GLTron continuous capture and independent release, 2026-11-17
Evidence: scratch/runs/20261117-gltron-continuous-capture-release-arm-gameplay/.
The predecessor's software read-ticket timeout and WebGL pre-DOWN arm refusal remain closed. No queries are sent to either endpoint.
After the first current-display request, the renderer retains each subsequent ordinary publication's source and output canvas caches together. Retention occurs synchronously after successful display, with no executing writer or GDI paint. A later request can read that exact publication while another native writer runs. The source is never reconstructed from the newer live producer. A new host request also refuses an already recorded native GPS producer error before it can wait behind a running slice. Identity, geometry, visibility, opacity, context loss, producer retirement and readback errors still refuse observation. First observations without a retained cache keep the existing ordinary-publication fences. No observer schedules a render, surface upload, guest entry or publication. Retained canvases are reused and only enabled after capture is requested.
The independent physical-release process connects before native launch. Each reviewed edge then arms over local IPC, with an immutable nonce and its own deadlines. Startup and WebSocket/target/Runtime phases are recorded. Both processes tombstone nonces; a real-browser test found and fixed the parent's old-reply match on a reused nonce. Cancellation before DOWN sends no input, idle parent EOF closes cleanly, and a stale browser start token refuses startup. Real native recipes, rather than protocol test stand-ins, qualify trusted physical UP and native USER return.
The arm bound remains 1000 ms. Fallback UP is sent at 1500 ms after arming, leaving delivery margin within the unchanged 1750 ms DOWN-to-UP maximum. Admission retains 500 ms before that send, and native context freshness remains 250 ms. Successful edges explicitly check the actual renderer UP timestamp against the actual applied DOWN. The software parent-stall recipe also retained the earlier boundary result: a real 1900 ms parent stall with trusted UP at 1748 ms after DOWN. The earlier send provides more margin.
Complete capture selects the immutable ticket before starting two independent reads: byte transport and a later full host screenshot. Both must join within the same original 3000 ms total deadline. A late branch cannot retain pixels or cause another endpoint call after UNKNOWN. Source/output still belong to one exact publication; the host screenshot records its separate later time.
Development failures are retained without success credit. A no-idle native producer first prevented any initialized display, then exhausted CDP responsiveness after Enter; it is not a qualification recipe. The faithful timer-driven native producer continuously presents while the game logic is idle, and records actual active writer captures. A development WebGL capture completed its ticket but exhausted the total bound during the serial host screenshot; concurrent byte transport repairs that serial overhead. A native retirement harness also attempted to serialize complete typed arrays through CDP; lifecycle results now report metadata and lengths instead of millions of numeric properties. These are distinct from native gameplay proof.
Focused tests cover active and parked writers, exact literal retained pixels, later publication isolation, queued damage, paint, cancellation, stale owner, process/instance replacement, producer errors and epochs, context loss, retirement, opaque output, concurrent transport and late-result discard. Real native/browser recipes and final original outcomes are recorded in the run's result JSON. Original allowances are used only after the fixes, complete recipes, focused tests, locked build and strict publication.
Foreign guest-thread, render-worker, NE and child runtime work is excluded from publication and immutable runtime inputs. Only the renderer review row is owned in the shared review file. Evidence records Git and binary patches; media are referenced by existing path and SHA-256. No source archive, frame archive, media copy, worker, box or public deployment is created.
The repairs were strictly published as b29ef843aee74fe366aac3698704e3238797ed9e.
All 35 focused tests and the locked build passed. Complete native/browser
software and WebGL recipes reached native Quit(0), including actual trusted
physical UP, a 1900 ms parent stall, cancellation before DOWN, and continuous
ordinary presentation. The final software recipe exercised the exact final
capture identity checks and the earlier fallback send. Native paused-writer
retirement and timer-driven WebGL context loss passed literal source/output,
retained bytes, invalidation and actual worker shutdown checks.
Each backend then used its one authorized fresh original from that published Git graph, with every helper exactly published, unchanged media, module bytes retained before execution, and prearmed independent owner/runtime watchdogs. Software reached the real paused arena. Its captures completed in 255–1263 ms, including the paused arena with an executing native writer and exact retained publication pixels. Release arms took 6–7 ms. Space then applied 10 ms after a fresh native owner observation, but dequeue saw that observation at age 293 ms and cancelled against the unchanged 250 ms context fence. Physical UP arrived. WebGL reached the real title menu. Its first Enter armed in 4 ms and applied 86 ms after the fresh observation, but dequeue saw age 259 ms and cancelled against the same fence. Its independent physical UP also arrived.
Both endpoints were permanently closed with zero subsequent controller requests. There was no retry, timeout increase, forced frame or fence change. Both original invocations ended through scoped browser cleanup with actual process waits and EOF; neither is native Quit. Neither backend demonstrated two live turns or a round consequence. The concrete remaining blocker is native USER context aging between fresh observation, physical DOWN and native dequeue under the original workload. Its scheduling cause is not established; the authored routes do not reproduce that original delay. Further work needs a faithful native latency reproducer under the unchanged 250/500 ms fences before any separately authorized new original attempt.