MW3 menu polling boundary

The failed 20261111-mw3-ready-observed-throttle-0255 target delivered 16 trusted motions totaling [-128,-23]. Renderer coordinates changed from [320,240] to [192,217]; the game's software cursor remained [320,240]. All four completed mouse reads preceded motion. This establishes an absent standard mouse-read stream during the target, with the recorder's existing binding, capacity and entry/exit checks clean. It does not establish input loss.

The retained dispatch trace has 60,950 matched API entries/exits, four com:1093 entries and four mouse completions, zero non-mouse GetDeviceState entries, four com:2864 Poll calls and 14,509 com:1094 GetDeviceData calls. The old recorder did not decode callers/devices for GetDeviceData or record game gates. Its aggregate categories have no interval timestamps. Consequently the trace cannot identify what closed the original mouse polling path. The active mouse reads themselves identify this=0x082220a0, type 7/deviceType 2, standard relative size 16, state buffer 0x600f68, caller 0x4fa56c, HRESULT 0.

Static evidence comes from mech3demo.exe, SHA256 df457ee6973f6f9d557c57aa3fc14258303486ec8b8742fd37b4369c1f37dbdc:

mw3-live-mouse-observation.cjs now accepts an optional decision decoder. mw3-menu-poll-decision.cjs supplies a read-only decoder for this exact game's gates. It records DirectInput method, caller, device identity/type/flags and mouse/keyboard gate values. A byte signature checks the authentic mouse gate and its call instruction before sampling. Existing full media, module and Worker identity checks remain necessary. Named and ordinal dispatch are covered.

Identical entry samples aggregate into a bounded map (default 64, maximum 256) with count and first/last wall/monotonic timestamps. Overflow increments drops and stops the recorder; decoder failures also stop it. The existing 45-second ceiling, logging restoration, completion matching and input admission remain unchanged. Entry samples are diagnostic evidence; they cannot authorize input or substitute for completed reads. Aggregates expose observed gate states but cannot reconstruct every transition's timing.

For the next separately authorized controller preparation, serialize sample into its pre-instantiation Worker prefix, alongside attachDispatch, and pass decision: sample in the existing attachDispatch options. Require snapshot.decisionEnabled === true, retain snapshot.decisions, and preserve the original Worker/module binding and all completion/drop/pending/error checks. The old generated prefix must be regenerated and repinned; merely updating the Node module cannot change its already embedded function. No original or new box is authorized by these instructions.

The compiled regression runs byte-identical game dispatcher, mouse read/cursor and buffered keyboard functions with synthetic objects, thunks and caller. It verifies the full media SHA before taking these small read-only slices. On the exact failed-run WASM and the current build, flag 1/count 1 polls both devices; flag 3, flag 0 or zero count leaves keyboard polling active and mouse reads absent. Shared physical deltas remain queued. Reopening only the game gate restores actual completed reads and the authentic game's scaled cursor update. This reproduces a causal mechanism; it does not assign that mechanism to the unobserved original gate. No emulator or controller gameplay behavior was changed.

Evidence: scratch/runs/20261111-mw3-menu-mouse-poll-gap-cause-0335/ retains the raw failure, verified prior-path/source references, disassembly, exact Git basis/patches, bounded-command actual waits and separate stream EOF. The next missing prerequisite for a runtime repair is a same-identity gate/caller sample during an actual gap. No timeout extension, replay or cockpit attempt follows from the current evidence.