GLTron startup readiness, 2026-11-14

The browser helper now waits passively for startup display readiness before its first capture ticket. The tested change is published as f6931d34b7ada7386d335624fe77f41803a2ea4f. The sole fresh original reached the rendered GLTron title/menu, then refused prebind output visibility. Its diagnostic shows the HTML fullscreen-consent strip overlapping the game. No ticket, personal binding or input was admitted. Two intact-cycle corners, a natural round outcome and native Quit remain unproved. There was no retry or second original.

Evidence: scratch/runs/20261114-gltron-startup-readiness/result.json.

Scope and admission

This is a startup-readiness change in tools/gltron-browser-binding.cjs and tools/gltron-browser-controller.cjs, with focused controls in test/gltron-browser-controller.test.cjs. It does not change host.js, lib/renderer.js, the current-display API, or callback census rows. The distinct idle-capture fix in 9baf6cec8122baa5698e1934240db8abe1617f84 and its final report bbceb4df42e84fef18dca1e80125c93417327251 retain their own scope.

Startup observes the first identifiable owner and keeps its process, worker, module, renderer, window and instance references fixed. A context/first clear/ present is insufficient: actual frontend draw/present and backend draw/transfer observations must exist. These observations are necessary conditions, not a gameplay or nonblack-scene certificate. Personal review still precedes permanent input binding.

The passive gate requires an actual ordinary complete-display receipt, the same owner, visible contained output, no pending host publication, no held/in-flight paint and no pending ordinary damage. It calls the production receipt validator with an ephemeral local owner request. The validator only marks that local request selected; no renderer state, ticket, pixel bytes, guest operation or repaint is produced. Poll pacing is not evidence of readiness. The existing 90-second startup bound and mechanical 180-second Quit reserve remain fixed.

Before an ordinary display exists, output choice is unknown: the normal compositor can select direct or filtered presentation. Output identity is selected once from the first valid unheld receipt, then checked through full worker/module validation and pinned permanently. A later output or owner replacement refuses. The initial output selection is not a rebind of a captured or closed session.

Capture rechecks owner/output/paint state in the same JavaScript turn as its ticket request. A known busy refusal has code CAPTURE_PRECONDITION_REFUSED and input state NOT_SENT; it is not UNKNOWN input. Every failed capture closes the chain. Cancellation and transport UNKNOWN remain terminal, including after late replies. Ordinary damage that arrives after startup can still use the existing current-display ticket's ordinary-publication path. The 3,000 ms capture limit, 5/30-second gesture freshness, one-use nonces, finite 1,500 ms pulses and independent 1,750 ms UP remain unchanged. No forced repaint/swap, hold release, raw slot read, guest clock/configuration change, solver or action loop was added.

HTML overlay boundary

Canvas bytes do not depict HTML overlays. Admission now scans intersecting painted light-DOM element boxes, including elements with pointer-events:none. It conservatively refuses them even when their stacking order might put them behind the opaque canvas. Empty transparent structural containers do not count as painted merely because Chrome reports an outline width with style none.

The Chrome control adds a real red HTML sibling over the output canvas and demonstrates admission refusal. This is separate from native-window overlap tests. This guard is a conservative light-DOM bounding-box check, not a general CSS/shadow-DOM raster-occlusion certificate. Canvas-only evidence must retain that qualification; the original's browser diagnostic is explicitly different from an admitted canvas scene.

Validation

Final production Chrome 151/WebGL2 control, synthetic producer, ordinary production compositor (real-chrome-v8/report.json):

Check Result
Exact predecessor helper during startup hold Premature ticket at producer frame 0; held-paint refusal
New passive wait through ordinary producer completion No new ticket/input; producer ended at frame 20; first new ticket unheld
Complete startup display capture 262 ms; full gradient window personally reviewed
Idle current display 46 ms; publication 19 unchanged; exact output and Chrome compositor pixels
Old next-publication idle ticket Still pending after its fixed 3-second budget; cancelled
Default CDP screenshot 41 ms; historical stall not reproduced
Late hold Known NOT_SENT refusal before ticket creation; hold unchanged
Actual startup timeout 128 ms with a 125 ms control bound; no ticket
Owner/output changes, hidden/clipped output, cancellation Refused; late publication cannot revive cancelled capture
HTML sibling overlay, pointer events disabled Admission refused
Actual CDP deadline and late reply 3-second UNKNOWN; still closed at 3,251 ms; later input calls refused

Four selected synthetic PNGs were personally reviewed. They show complete gradients and the exclusive presentation's actual black bands; they are not game proof. No fixture input occurred. Seven preceding synthetic development controls are retained, including the initial conservative DOM false positive, normal output selection transition, a report-serialization bug and an overstrict ordinary-damage capture guard. These were closed controls, never originals or reopened sessions. Intermediate diagnostic controls v2/v4 lack independent full source pins; final v8 has exact Git SHA, source hashes and relevant binary patches.

Final focused controller/loopback/actual-child tests pass, covering passive readiness, actual timeout, owner replacement, output selection, terminal late holds, replacement/UNKNOWN/cancellation/no replay, 5/30-second ages and the irreversible Quit cutoff. Five focused renderer files pass (20 entries, including 16 endpoint cases). Final-source flock /tmp/wine-build.lock bash tools/build.sh passed in 28.37 seconds. Real Chrome blank-page trusted DOWN/UP controls passed for every allowed key, a 1,500 ms pulse and independent 1,750 ms EOF fallback. Final code was committed and strict pull/push completed before the original.

Sole original and custody

The original launched at 21:33:07 UTC with its independent 600+20-second watchdog already armed, after Minesweeper's native/watchdog/supervisor/outer actual waits, both EOFs and heavy release. Its exact 677 input pins include all 92 existing media files, without copies/downloads, plus Git SHA and relevant dirty patches. Actual recovered guest and renderer module SHA-256: c94ccf1b179f18bd4b47867445afed8cef2f1393b19aabfd259c7f3b94a7be09. The actual host used 1,000 blocks per slice and normal real clock.

The early owner observation already reported output visibility false. Startup waited until actual frontend draws rose to 2 (backend draws 3, present/transfer 1), then the prebind output check refused before the first ticket. The personally reviewed full-browser diagnostic shows the rendered GLTron title/menu and an HTML fullscreen-consent strip across its top. This corroborates an unsupported canvas-only visibility condition; no exact live DOM refusal-reason inventory was retained, so the strip is not claimed as the sole evaluated failing clause. No admitted gameplay scene, known human identity, corner, natural outcome or native Quit is established. Browser close is host teardown.

Native controller exit 1, Chrome exit 0, watchdog exit 0 (unfired), outer child exit 0 and outer caller exit 0 all have actual waits and both EOFs. Supervisor zero means successful custody, not game success. The heavy lock is released, all four recorded runtime/browser/supervisor/outer PIDs are absent, HTTP port is closed and profile removed. All 677 source/media pins remain unchanged. The closed direct predecessor's 356 files and current-display predecessor's 260 files retain their exact names, sizes and SHA-256 values. Neither was reopened. The ancillary board watcher uses tool-managed pipes; separate pipe EOF evidence is unavailable and is not included in the runtime wait claims. Intermediate synthetic v2's tool-managed outer caller also has unavailable separate EOF flags; its native/watchdog/supervisor waits are recorded independently.

Only five selected PNGs remain in this run; three byte-identical unselected development copies were removed with exact recovery hashes/paths recorded. No source/media/WASM archive, frame series or box was created. Box ID/expiry are null, state NOT_CREATED, stop receipt N/A. No full-27, FPS, audio, parity or physical-GPU claim is made. This bounded task ends with a tested published startup fix and the precise missing gameplay proof above.