NFS3 current original: native setup reaches race and driving
The current original starts and drives through the ordinary registered route. No native timer abort recurred, and no production emulator change was needed. The historical abort described in the timer note remains unexplained; this successful launch supplies no retrospective initializer or clearing-branch trace.
Evidence: scratch/runs/20261111-nfs3-current-original-race-timer-0850/.
The canonical game/nfs3demo.exe is 1,577,984 bytes, SHA-256
0defab3eeb22ee4b6e0007a4d5b26a99d868008ba77e2b9bd3ef770e924548ad.
The staged runtime starts from main d41587412ac3216bd709474ae50d2e3133561b18;
the current production WASM is 2,189,565 bytes, SHA-256
d0481f03459ec4f612864c57a290a12afb1e22fb09cd476a190dca967a829a10.
All 715 initial runtime/media pins were rechecked unchanged after execution.
Supplemental dependency pins, relevant binary diffs and untracked-file patches
record recovery through Git without a source or media archive.
Ordinary entry and observed consequence
One fresh browser profile runs index.html?candidates&threads, with real guest
Workers. The registered catalogue installer exits 0. The first native setup
invocation selects its own Software default and exits 0. After reload, the second
native setup paints the chooser. Full images were reviewed before each mouse
or keyboard action. The dropdown offers Direct3D Device 1; selecting it,
acknowledging the original unsupported-card warning and clicking setup OK makes
the utility write its native selection and exit 1. Another page reload precedes
the unmodified nfs3_chosen_renderer launch.
The ordinary process policy reports mmTimerThread=true and
asyncMultimediaTimer=false. No readiness, counter, clock, RNG, scene, guest
memory or save state was seeded or edited. No isolated timer fixture was rerun.
The original proceeds through its Corvette loading screen to an actual race.
| Reviewed image | Visible consequence |
|---|---|
prerace.png |
Cockpit on road, 0 MPH, 0.9k RPM, 0:02.53, lap 1/1, second of two |
accelerate.png |
After an 8-second ordinary Up hold and release: 38 MPH, 4.5k RPM, 0:09.78; house and road advance |
steer.png |
After a short Up+Left hold and reverse-order releases: steering wheel turned left, changed road alignment, 16 MPH, 0:15.15 |
The retained command ACKs and page input logs record Up down/up and then Up/Left down followed by Left/Up up. These establish acceleration and visible steering/track progress; no lap completion, finish, audio or FPS claim is made.
The browser directly executes the released Chrome binary with explicit SwiftShader flags, without the fingerprinting wrapper. Game D3DIM telemetry reports a legacy WebGL endpoint, ANGLE SwiftShader, and zero software-D3DIM fallback completions/errors in the driving snapshots. This is a software GPU route, with no physical GPU or historical renderer identity acceptance implied.
Observer boundary and recovery
Startup and all driving captures use no guest Worker attachment, debugger,
import wrapper or block tracing. Ordinary page-level state reads and screenshots
still have observer cost. An optional post-driving Worker/module census
later times out at Runtime.enable; it exits 1 and is retained as UNKNOWN,
without replay. Browser shutdown followed before that outer census process
was joined. There is no successful executing-module census or native timer
entry/API/counter trace from that attempt. Runtime identity is the pinned served
closure; the identical module is recoverable from the preceding AVP2 run's
final-module.wasm.gz, freshly inflated and hash-checked here. Its matching
locked build and focused normal/compat regressions are referenced, not rerun.
All 101 remote artifact paths were downloaded and independently size/SHA-256
checked before stop. PNG CRC and complete IDAT inflation checks pass.
recovery-map.json and duplicate-map.json preserve byte-identical paths after
deduplicating early capture downloads. Unique logs and the failed census remain.
Chrome exits 0. The runtime owner actually waits for controller exit 0 and both stdout/stderr EOFs. A separate remote check finds controller/Chrome PIDs absent, the profile removed and the server port closed. This is browser/runtime disposal, not a claimed native guest ExitProcess during the race.
The independent watchdog was observable before the fresh private allocation, with an immutable deadline of 09:36:09 UTC. After evidence verification, delegated normal stop completes at 08:58:32 UTC, forced=false, error=null. Final box bx_9nwfkcya is independently confirmed ARCHIVED, expiry null (initial expiry 2026-11-11T09:36:09.832Z). Stop receipt: stop_7be83a65f7c2669c. The watchdog exits 0 unfired with an actual parent wait and both EOFs. No box remains billing; no resume, deletion, public deployment or workers occurred. Next useful work is ordinary race completion; timer-cause instrumentation is needed only if the native abort recurs.
Follow-up: native Hot Pursuit race ends in arrest
scratch/runs/20261111-nfs3-native-lap-finish-0910/ reaches a real losing
race outcome: the original game displays ARRESTED, then its demo end card.
This establishes native pursuit consequences and race termination, not a
completed lap or finish-line crossing. The last cockpit image still says
lap 1/1, at 7:48.81, with tickets 1/2 before the final arrest.
The same catalogue → native setup → Direct3D Device 1 → original warning OK
→ setup OK → reload → nfs3_chosen_renderer route was used. The canonical
EXE hash above is unchanged. The served runtime is pinned at main
6fdf10b4697f1d63de4af581ae9239705692e336; production WASM SHA-256 is
5512a65e2ce88cc76b39b3246d35f13919d9df060d3670a9a85cae9c4100f638.
The intervening Surface7 priority HRESULT fix has matching locked-build and
normal/compat regression evidence in
20261111-surface7-unmanaged-priority-hresult-0900. This worker changed no
production code. Served closure pins, Git patches, and retained tested modules
record identity; no executing-Worker/module census was attempted.
Ordinary Up, Left/Right, Down and R controls produce sustained driving, roadside collisions, a rollover and native road recovery. The first race was restarted through page reload after a pause-menu navigation attempt stayed in the menu; those images and input attempts are retained. The second ordinary launch randomly selected Hot Pursuit without weather. No RNG, guest memory, clock, readiness, save or race state was edited. The page-only driver never attached to a guest Worker. Chrome uses the explicit SwiftShader route above.
Review-boundary limitation: twelve R→Up command pairs were dispatched within
one orchestration call. Their intermediate recovery images were inspected
after dispatch, although the generated review annotations say “inspected.”
Those annotations overstate pre-action review; the pairs were based on the
preceding reviewed crash image. review-audit.json lists the exact pairs.
The first prerace annotation also mistakenly called the introductory exterior
view a cockpit; the following command explicitly corrects it. These limitations
do not change the retained native arrest/end-card images or input ACKs.
Reviewed image under recovered/ |
Visible consequence |
|---|---|
drive12.png |
87 MPH, 0:13.06, lap 1/1, second of two |
drive49.png |
“Wild One arrested”; position changes to first of one, 52 MPH, 3:36.96 |
drive56.png |
Native TICKETED cutscene |
ticketsettled.png |
Cockpit resumes, tickets 1/2, 4:18.31, first of one |
drive75.png / drive76.png |
Rollover, then ordinary R recovery to the road |
drive78.png / drive79.png |
Lit tunnel traversal, then open road at 82 MPH, 7:11.40 |
drive83.png |
Last cockpit context: 7:48.81, 1 MPH, lap 1/1, tickets 1/2 |
drive84.png / drive85.png |
Native ARRESTED cutscene, then demo end card |
No checkpoint counter change, completed lap, win, audio or FPS result is claimed. No save option was offered on the observed outcome/end card, and no save was obtained. Enter dismisses the end card to a black image; the page still reports a running process without a trap. Native ExitProcess after that dismissal is unverified. Browser/controller disposal subsequently completes normally; this is a distinct lifecycle observation.
All 666 remote artifact paths were recovered and size/SHA-256 checked before box stop; 109 PNGs pass chunk CRC and complete IDAT inflation checks. The recovery and lossless-compression maps preserve original paths and hashes. All 726 runtime/media/driver pins are unchanged at final recovery. An initial controller start failed before Chrome because its pin manifest was omitted; the actual exit was confirmed, the missing manifest uploaded, and a repaired start used the same fixed deadline. That failed start and the premature read-only shell check remain preserved without replaying an uncertain action.
Chrome and controller exit 0 with actual parent wait and both EOFs. Independent checks find their PIDs absent, profile removed and server port closed. The watchdog was observable before allocation and retained its immutable 09:48:22 UTC deadline. Delegated normal stop completes at 09:36:50.990 UTC, forced=false, error=null, receipt stop_13d04591d8c9db0c. Box bx_q9gvknan is independently confirmed ARCHIVED, expiry null (initial expiry 09:48:23.104 UTC). The watchdog exits 0 unfired with an actual parent wait and both EOFs. No box remains billing. No workers or public deployment occurred. Next useful work is a finish-line crossing; lap completion remains open.