Stronghold scene20: fresh input-free samples reject stability
The dim actual layout is not stable across the next 256 input-free steps. The sole Chrome reproduced it at11136 steps, then showed brighter setup panels, an upper-right map and no unavailable text at11392. The complete-output stability guard rejected that second image and immediately closed Chrome. No preset press, Start or world input followed. This establishes an original rendered transition, not a rendering defect, a fade mechanism or strict scenario readiness.
Evidence: scratch/runs/20261104-stronghold-fresh-layout-stability/. Predecessors: actual-layout audit, commit 94d89d795cd095014aeb6a48dff993f24b9c6464, and natural scene20 rendering, commit 38238efd5b786b6a00b594ec5a6f775a3ce90787.
Personally viewed fresh captures
I personally viewed both recovered original captures in a separate tool call before writing their hash/step/observation reviews in personal-stability-review.json:
- At11136,
route-native-boundary-6.png, SHAb5adb615b443746c40f3a92625e563701a168cb81856ebcf22511fccc170096c: buttons1/2/3, “Not available in the demo”, dim Human/Lord Crusader, Deathmatch, gold and player panels; no upper-right map. - At11392,
route-native-boundary-7.png, SHA085a70c1d54fb61cda112692db2c7ee17e6b2bd2003626759f547abce2e10bbc: brighter, sharply bordered panels; readable Human/Lord Crusader and Deathmatch, gold10000/10000, Crusader Demo player and an upper-right map. The unavailable wording has gone.
Four complete64-step ACKs separate these images. Input telemetry is drained at both; there are zero transport events in the interval. The numbered-button rectangle remains byte-identical while the complete game rectangle changes 1,129,717 RGBA bytes. This independently demonstrates why button-only acceptance is insufficient.
Observation policy and first failure
The run-owned candidate identity hashes every RGBA pixel of the measured game-output rectangle [10,194,860,668], including all panels, controls, text and background. Its retained-reference hash is ef9494702084b55824328cd484a347f0eac79f7f32e106e821c36eb84153f768; the fresh11136 image matches exactly. It is explicitly an observation identity, not input authorization from one image. The next sample hashes to f156657d1308d2e3af2b59793a1184cb4c12619cfc1aebec2acd195914135e0c.
After the original balanced Custom Game gesture, only the already recognized hover is allowed while awaiting the candidate, within the inherited4096-step route budget. Either an immediate or delayed candidate starts a planned three additional256-step input-free sampling groups. Existing identity, presentation, ACK, input, observer and fault guards apply throughout. The first such group rejected its changed output, so later groups and final personal-review authorization were never reached. No recognizer was changed after that failure.
First failure: 10:32:53.909Z, first unexpected actual stability scene: unknown. Chrome12100 closed at 10:32:54.409Z, exit0. No second browser, reload, forced guest call, guest write, repaint or additional trace range was used. The later map image is retained as an unqualified new state; no Zephyr/Start/world claim is made.
Tests and preflight
Before Chrome, the exact unchanged 27 observer,20 freshness and12 route controls passed, alongside original instruction/native-route tests and actual retained WASM/original-EXE non-ID2-negative and ID2/request20-positive execution in the actual Worker/broker fixture. The fixture's intentional instruction execution belongs only to isolated preflight, not the original browser route.
Actual CJS/dynamic ESM/Puppeteer imports, syntax, complete helper hashes, 2415 retained runtime pins, the98-page/12-Worker cache graph and five pinned HTTP/isolation responses passed before launch. The assigned expiry was verified through the API. The600000ms TTL guard and inherited300000ms expiry reserve were retained, exceeding the required180000ms recovery reserve; Chrome was bounded to six minutes. Allocation/deadlines were appended before launch. No runtime touched the stopped predecessor box.
Candidate controls reject five retained wrong scenes and nine panel/text/button/background/edge mutations. Four modeled sampling cases passed before launch: delayed arrival, stuck hover, unexpected waiting image and changed actual sample. An initial test caught an immediate-arrival sampling bypass; it was fixed and retested before launch. Post-recovery independent fresh11136 positive and the same nine mutations pass; the fresh11392 transition rejects. These are candidate-identity controls, not a stability positive. Offline fixture-path and transport-count audit corrections are retained explicitly.
actual-evidence-check.json verifies 178 complete ACKs /11392 steps,1052 served-pin matches, six paused-before-start Worker preloads, four live armed instances,26 observer reports, two explicitly pre-recording retirements, zero observer errors/drops and eight rejecting proof mutations. Original callbackID2/branch/request20/down/release ordering is proved; final current/requested scene remains20/20. Callback absence remains unqualified.
Recovery and custody
All 262 immutable closed files /19,489,587 bytes were recovered and SHA-verified by10:33:10Z, with the unique actual-Worker result, exact executed driver and driver log recovered separately. Post-close checks find Chrome/driver absent, actual HTTP44145 closed and all2415 pins unchanged. The last API observation at10:34Z shows bx_jret53j4 RUNNING /API-idle, healthy and hydrated, expiry 2026-11-04T11:05:35.153Z. Coordinator owns shutdown and receives custody; no lifecycle action was taken.
Runtime source remains 1188e9bf57ef35bd3363fb88b7ca0da442ce696d, normal WASM cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee. Exact checkout Git SHA, relevant binary diff, owned-new-file patches, small-input hashes and artifact seals preserve provenance. No source archives, module/media copies, production code, authentication/configuration changes, local heavy runtime, subworkers, successors or deployment were introduced. Only this note is committed; no production full-build claim applies.
The remaining gap is successive input-free stability and complete independent recognition of the later11392 state before any preset gesture can be authorized. This single assignment stops at its required first failure.