GLTron immutable attestation and bounded pre-ticket retries

Evidence: scratch/runs/20261120-gltron-pre-ticket-cost-repair/result.json.

The source regression fails on the old binding at the unchanged five-second PREBIND deadline. The candidate prepares immutable attestation before PREBIND, retrieves module bytes once per actual Worker/script/instance incarnation, and reuses that proof only within the bounded pre-ticket intent. Every arm still reads the live page owner, output, context references, publication and paint readiness. Separate intents and post-capture validation retain actual Worker instance and renderer endpoint reads. Context lifecycle events, new wasm scripts, retirement, replacement and hidden output invalidate admission. No scene pixels, publication timestamps or readiness results are memoized.

The deterministic cost regression uses the measured 594 ms identity and 158 ms bytecode request costs. Old source expires before its third arm. New source prepares in 1346 modeled ms, reaches all three arms, and sends zero Worker identity or bytecode requests inside PREBIND. Eighteen focused invalidation cases pass, including renderer replacement and invalidation between positive NOT_READY arms. This is a source regression, not a browser latency benchmark. The existing native-input, capture and real-child controller regressions pass; the full flock /tmp/wine-build.lock bash tools/build.sh exits zero.

The fresh original uses the exact f696 runtime, a600 module and 92 existing media pins with the existing three passive observer overlays. Shared RPC and native-input source are read-only. Four fresh main-menu/current-display captures pass using the repaired binding. The selected complete browser images were returned for personal review. Host fullscreen consent succeeds. Two expired local release windows refuse before guest input; a subsequent timely Enter opens the visibly reviewed GAME submenu. A further Enter selects START GAME with independently joined UP and physical neutral. The existing release ordering and deadlines were retained; no release-window extension was made.

The next, distinct round-pause intent exhausts the original five-second window. It observes publications 373 through 385, reaches readiness once, then receives 66 positive pre-ticket NOT_READY arms for held/in-flight paint. Across the observed window, all 172 subsequent requests are Runtime.evaluate on the page session: zero Worker requests and zero module-bytecode requests. No capture ticket or Space DOWN is issued. Thus the repeated identity/bytecode cost is removed, but ordinary paint readiness remains a newly evidenced dependency. The deeper reason for that held-paint interval is unknown. This scope closes permanently; there are no later endpoint requests or unchanged static campaign. Two turns, pause/resume, visible round outcome and ordinary Quit are unproved.

Supervisor and independent watchdog join with both EOFs; the supervisor reports leaked:false. Browser, Xvfb, Openbox and release children have actual waits. The outer tool returns 143 separately from the saved supervisor exit 1 and watchdog exit 0; its signal cause is unknown. Runtime disk monitor minimum is 1,793,978,368 bytes. Selected screenshots, compact request/reply joins, unique receipts, owned patches and recovery references are retained; no bytecode payloads, media copies or source-tree snapshot are retained.

No box was created: ID, expiry and stop receipt are null. No agents, workers, Python, Wine, public deployment or Windows reference run were used.