Redline native relative aim and car-door boundary — 2026-11-08
Native mouse turning works; vehicle entry and driving remain unverified. Trusted pointer-locked mouse deltas reach the original game's owning DirectInput mouse state, and personally reviewed views show small heading changes while retaining the car. A forward approach reaches the rear quarter and the native instruction “TO ENTER A CAR USE THE DOORS.” A later forward hold returns 90 owning VK38 high-bit reads but produces no reviewed displacement, including after passive settling. Gameplay input stops at that measured approach boundary. No visible side door, door contact, changed vehicle HUD/camera, acceleration/braking/steering consequence or vehicle exit is claimed.
Evidence is retained in ignored scratch/runs/20261108-redline-car-door-native/. Principal records are relative-qualification.json, mouse-consumption-summary.json, poll-approach-right-rear.json, poll-past-rear.json, retrieval-prestop-verified.json and normal-stop-confirmed.json. Reviewed images are under review/; complete recovered originals and traces are under retrieved/original/, with large logs individually losslessly compressed.
Local qualification and original controls
The shipped test/binaries/candidates/redline-demo/original/MAINDIR/Readme.rtf recommends mouse driving, documents left mouse for fire, right mouse for forward movement, arrows for movement/acceleration/braking/steering, and entry by walking into a driver/passenger side door. No binding, remapping, guest/configuration write or forced entry is introduced.
Before box allocation, actual Chromium executes a small ordinary DirectInput PE in the production guest Worker. A trusted click acquires native pointer lock and an independent timer releases its button. A trusted (7, -3) movement persists for 2300 ms without a guest poll. Subsequent native GetDeviceState calls accumulate exactly (7, -3) once, with zero additional displacement on the next slice and the same Worker endpoint. The fixture's guest code owns its input results; it is never substituted into Redline.
The first local experiment demonstrates that CDP's explicit movementX/Y fields are ignored here: its trusted event reports zero motion. The corrected route changes CDP coordinates, producing trusted DOM movement deltas while locked. Both outcomes are retained. The original-game helper admits only nonzero integer deltas bounded to 24 counts per axis, requires existing native pointer lock and a previous CDP position, and sends motion with no buttons held. Invalid/unlocked controls produce zero dispatches. Pointer lock in the original is acquired through production browser-input handling of a native click; no scripted lock request is injected into the original game.
The qualified 5000 ms maximum keyboard helpers are byte-identical to the predecessor. Original movement here uses 3500 ms holds; no larger ceiling is introduced. Independent key/button release, strict 18-field identity, module/image checks, freshness, command/CDP/ACK/UNKNOWN/browser/box deadlines and no-replay behavior remain intact. Owner/stale/exact-ACK/sticky-UNKNOWN tests pass. Actual-WASM observer tests verify exactly-once original imports, preserved returns/exceptions/instance/memory, and read-only capture of the mouse result buffer on the next import. Exact-ready startup and the existing Worker DirectX-vtable boot regression pass. No emulator code changes or performance/semantics repair are claimed.
Two local setup errors are retained with their corrections: a missing copied ACK helper and an outdated startup-test stub lacking the new passive mouse-event reader. Neither issues original-game input or creates an original command UNKNOWN.
Actual native mouse consumption and reviewed approach
One fresh browser follows individually reviewed Training Levels → Vehicle Training, spawning on foot with the car rear directly ahead. A native center click acquires pointer lock; its independent button release occurs after 150 ms. That click is not a firing success claim.
| Relative command | Trusted DOM UTC | Counts X | Delay to owning mouse-state poll |
|---|---|---|---|
| aim-right-seven | 04:38:51.545 | 7 | 3 ms |
| aim-clear-right | 04:39:35.249 | 14 | 3033 ms |
| face-flank-one | 04:41:24.348 | -24 | 2981 ms |
| face-flank-two | 04:42:24.614 | -24 | 1 ms |
| face-flank-three | 04:42:57.092 | -24 | 0 ms |
| line-along-side-one | 04:43:49.630 | 24 | 2 ms |
| line-along-side-two | 04:44:23.772 | 24 | 2 ms |
All seven trusted events have pointer lock, zero buttons and zero Y movement. Each has exactly one matching retained nonzero owning 16-byte mouse-state result, followed by zero. Native return address is 0x48f066, buffer guest address 0x0881dafc; the original imports are called exactly once. These are read-only post-handler buffer observations at the next import, not an independently captured HRESULT. They establish actual game input consumption, rather than merely host delivery. The three-second cases demonstrate that relative motion survives a slow interval until this game's normal poll.
Presentation can lag consumption. The first capture after the 14-count adjustment and after the first -24-count adjustment shows no new heading; passive settle captures then show the changed heading. No duplicate input is issued to obtain those views. Subsequent dependent inputs use the reviewed changed views.
The initial forward interval is 04:40:40.407–04:40:43.906Z: 129 owning VK38 high-bit reads, 43 completed slices, and a later zero. Reviewed approach-right-rear.png shows the car close at the left and the native door instruction. Small left turns expose more rear-quarter body and tail light, but do not expose an identifiable side door. Two reviewed right adjustments restore a forward line beside that visible rear quarter.
The final forward interval is 04:44:54.211–04:44:57.711Z. All 90 in-window owning VK38 polls return 0x8000, with 32 completed slices. The first high result is 04:44:57.088Z; the last is 04:44:57.230Z. Native stack return is 0x48eefe with a 256-byte keyboard-state argument. Zero results bracket the window at 04:44:54.208Z and 04:44:59.889Z. Trace retention covers both endpoints without observer errors. This is not a missed short pulse.
Reviewed line-along-side-two.png, past-rear.png and past-rear-settled.png retain the same car rear-quarter and terrain alignment, on-foot weapon and health100. The final passive capture completes 04:45:47.112Z without showing displacement. Gameplay input stops: no repeat, longer hold, extra firing or control escalation follows. Collision with the rear/corner is a possible explanation; collision geometry, guest simulation state and underlying cause are not established. Door-contact entry is never reached. The next dependency is a reviewed path that clears this rear/corner boundary and exposes a side door; any additional native strafe route should first qualify its actual guest consumption locally.
Provenance, retrieval and shutdown
Tested Git SHA is 9842f2e8aff8cdcb6ac8691ba5b6dc38fb767661. relevant-source.patch, new-wininit-boot.js.patch, run-input patches and dependency-helper pins preserve recovery; Git is the source archive. All 694 non-run runtime/media pins equal the predecessor, including 65 rehashed original media files. The 705-file launch closure is verified remotely. Nine bounded segments assemble original Redline.bgd SHA-256 840dc1421e88708bb64de686e6a996e2638c27693b25ecdb54237824c64808aa and are removed before launch. No source/media archive or repeated media copy is retained.
Actual guest and renderer module SHA-256 remains f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121; recovered modules are retained. All 18 owner identity fields match from Training spawn through final capture. The owning provider is WebGL2, legacy/WebGL, 640×480, ANGLE Intel UHD620 / Mesa23.2.1, fallback0. There is no paired-backend or FPS conclusion.
Verification independently rereads 362 artifact references / 325 unique hashes / 254,996,024 reference bytes, checks 23 PNG CRC/IDAT streams, all 23 exact successful ACKs, seven relative movements, both forward polling boundaries, original media and browser/process drain before requesting normal stop. Retrieved transfer storage is about16.8 MB. All sampled final physical keys are up; the acquisition mouse-up and motion events record buttons0. Final input-queue length was not independently sampled. No original command UNKNOWN, refusal or replay occurs.
Chrome exits0 at 04:46:10.601Z. Drain confirms controller6276/Chrome6288 absent, HTTP42047/CDP33639 closed, and profile/writer lock absent. Fresh box bx_jb827hmf is normally archived04:47:55.464Z, receipt stop_ab07fb642497d085, completed04:47:55.485Z, forced:false. Final expiry is null; initial expiry was04:59:46.650Z. Observable independent watchdog and normal-stop delegates are armed before allocation; both exit0 and supervisor has no remaining children. Local disk stays above5GB. No RAM gate, old403 action, resume/delete, nested worker or public deployment occurs.
No Redline corpus manifest entry exists, so evidence remains investigation metadata rather than a result for an unrelated candidate. Vehicle entry/driving, FPS, audio, parity and full playability remain open.