NFS3 native timer: unchanged module passes the offline routine fixture
No production mismatch was reproduced, so no production change was made.
Evidence: scratch/runs/20261109-nfs3-native-timer-cause-0724/.
This is a bounded invocation of original timer instructions, not an original
game launch or a reproduction of the historical startup environment.
The previous original abort remains unexplained. Its prior API returns, readiness-counter history and executed clearing branch are UNKNOWN. Its missing per-case startup manifests remain missing. Neither attachment nor a zero clock has been established as its cause.
Executable path and experiment
The existing test/binaries/candidates/need-for-speed-3-demo/game/nfs3demo.exe
(1,577,984 bytes, SHA-256
0defab3eeb22ee4b6e0007a4d5b26a99d868008ba77e2b9bd3ef770e924548ad)
contains the timer initializer at 0x48dad0. Its original caller at 0x4f6e2e
loads 0x562ae8 into EAX before calling it. The fixture parses that existing
PE, never runs its entrypoint, and uses a small generated caller to perform
the same load and call. The PE's initial field is zero; the initializer itself
selects its normal 100 Hz fallback. No readiness/counter field is seeded.
Both cases instantiate the same retained production CPU bytes, SHA-256
7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf
(2,182,452 bytes), referenced through the preceding run's runtime-pins.json.
Production JavaScript dependencies are hash-checked and recovered in memory
from their retained files or Git 0a38cf48438096482d059c5d8f27c048474c83b9.
The fixture executes the retained browser-shell timer-policy assignment and
host timer-mode boot block, with the actual nfs3_chosen_renderer entry and an
empty query. They select mmTimerThread=true, asyncMultimediaTimer=false.
This is ordinary production configuration, not an experimental timer override.
Actual production createHostImports and ThreadManager implement thread,
event and wait behavior. The fixture uses the supported cooperative backend,
with the real default host clock and ordinary manager sleep/wait handling.
It does not reproduce Chrome's Worker backend, whole-game prehistory or the
old attachment procedure. No API success, branch, clock value, scheduler state
or guest timer state is forced. Calling the routine in this isolated context
is a deliberate fixture boundary, not evidence that the historical original
reached it with identical state.
| Observed boundary | Diagnostics off | Diagnostics on |
|---|---|---|
timeGetDevCaps, return at 0x48db2c |
0 | 0 |
CreateThread, return at 0x4ddcde |
0xe1001 |
0xe1001 |
Native thread wrapper, return at 0x48db99 |
1 | 1 |
timeBeginPeriod(1), return at 0x48dbbe |
0 | 0 |
Readiness 0x562ae8 on return |
100 | 100 |
Readiness counter 0x79a1a4 on return |
1 | 1 |
Retry counter 0x55c9dc on return |
0 | 0 |
Both cases execute the existing native callback and helper-thread machinery.
The diagnostic trace contains 74 distinct block/state observations, including
the unchanged 0x48d670 helper and 0x48d810 TimeProc. The off case has no
log_eip wrapper or enabled block trace; both cases sample exports at ordinary
CPU slice boundaries. This establishes matching observed timer outcomes,
not whole-game passivity, identical elapsed times or full register parity.
The measured wait reaches 0x48dc7b with EBX's captured counter zero and the
current counter one, then takes the changed-counter path to 0x48dc9f and
returns with readiness intact. The timeout cleanup at 0x48dc84 is not taken.
The static abort predicate at 0x4f044b was not executed by this fixture.
Rejected fixture evidence and limits
Earlier fixture iterations omitted the retained launcher's timer-mode boot
policy. The raw module's zero-initialized mode then supplied no independent
timer callback delivery in this isolated loop: the counter remained zero,
the native timeout called cleanup, and readiness cleared. Those logs are
retained under no-launch-policy-*; they are an artificial fixture failure,
not old production FAIL evidence. The helper thread initially waits on its
event; the later zero-EIP log follows timeout teardown and does not demonstrate
a missing initial thread argument. Direct observations confirm the real
argument is present before its first slice.
The first diagnostic preparation installed a host wrapper after instantiation,
so it captured no imported trace calls. preflight2-* retains that invalid
diagnostic attempt. The final wrapper is installed before instantiation.
preflight1-* retains a dependency-hash refusal before CPU execution. No
rejected run is promoted into original execution or a repair comparison.
The missing evidence is a historically bound record of initializer entry state, process timer policy, API arguments/returns, thread/TimeProc progress, counter changes and the exact clearing branch on the aborting original's executing Worker, together with a matched observer-off original control. The isolated passing pair cannot supply those facts. There is no justified old FAIL/new PASS production patch and no unchanged original retry.
Recovery and lifecycle
base-commit.txt, small run-input patches, relevant-source.patch, source/media
references and SHA/size verification retain recovery without source archives
or media/module copies. The old modules and evidence remain at their existing
paths. Logs are gzip streams from their first byte; integrity verification
fully inflates them, checking gzip CRC as well as SHA-256 and size.
Each local fixture has an independent deadline watchdog prearmed before its
CPU child. The supervisor separately waits for child exit and joins stdout
and stderr gzip writers; both final children exit 0. Cooperative WASM thread
instances end with their owning child process, including the idle timer
thread; this is bounded fixture disposal, not a claimed guest shutdown.
Watchdogs are terminated and actually waited for. No Chrome, original launch,
box, native-custody change, peer provider edit, worker or deployment occurred.
Box ID: none; state: NOT_CREATED; expiry and stop receipt: N/A.
Free space on /home remains above 1.5 GB. Single task stopped.