GLTron current native capture qualification, 2026-11-17

Evidence: scratch/runs/20261117-gltron-current-native-capture-round/.

No original has been launched. The continuous authored Win32/OpenGL fixture has produced a bounded public current-display observation and a complete host screenshot, but the retained source was entirely white. The later host image shows the triangle. These observations have different timestamps; they do not establish an incorrect readback of the same display publication. No production capture/input repair, initialized-scene qualification, trusted guest input, two turns, natural round outcome or native Quit is claimed.

tools/gltron-current-native-capture.cjs preserves the earlier authored PE byte for byte (SHA-256 a282fde427480359786e060285a52a0cb94517e2df9ab7a40a9b91bb90e43bbf). It continuously pumps native messages and issues GL clear/triangle/swap calls, with native temporary display mode and unchanged 1,000-block slices. It uses the current Git runtime and canonical WATX compiler in memory, with no foreign worktree source, source archive, original media or shared build outputs. The actual guest/render Wasm bytes are recovered with Debugger and checked against the compiled module. Unique tested modules are retained once.

The browser uses the existing local Puppeteer/Chrome route, explicit ANGLE SwiftShader launch options and real clocks. WebGL reports an Intel/Mesa string; that string is not independent physical-device attribution. No physical GPU, FPS, audio, parity or full-27 inference is made.

Bounded observation

The diagnostic native-bounded used Git 799439f0a8e508f22d760666438acba811147315, with its recorded owned helper patch. Runtime source/module and helper patch are separately identified; this development run is not a published-helper qualification for an original.

The single current-display request armed during a real writer hold and pending host presentation. It completed on ordinary publication 5 in 1,539 ms:

The separate unchanged full-host request Page.captureScreenshot({format:'png',captureBeyondViewport:false}) completed in 2,299 ms. Personal review shows the ordinary fullscreen consent strip, native yellow triangle/dark blue background, shell and runtime log. Canvas-only input admission remains invalid: the real touch-cursor DOM overlay intersects the output and is neither hidden nor exempted.

Current native GL uses the GPU stream route. The legacy GLTron binding's options.api === 'gl' lookup reports no GL endpoint on the observed current route, despite an actual render endpoint and device draws/transfers. This source/observation mismatch is a further unqualified owner-binding boundary; no legacy-GL identity is substituted for current GPS ownership.

Earlier preparations and limits

Two browser preparations stopped before DevTools opened. The new harness's initial direct-spawn startup check did not account for this installed browser's own timezone probe. The existing Puppeteer launch preparation reaches the native scene; capture/input deadlines were not increased.

native-puppeteer first awaited the ticket in a single CDP evaluation and became permanently UNKNOWN. The diagnostic native-final-poll armed once and observed pending publication 7 while slices advanced 22 to 32. A final individual poll could outlive the total observation budget, so its 5,426 ms failure is diagnostic only, not a 3,000 ms qualification. The final helper adds an independent overall 3,000 ms close timer; an earlier preparation was cancelled before browser acquisition to preserve permanent-close teardown. No endpoint was rebound, queried after UNKNOWN, or reused for an original.

The final helper observes existing current-display tickets only. It never forces idle, flushes, calls manual render, schedules a repaint, releases a writer or changes a clock to obtain pixels. A completed ticket's historical bytes are independent of current input authority. A full native input/focus/ owner/generation/readiness positive and cancellation, late paint, context loss and retirement matrix have not passed here. The unchanged 500 ms confirmation, independently prearmed 1,750 ms physical UP, 3,000 ms completion, and 5-second game/30-second menu freshness remain prerequisites. The old controller's 2,000 ms UP assertions are not accepted as qualification.

The conditional original is withheld, so its fixed 1,200-second native guard, 1,220-second independent watchdog and 180-second Quit reserve are unconsumed. All original media/defaults are untouched and no original endpoint was opened.

Checks and shutdown

The final native-published check serves only published Git fa5ed3fc81352fb039ef370889c089019f6189af, including all five exact helper pins and empty relevant patch. Both actual Worker modules match cff5da9a64e3df960918df28ae7e06bd1db7cfae49d189f4281d82ea09e7f274. The current GPS owner is explicitly observed: gl-gps, endpoint 1, epoch 1, 101 native presentations, live 480 by 360 layer, actual focus/visible document. The legacy binding still reports render endpoint none.

Its public ticket completes in 915 ms on ordinary publication 5, again retaining exactly 172,800 opaque white source pixels. After publication 7, both retained arrays and original timestamps are unchanged. The subsequent full-host screenshot becomes permanent UNKNOWN at 3,001 ms. There are zero later CDP requests, zero input, no adopted late PNG and no rebind/replay. This is the terminal current-source prerequisite refusal; the earlier 2,299 ms host success is not a causal repair and is not substituted for this failed capture.

Native process actual exit 1 is preserved. Both independently prearmed watchdogs have actual exit 0 and observed pipe EOF. Browser actual exit is 0 with both pipe EOFs. The separate outer supervisor actual exit 1 and both EOFs are recorded by completed Popen.communicate; its saved regular files are identified as logs, not pipe-EOF observations. Scoped processes and profile are absent. No actual Worker target-destruction receipt was obtained after terminal transport close; the empty event list is not promoted to one.

The canonical flock build passed; it is a separate shared-tree gate, not the immutable native fixture module. Its post-build shared diff and module hashes are retained. Fixture byte equality, JS syntax and the focused retained-image/ timing/actual-wait verifier passed. The verifier explicitly returns qualified:false and does not turn a partial observation into gameplay proof.

The reused published Boids supervisor prearms independent preparation/runtime watchdogs, uses pinned process identity, and retains actual subprocess waits and observed pipe EOF separately from regular-file logs. Development nonzero exits/signals remain recorded. Successful bounded observation has native, both watchdogs and browser actual wait 0 and both pipe EOFs. Profiles and scoped processes are removed. Browser teardown is not guest native Quit. Worker target destruction is not inferred from an empty CDP closure-event list.

The next real boundary is an initialized native GL publication whose complete retained pixels match an independent literal oracle, with actual current GPS owner identity, followed by the full native input/fault qualifications. An intermittently successful host screenshot alone does not satisfy that boundary.

No boxes, other workers or public deployment. Box ID null; state NOT_CREATED; expiry null; stop receipt N/A. Git and recorded binary patches are the source archive; only named selected images and unique tested modules are retained.