authored proof. The exporter instead reads atomically committed shared records and returns from its own timer before Chrome sends its console batch to an external append-only file. A page spinning inside capture and a producer blocked in Atomics.wait both retained their final entry stacks through actual browser exit, both pipe EOFs and external file fsync. Another blocked producer retained two records and reported exactly twenty capped drops. This is transport qualification, not reconstruction of missing original GL inputs.

The separately authorized single WebGL original reached the GAME submenu. Its one physical Enter returned through native USER in 130 ms, with independent UP at 222 ms. START GAME admission expired before DOWN: the bound owner remained healthy but its last USER observation was old. The participant confirmed cancelNoDown; no later transport query or second original followed.

The external tail preserves an 802 ms interval between the last presentation transfer return and publication entry in the repaint path. It also preserves an 855 ms gap between main Worker wait boundaries; the later stack resolves through handle_ReleaseSemaphore using the exact retained module's name section. Neither interval alone separates JavaScript/native canvas work, guest execution, GPU driver blocking or OS scheduling. No OS CPU, historical 444 ms cause, production repair, live turn, round consequence or original native Quit is claimed. The older request-ticket timeout was not reproduced by this attempt.

Review failure and qualification limits

The software original's first Enter was incorrectly written in the same tool call that requested its image, before the returned image was personally inspected. Its review text also incorrectly described a title prompt; the actual initial whole image was black during initialization. This violates the required review ordering. Input stopped after recognizing the mistake. Its 63 ms native return and subsequent visible main menu are diagnostic evidence, not reviewed gameplay qualification. There was no second software original. The same ordering error invalidates reviewed-input claims for preliminary authored-software-v2 and authored-final; explicit correction records remain.

Before the WebGL original, authored-reviewed-webgl used separate image-view and subsequent intent tool calls. Reviewed Enter returned through real owning USER in 67 ms with independent physical UP, and the next whole image showed the red background. Reviewed Escape subsequently cancelled at token zero and scene age 331 ms; input stopped. This validates the observer and cancellation path, not native Quit or equivalence to the original workload. Existing active/parked, context-loss, future-request, stale-owner, noDown and parent-stall proofs remain unchanged predecessor proofs; they are not newly requalified by the overloaded diagnostic fixture.

Original WebGL phases

The initial black image received only an observation request. The next whole image showed GAME highlighted. After separate personal review, Enter returned through owning USER in 136 ms. A later whole image showed START GAME in the GAME submenu. The separately reviewed Enter for that selection cancelled.

Phase Wall timestamp, ms Meaning
Last owning USER import entry/post 1794923174330 Slot 0, pinned guest Worker
Page observes native context 1794923174332 Authentic observation
Owning USER import returns 1794923174333 Empty native poll
Physical DOWN applied 1794923174361 Observation age 29 ms; CDP round trip 11 ms
Last early status 1794923174399 DELIVERY_PENDING, token 0
Next status cancels 1794923174772 Scene age 440 ms, token 0
Next owning USER entry/post 1794923174777 444 ms after prior API return
Page observes that poll 1794923174780 3 ms after Worker post
Next owning USER returns 1794923174783 Cancelled DOWN absent; no token allocated to it
Later synchronous GPU drain begins 1794923174952 After the cancellation
Later GPU drain returns 1794923175575 623.46 ms elapsed
Independent trusted physical UP 1794923175762 1401 ms after DOWN

The gap between two status calls also includes roughly 373 ms of elapsed time. That does not establish page CPU time or a specific stalled function. The observer does not fully bracket gpu_gl_call/GL batch waits, guest execution, OS runnable delay or browser scheduling. There is no demonstrated cause for the 444 ms interval, and the authored software render-wait example does not supply one. No speculative production edit or old/fixed claim follows.

Both original display sources retain the default 800×600 dimensions. Public GPU record counts are retained per complete frame: the last WebGL menu frame before cancellation has 30 DRAW_UP and 30 BIND records, two CLEAR, two RASTER, and one each TARGETS, MARK and PRESENT. Later frames have a different mix. Only the first 64 records retain full fields/input bytes, capped at 16 KiB; subsequent per-frame mixes and slow-call timings remain available. Raw GL batch mix and later dynamic renderer input bytes are missing. This is not a faithful replayable original workload, and elapsed renderer calls include any scheduler delay. Remaining work is attribution of the pre-USER interval and complete renderer-workload capture under a separately authorized assignment.

Source, bounds and closure

Diagnostic observer commit: 051d8f1b621a7d2b0cc4e0315f320786fadad90f. Both originals serve the unchanged immutable 60371e9f3bf83f1fbcfab790f1e34977b06e9ec7 runtime graph. Its complete host/lib/WAT graph matches the retained predecessor runtime; the sole earlier helper difference adds authored fixture options. Both real Workers receive the retained 6de8753b667362e909292536a340219ccbf90eedccb2b8405c31975193f5e78e module, pinned before execution. No foreign WIP is served. Per-run exact Git, relevant binary patches, helper pins, all 359 media SHA references and actual Worker module hashes are retained. Media and source trees were not copied.

Freshness 250 ms, delivery 500 ms, independent UP 1750 ms and complete capture 3000 ms remain unchanged. There is no forced pump, restamping, guest timer, configuration or EXE change, private game read, OCR, solver or guest mutation. The WebGL cancellation has zero subsequent main-controller requests. Later passive console events come from the prearmed observer while independent UP and process joins finish; they do not reopen the endpoint.

All six run scopes have actual native/browser/outer-owner/watchdog waits with both EOFs and empty owned scopes, including failed and review-invalid runs. Cancellation cleanup is browser termination, not original native Quit. The selected whole host PNGs remain directly viewable. Supporting PNGs and closed verbose logs are individually compressed with exact-byte SHA-verified recovery recipes; no source archive, frame bundle or duplicate media is made. Temporary profiles and Chrome temporary files were isolated on tmpfs and removed after scope closure. Shared disk-floor crossings stopped the first authored run before input. Launch checks required more than 1.5 GB free; the final shared-disk sample fell below that floor after all task runs stopped. No box was created: ID/expiry null, state NOT_CREATED, stop NOT_APPLICABLE.

The 18 focused native-input/capture/GPS/observer tests and canonical flock /tmp/wine-build.lock bash tools/build.sh passed. No new code was changed after those checks. The task stops with the explicit attribution gap, the software review failure and both original allowances exhausted.