NFS3 native game-control observation, 2026-11-18
The fresh original localizes part of the input path but does not establish a runtime defect or a game-behavior repair. Up and Right reached NFS3's scan state and generated control-packet bytes; Escape reached scan state without visible Pause. The unresolved boundary is packet admission and simulation consumption, including the game's Pause decision. USER delivery alone was already insufficient evidence and remains so.
Evidence is scratch/runs/20261118-nfs3-game-control-poll-route-cause/, with result.json published last. The observer/test commit is c9d0c2cf720cc13b962ea8b9c2565de0b480d319. No input runtime file was changed. The published 70062 frame-routing oldfail/newpass and actual rendered Pause/Continue/ordinary-Quit fixture, async9191, and loader0888 proofs were reused offline; no closed scope was reopened.
The bounded, opt-in observer records actual native dispatcher entry/return, successful stdcall completion, thread, DirectInput configuration, returned state/data, and selected read-only original-game bytes. It wraps synchronous construction and asynchronous instantiation, reads results after the existing handler, and adds no native polling, guest writes, yields, or forced state. A preparation defect where inherited report writers could overwrite one another was corrected with process/host-thread filenames before publication and the original. This is an observer correction, not a game-control repair.
The focused test verifies constructor/async wrapping, ordinal resolution, incomplete-return exclusion, bounded overwrite, and identical memory, returns, and 12 import calls with observation off/on. An authored real native software/cooperative PE additionally exercises GetKeyboardState, standard keyboard DirectInput state, custom DIK C8 at offset7, buffered down/up data, and main-thread USER versus worker-thread DirectInput state. The final off/on executions both pass guest assertions and exit0, with 678 API calls, 66 batches, and identical whole PNG SHA256 8e22aa095718e8bea2ecf77eacdf270a24f4b6c8a6da1269b45b1c079f4430af. The on report has 285 rows and no observer errors. The worker sees held custom state; buffered data is observed on the acquiring host context, with two offset7 records, values128/0. The other host context returns zero buffered records. That authored cross-thread observation does not establish an NFS3 defect: the original's measured DirectInput device is a mouse on thread1, not that keyboard configuration. Earlier failed GUID, scan-code, report-writer and compile preparations remain explicitly excluded from passing controls.
The locked build and focused observer test passed before the code commit, followed by strict pull/push. The original used Git source c9d0c2cf7, whose runtime src/, lib/, host.js, and test/run.js are identical to published70062. Its retained module SHA256 is 7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce; actual source-loader records cover main and inherited execution hosts. Genuine chooser pins and all135 media pins were checked unchanged before and after. No registry seed or guest-state edit was used.
Exactly one fresh registered nfs3_chosen_renderer original ran. Whole frames were manually reviewed through loading, exterior intro, countdown, and active race before input. Each finite gesture followed a fresh reviewed whole frame and retained the existing identity, operation, freshness45s, native confirmation800ms, and independently armed release800ms guards. The following milliseconds are intent-file time to the actual boundary, not game-effect timing:
| Gesture | Host apply | Native USER return | Game-consumer evidence |
|---|---|---|---|
| Up | 63 | 70 | scan48=1; generated packet [0,255,0,0] |
| Left | 97 | 102 | no native snapshot overlaps the hold; unmeasured |
| Right | 68 | 71 | scan4d=1; generated packet [127,0,0,0] |
| Escape | 156 | 209 | scan1=1; running race, no visible Pause |
| Diagnostic Up | 241 | 258 | held read-only queue/car snapshot; packet generation not sampled during this hold |
All five USER returns are thread2/HWND0x18001, and all independent release helpers actually exit0 with separate stream EOFs and acknowledged release. A malformed local command before the first actual Up failed before sending an endpoint request. There was no transport UNKNOWN, replay, deadline widening, or input after closure. The second Up was a distinct finite diagnostic gesture after frame review, not a replay. Its release request was independently armed before the read-only inspection.
Actual original DirectInput configuration is a 16-byte, seven-object mouse format, cooperative flags5/HWND0x18001, with successful acquire. Loading-time GetAsyncKeyState calls query F11, not the driving keys. The sampled installed bindings include scan48/4b/4d. Native transmission state becomes1; this does not prove a gear issue or identify a reviewed UI selection.
At epoch milliseconds1794968511352, 97ms after the first Up USER return, two existing mouse API returns expose scan48=1, pollCounter3114, and generated packet [0,255,0,0]. At1794968559115/116, Right exposes scan4d=1, counter4502, packet [127,0,0,0]. Escape exposes scan1=1 at1794968582669, counter5019. These are game-produced bytes beyond host apply and USER return, sampled at native mouse API completion; they are not execution hits inside the private packet or simulation routines.
The 4096-row ring explicitly overwrote25421 of29517 observations in the final live snapshot, while configuration and aggregate API totals remain separate. Actual mouse-return timestamps and successive game-counter deltas are retained in measured-consumers.json. The counter advances between returns, but individual keyboard-binding read times are unmeasured. Some mouse-return gaps exceed the800ms hold; absence of a held Left sample is therefore inconclusive. No artificial game polls were added.
Every retained simulation-packet sample at0x79a4a8 is zero. A separately tested read-only queue/car inspection preserves its test memory byte-for-byte. During the second Up, inspection at1794968693044, 173ms after USER consumption and before the release due time, finds player0 queue sequence768, four pending neutral packets, and selected-car control bytes [0,0,1,0]. The previous native mouse snapshot is1794968691748 and the next is1794968694274, after release. Consequently this snapshot cannot establish that a nonzero packet was generated during that second hold, never admitted, or never consumed between samples. Configured callback/apply pointers are retained as data, not claimed execution evidence. The queue had advanced from sequence696 in the earlier inspection; no gate-failure inference follows from neutral entries alone.
Reviewed whole frames show an active HUD at0:01.0/0MPH before Up and0:02.25/0MPH afterward; subsequent frames show1MPH and eventually2MPH with no attributable steering effect. Escape is followed by running-race frames at0:08.87 and0:11.25, not Pause. Natural speed drift is not attributed to throttle. No assumption about sufficient throttle duration, binding failure, or transmission explains this result. Continue and ordinary Quit were not attempted without a verified Pause target. Final renderer counters report647 presents/61801 draws, zero fallbacks/readbacks/declines; presentation is distinct from control effect.
The concrete blocker is a missing execution-boundary observation linking generated packet admission to the selected simulation consumer and the Pause decision. No relevant runtime defect was isolated, so no behavior fix or second original was justified. The fixed390-second native lifetime ended normally at the harness limit; native, supervisor, and independent finite/disk watchdog all actually waited0, both streams reached EOF, and no tracked process leaked. A later cleanup SIGTERM attempt returned ESRCH and sent no signal; it is not credited as shutdown. All51 known owned process IDs are nonlive, the original listener is absent, and final home free space is over2.64GB, above1.5GB throughout the watchdog observations. Selected named whole PNGs, unique authored PE, exact Git manifests/owned patches, module/media pins, actual waits and recovery references are retained; no media copies, source snapshots, or frame series were made. Box ID/state/expiry/stop receipt: null/NOT_CREATED/null/null.