Alice original demo: local browser registration and reviewed startup

The already-local American McGee's Alice demo is registered as alice_demo. On the local browser page, enable the existing experimental candidate picker with ?candidates, select “American McGee's Alice Demo (2000, experimental)” and use ordinary Launch. The public inventory fence remains intact. This is not a retail qualification or a deployment.

Evidence is in scratch/runs/20261118-alice-local-original-browser-gameplay/. The executable is 1,454,142 bytes, SHA256 61857b521eb7260c28b3ae6960fa8ee18d4d3679b6190f5221131e382e0f3a72. Its version resources report FileVersion/ProductVersion 1.0, comment 1.0a2; the PDB path identifies demo_release, the readme is dated November 27, 2000, and the PE timestamp is November 28. All fourteen declared companion files were present and individually pinned, including the original pak0.pk3 and both game DLLs. Existing media was unchanged. No replacement assets, configuration, registry changes, launch commands or downloads were introduced.

Registration uses the existing browser manifest, required-file validation and relative mouse support. Main and ordinary Worker preflights validate all mounts before guest execution. An actual browser selector inspection confirms the option and registry fields. Actual guest and renderer Worker WebAssembly byte readbacks match the tested module, SHA256 46c9b8e4af91a8770d454ff529e21d6a17553c31ad57aea30ca4a8aa6bf3e311. The graph was frozen from committed revision 2b8ba1b675a6fc0ce286d48a6261dc04acef62dd with only the published Alice registration/helper overrides. Shared foreign WIP and the concurrently generated build module were not adopted for runtime.

An authored browser GL control reached initialized publication, but public capture exceeded its three-second display deadline. That browser scope closed UNKNOWN, without input or later endpoint requests. It is not a successful browser capture qualification. The independent native software GL-stream route passed personally reviewed authored before/after pixels, a consumed W pulse, independent physical UP acknowledgement and ordinary Escape exit. Hardware OpenGL packages were unavailable, so this is no hardware GL claim.

Registration and the finite input helper were published as b12608a7b after focused tests and a locked full build. The first original reached the console and game window, but premature presentation capture exhausted its unchanged three-second readiness budget with focus.mainTopLevel. It closed without a capture, binding or input. No endpoint was queried after UNKNOWN.

A separate finite passive startup wait was then added to the Alice helper. It calls only the existing startup-readiness observer, retains complete snapshots and verifies process identity, physical UP, screen and ordinary USER ownership. It cannot capture, bind, focus, pump guest messages or deliver input. Its deadline is at most sixty seconds and precedes the fixed Quit reserve. The original three-second capture and input deadlines are unchanged. An authored executable spends 4.5 seconds in an ordinary GetTickCount loop before creating its window: immediate capture fails at three seconds, while passive wait followed by a new bounded capture passes. Both runs use identical executable bytes and no input. Unit controls also reject foreign PID and held input. These tests and the locked full build passed before publication as 98ff811e9 and the one permitted additional original.

The additional original produced a personally reviewed whole startup-console PNG (affected-extra/main-menu.png; despite the filename, this is not a menu). Startup then created the separate live, visible game HWND 0x10007 and naturally focused it. Initially the live, visible console HWND 0x10001 remained main. The initial passive readiness samples reject precisely focus.mainTopLevel: focusTopLevel is 0x10007 while mainHwnd is 0x10001. Both windows have the same process/thread identity; initially ordinary USER ownership and actor context are healthy. Later the original naturally hides the console, still retained as main, causing main.visible and owner.ordinaryUSER failures. It then acquires mouse capture on 0x10007, adding input.capture. Physical keys remain UP. The helper closes on that capture transition after 510 passive samples (about 55 seconds), rather than waiting out or extending its deadline. The console reports successful GL context creation, renderer/client initialization and sound initialization. This establishes a concrete admission blocker, not a claim that rendering or game loading has failed. Native scene admission requires focus to belong to main when no modal exists (lib/cli-native-scene.js); window creation retains the first top-level as main and ShowWindow does not replace an already visible ordinary main (src/09a5-handlers-window.wat). No focus/main promotion or identity fence was changed to bypass this condition. Gameplay pixels, movement, jump, enemy consequence and ordinary original Quit remain unverified.

Both originals use normal Workers, real ticks, the exact frozen module and unchanged original media, with native supervision at 900 seconds, an independent 920-second watchdog and a 180-second Quit reserve. Closed scopes are cancelled through the owned supervisor and actual native/watchdog waits and both EOFs are retained. No original input was bound or delivered. There were exactly two originals, and neither is eligible for replay. No test boxes were created: ID, expiry and stop receipt are null, state NOT_CREATED.

Recovery uses Git, not a source archive: restore the graph from revision 2b8ba1b675a6fc0ce286d48a6261dc04acef62dd in memory, apply only Alice's published registration/helper changes, compile with the recorded compiler and verify the module hash above. The retained preparation scripts, module source hash list, source-read receipts, media pins and own Git patches document those inputs. The original local media must pass those pins; no media copy is retained.

Evidence retention incident: the run PNGs, binary files and the full 510-sample readiness JSONL disappeared before finalization; cause UNKNOWN. The exact module and authored fixture were reconstructed from Git and the retained recipe without executing a guest, with hashes matching pre-run pins. Three selected personally reviewed PNGs (console, authored input consequence, authored delayed readiness) were recovered byte-for-byte from this session’s persisted image tool responses, each matching its original capture hash. The final two readiness snapshots and sample count were recovered as exact prior tool output; the full 510-sample JSONL remains missing. Recovery receipts record source line and record/output hashes. No guest, endpoint, capture or image transformation was used for recovery. This limits independent review of the complete passive timeline, but preserves the selected original pixels and final observed failure details.