Alice: natural production-browser startup reaches controllable player HUD
Run scratch/runs/20261118-alice-current-browser-player-hud/ reached the actual
third-person player HUD through natural startup cinematics, New Game, Easy,
loading and the complete opening intro. Reviewed original2/player-neutral.png
and original2/after-w.png show Alice moving forward beneath the timber beams
after the recorded W pulse, before the later A/S harness problem. This is player
movement, not the autonomous intro camera. Jump, attack and ordinary Alice Quit
remain unverified. The terminal blocker is the browser controller's release
prearm deadline; no Alice runtime defect is established.
The frozen published source was 57af0a44ca06dc27e930c4e2e0fd94cfe8dbd847.
Its module was 2,411,934 bytes, SHA-256
1a8a78d4dd4508604ac4949eb54df79ede5496786074984ebf35e2275ceeaf44.
prepared.json records the Git inventory, tested.wasm retains the unique
compiled input, and worker bytecode/HTTP receipts record the actual normal
guest/render Worker graph. The normal URL was
index.html?debug&threads&gl-renderer=webgl&app=alice_demo. This was a local
browser serving immutable published production source, not a deployment or a
claim about a remote public website. Existing sixteen media pins were rehashed
unchanged; media-pins.json records paths and hashes. Foreign worktree changes
were excluded from the served graph and recorded in relevant-worktree.patch.
No game configuration, registry, media or guest state was patched.
Before either original, sequence-plan.json fixed a 60-minute runtime, a
five-minute ordinary Quit reserve, and an independent finite native supervisor.
There was no extension. The same source/module normal-browser authored fixture
passed reviewed Enter, independent release, visible color change and ordinary
Escape/actual ExitProcess(0) in calibration2. That is calibration evidence,
not Alice Quit evidence. Calibration1 rejected input before DOWN because a
Quit-reserve check was mistakenly applied to the short fixture window.
The first original closed during capture before any input. Its supervisor
recorded exit 1 and no leaks; the outer tool returned 143, with no retained
browser wait/EOF receipt or established controller cause. original-closure.json
keeps that gap explicit. This closed scope was not queried or replayed.
The second original began at 07:20:34 UTC and closed at 07:43:32. The natural menu, difficulty menu, corridor/Rabbit intro and grounded HUD were individually reviewed. W, Space and navigation inputs used an independent release process armed before DOWN, with document/process/window/Worker/module lifetime checks. CDP replies are transport replies; they are not fabricated guest acknowledgments. Space inputs and a selected capture during the hold did not resolve clear airborne motion. Unarmed Alice was visible; knife acquisition and attack effects were not established.
There was a concrete run-local release-helper error: A/S DOWN used KeyA/KeyS
but UP used a/s. Transport replies therefore concealed physically held A/S.
keycode-oldfail.json and original2/jamb-progress.json retain that failure.
Movement from clear-upright through retreat-from-jamb is contaminated and
cannot establish Alice navigation behavior. Matching ordinary UP events cleared
both, as recorded in matching-release-state.json. The helper was corrected to
matching codes and now checks passive physical state after release. Fresh W on
the same original then visibly advanced Alice into the tunnel:
neutral-after-release-correction.png → corrected-forward-result.png;
corrected-forward-up.json observes W released. The matching-release direct
tool wait returned 0, but separate stdout/stderr EOF values were unavailable.
The following corrected Space release also observed physical UP; jump remains
visually inconclusive. This was a controller correction, not emulator code.
The last proposed Mouse1 operation was NOT_SENT: its delivery receipt has
at: null, error release prearm deadline. The child eventually armed and
released with actual wait 0/both EOF, but missed the parent's fixed 1,000 ms
prearm acceptance. The parent closed the browser on that failure, before the
planned ordinary Quit. Browser.close/host cleanup must not be called game Quit.
original2/failure.json, ordinary-mouse-attack-delivery.json and its release
wait preserve this bounded next cause. No subsequent endpoint request was made.
Next work must make the prearm acceptance accommodate measured attach/identity
latency while preserving an independently finite release and rejecting expired
leases before DOWN; calibrate it before a fresh original. This run does not
justify an interpreter optimization or guest polling change.
Second-original browser, Xvfb, Openbox and supervisor actual waits have both stdout/stderr EOF. The supervisor returned 1, reports no leaks, and did not hit its immutable deadline. Selected whole frames, input boundaries, trusted-event logs, immutable target identities, and cleanup receipts remain in the run. No boxes were created: ID/state/expiry/stop receipt are all null. No Wine, agent workers, other games, public deployment or optimization variant was used.
Recovery uses the recorded Git commit plus prepare.cjs, the existing media
paths/hashes, retained tested.wasm, and run-local controller inputs. There is
no source-tree/media snapshot. Focused existing Alice prepared-control,
native-window and local-original tests passed (3/3); the run-local old-fail /
new-pass evidence check and locked build are recorded alongside the result.