LMarbles finite-intent freshness admission, 2026-11-16

The controller now rejects insufficient remaining operation time before DOWN, normally before prearming its independent release child. The one authorized fresh Normal / Original / Tier 1 Puzzle 8 session reached Moves 96, but every subsequent board intention was NOT_SENT. The manual full-image review turnaround left too little of the five-second capture window. Escape was also NOT_SENT. There was no clear, tally, Puzzle 9, or ordinary native Quit.

Evidence: scratch/runs/20261116-lmarbles-freshness-release-admission-repair/result.json.

Admission and immutable bounds

The predecessor's sg1 had only 121.8 ms remaining before DOWN. Its 100 ms pulse left 21.8 ms for UP acknowledgement; UP took 190.4 ms overall and arrived 68.6 ms beyond the immutable deadline. That original remains permanently closed. No old endpoint was contacted or restarted.

Strict admission now requires strictly more than max(1.750, 0.500 + pulseSeconds + 0.500, 1.650 + 0.500) seconds. These are the existing physical-UP bound, confirmation/pulse/confirmation path, and independent fallback plus confirmation allowance. For all supported 40–1000 ms pulses the result is 2.15 seconds. This is not a measured-latency timeout adjustment.

The deadline is computed once as the minimum of the existing three-second operation window, capture freshness expiry, and personal review freshness expiry. It is never refreshed. Admission is checked before prearm, after arm, and again immediately before dispatch. The latter checks also require enough room before the already fixed fallback for confirmation and the requested pulse.

An armed release child can be disarmed before dispatch. It independently checks the no-dispatch/disarm records and exits without HTTP or physical release when no DOWN was authorized. Dispatch is durably marked before sending DOWN; once effects are possible the independent UP remains mandatory. Slow ACK, invalid focus/owner, UNKNOWN, release failures and post-effect timing failures still permanently close the controller. They cannot become NOT_SENT or permit replay.

Confirmation 500 ms, independent physical UP 1750 ms, observation 3000 ms, game freshness 5 s, menu freshness 30 s, and the 300 s Quit reserve remain unchanged. The reserve is irreversible, including when an explicit Quit intent is rejected as NOT_SENT.

Verification

Seven completion-budget tests pass, including equality at the admission boundary with zero request or child creation, both sides of that boundary, scene freshness, and actual 3600/3620/300 prelaunch watchdog/reserve records and complete waits. Eleven existing reserve/lifecycle tests pass.

Eight focused native tests exercise actual threaded test/run.js, real USER windows and guard consumption, the production HTTP transport, and an independently executed UP child. They cover sufficient room, actual elapsed capture age, budget spent during prearm and immediately before dispatch, a fresh capture and new nonce after proven NOT_SENT, slow ACK, lost ACK/UNKNOWN, focus and owner. Positive consumption, release and confirmation pass within the unchanged bounds. Predispatch refusals issue no DOWN; armed refusals issue neither DOWN nor UP HTTP. Every fault control attempts capture, rebind and another pulse after closure and proves zero later requests.

The final focus control changes actual native focus through the exported USER setter. The final owner control replaces the fixture's bound renderer window reference. Production guard receipts are CANCELLED/unconsumed with respectively reviewed native scene changed and owner/epoch/queue/device/window changed. An earlier diagnostic control attempted a forbidden second bind; its expectation of a failed UP child was wrong because the permanent-bind fence rejected it before changing ownership. That assertion failure is retained. The corrected actual-window replacement control passes. These are fixtures, not game inputs.

The required flock /tmp/wine-build.lock bash tools/build.sh passed with actual native/watchdog/supervisor waits and both EOFs. Shared outputs were released promptly; subsequent runs did not read them. Exact 86-file media/CRT checks and seven manifest corruption controls pass; postflight verification also passes.

Startup adoption and fresh original

The runtime uses published Git 08d860b1809f678b912f21f2c85689ee9f21c3dc, a canonical independently compiled sealed module 83703369cc49c3e16077a4abd43def27262487f83dd52583d1b433adf7095f5f, and the predecessor's unchanged indexed-source loader/supervisor. Its exact missing and tampered source/index/module validation is retained unchanged. The predecessor's observed cold startup result and native two-guest-Worker fixture are recorded in 20261116-lmarbles-threaded-desktop-startup-repair; no unchanged startup or GL suite was repeated here. The new original itself instantiated both actual guest Workers and displayed its ordinary menu and Puzzle 8 scene. Authentic SDL and the exact original media inventory are unchanged.

The new original recorded 3600 native seconds, 3620 independent watchdog seconds and the 300-second Quit reserve before launch. Four personally reviewed setup inputs completed; physical UP acknowledgements took 123–184 ms. The complete Normal/Original setup, level-8 confirmation and Moves96 board were reviewed. There was no old progress, save, solver route, private board read, OCR, autoplay, or queued action. Full-frame JPEG display conversions were tried to reduce review turnaround; intent hashes and timestamps still refer to the original PNG capture. No crop or timestamp refresh was used.

Five board/Escape intentions were cleanly NOT_SENT: two captures had expired, and three retained only 1.079–1.625 seconds at admission. None reached prearm or DOWN. Each new intention followed a new capture and nonce after its predecessor was provably unsent. The final Escape intent irreversibly entered Quit phase but also lacked admission room. The last complete image still shows Moves96. This is an observed manual-review freshness blocker, not UNKNOWN and not a physical release failure. No second original was launched.

Owned local cancellation shut down the host; native exit0 is not game Quit. Native/watchdog/supervisor actual waits and both EOFs are retained. All 204 recorded process IDs and the original port are absent; 167 paired process waits are indexed. The source/module holder's invoking tool exited143 without a normal close receipt; its PID/descriptors are absent, and native launches had already verified and resealed their own inputs. The board watcher also returned143; the tool API does not separately report its pipe EOFs. These limitations are retained without fabricating shutdown receipts.

No box was created: ID null, state NOT_CREATED, expiry null, stop receipt N/A. Git plus owned binary/new-file patches are the source archive; no source tree, media copies or WASM dump was archived. Only selected named images are retained. Any further original attempt requires separate authorization and must preserve these bounds. This worker releases its files and stops after publication.