Stronghold scene20: another intermediate output stops qualification
The sole original-game attempt reached bright setup panels and an upper-right map at11200, but “Not available in the demo” was still visible. This complete output matches neither the retained dim11136 nor later11392 positive. The first unexpected capture stopped the driver immediately. Successive later-state stability remains unqualified; no preset press, Start or world input occurred.
Evidence: scratch/runs/20261104-stronghold-later-layout-stability/. Input: fresh layout stability, commit d4e1999e5391ace5152de97f50be62719f5d909e. This is a failed observation prerequisite, with no guest rendering-fault claim.
Model and unchanged guards
Before launch, run-owned observation code admitted both exact retained complete-output identities: earlier dim output as a waiting state, later bright output as a candidate. It required three further256-step input-free groups with later identity before personal review. The original4096-step route ceiling,3072-step arrival loop,64-step complete per-guest ACK guards,16384 navigation ceiling, observer history/range, drainage, input, presentation, fault and freshness guards remained. The known earlier-to-later change was neither a failure nor input authorization.
Five modeled cases passed: immediate candidate, delayed candidate through retained earlier output, stuck hover, unexpected arrival output and changed later stability sample. Complete-output identity controls passed the retained later positive, five wrong scenes and ten map/panel/text/button/background/edge mutations. A separately implemented tiled recognizer covers every pixel of [10,194,860,668]; its14 controls include both retained states and ten rejecting mutations. It runs before the later-state personal-review gate. The ordinary preset1 route would remain behind these gates and a fresh physical hover/press/release guard; it was never reached. These are model controls, not measured stability.
The exact unchanged27 observer,20 freshness and12 route controls passed. Actual retained WASM/original-EXE instruction execution in the isolated Worker/broker fixture passed its non-ID2 negative and ID2/request20 positive. No such execution was forced in the browser route. Actual CJS/ESM/Puppeteer imports, helper closure hashes,98-page/12-Worker cache graph, five HTTP/isolation responses and all2415 runtime pins passed. Missing individual predecessor text fixtures were recovered with gzip -dk and original SHA verification using the supplied recovery indexes; archives remain.
Actual run and personally viewed capture
The one Chrome22760 launched at approximately10:46:26Z on bx_jret53j4. The unchanged600000ms TTL check and300000ms expiry reserve passed; its six-minute deadline preceded hardclose10:55. I personally viewed each requested title/menu/ChooseType/Custom Game capture in a separate tool call before writing its capture-bound acknowledgement. The original balanced Custom Game gesture was followed only by input-free stepping.
At 10:48:43.599Z, command8 failed: first unexpected natural scene: unknown. Chrome closed normally at 10:48:43.955Z, exit0. There was no retry, reload, second Chrome, forced guest call, guest write, repaint or guard revision after failure.
After recovery I personally viewed route-native-boundary-6.png in a separate call before writing personal-final-review.json. At11200 it shows Human/Lord Crusader, bright panel borders, the upper-right map, Deathmatch Game, gold10000/10000 and Crusader Demo, together with the unavailable wording. PNG SHA: fc39f8676ac2c1739c53f3b294485ac950fb2699861781865ada84620c434b7c. Complete-output digest: 820a52df229e09525333e254dd3b00b91c63a66c4166ca6183b0ae27acd4de91.
It differs from the retained earlier positive by 1,128,107 RGBA bytes and the retained later positive by 4,753 RGBA bytes. Independent full-layout recognition rejects it. This new intermediate state was not added to the running recognizer or used to authorize input. The concrete remaining dependency is a prelaunch bounded transition model grounded in this additional retained state, followed by actual successive later-output captures and independent recognition within the existing budgets.
actual-evidence-check.json validates175 complete ACKs/11200 steps,1052 served-pin matches, six paused-before-start Worker preloads, four live armed instances,20 reports, two explicitly pre-recording retirements, zero observer errors/drops and eight rejecting proof mutations. Original callbackID2, branch/request20 and down/release ordering pass; final current/requested scene is20/20. Callback absence remains unqualified.
Recovery and custody
All 254 immutable closed files /17,955,727 bytes were recovered and SHA-verified by10:48:53Z. The exact executed driver, unique Worker result and driver log were recovered separately. Post-close checks show no Chrome/driver, actual HTTP41787 closed and all2415 pins unchanged. The initial copied recovery helper checked the predecessor port44145; the final custody check explicitly corrects that to actual41787. An inherited prelaunch custody field said newChromeLaunched:true before launch; that field is not prelaunch evidence. The actual launch identity and final custody records establish the attempt.
Box bx_jret53j4 is left RUNNING, API-idle, healthy and hydrated, expiry 2026-11-04T11:05:35.153Z. Shutdown remains coordinator-owned; no lifecycle action occurred. Custody returns before recovery10:59 and shutdown checkpoint11:00.
Runtime source is 1188e9bf57ef35bd3363fb88b7ca0da442ce696d, normal WASM cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee. Exact checkout Git SHA, relevant binary diff, owned small-file patches and evidence seals retain provenance. No source archives, media/module copies, production edits, authentication changes, local heavy runtime, other boxes, subworkers, successors or deployment were used. Only this note is committed; no production full-build claim applies. This assignment stops at its first failure.