MW3 throttle: missing compiled observation, 2026-11-11

Run: scratch/runs/20261111-mw3-throttle-consumption-cause-0125/. Baseline: 20261110-mw3-batched-native16-firefly-0050, tested main 2c29f727af55c20159749bbef21610c5215cc2a4, module SHA-256 75ecda5701c15f6fe2c269455ead4e15457b90653733df15d35674b08e8849d5.

The retained run does not identify an emulator throttle defect. Trusted native key edges are present; actual MW3 keyboard completions for those edges are missing. The files named consumed-approach-firefly and consumed-set-forward-throttle contain claimed before execution\n: command reservation markers, not keyboard records. browser.cjs arms its poll/held-frame observer only for Space. Its successful Space fixture cannot establish that MW3 sampled either movement command.

Exact route and control meaning

The shipped demo readme specifies Up Arrow = forward and top-row 0 = 100% throttle (1 = 0%, 2 = 15%, through 0 = 100%). Source media is referenced in the run evidence, never copied. Executable: existing binaries/shareware/mw3/ex/Program_Files/mech3demo.exe, SHA-256 df457ee6973f6f9d557c57aa3fc14258303486ec8b8742fd37b4369c1f37dbdc. The following are static default bindings, not a dump of the live game's potentially changed configuration.

Boundary ArrowUp Top-row 0
Pinned native helper code ArrowUp, VK 38 code Digit0, VK 48
Production renderer/DI translation DIK c8 DIK 0b
Demo default binding initialization, 40dc0f..40dded action 12 hex action 1e hex
Game interpretation per-update incremental throttle full-throttle preset callback

The actual keyboard path is renderer physical state plus buffered edges -> di_keyboard_buffer -> compiled IDirectInputDevice_GetDeviceData in src/09a8-handlers-directx.wat. Its DIK/VK table distinguishes extended Up (c8, VK38) from numpad 8 (48, VK104), and top-row 0 from numpad 0.

The demo creates SysKeyboard at 4f729f, stores it at 5fc90c, installs the standard 256-key format, requests a 128-event buffer and acquires it. 46a9cc -> 4fa160 -> 4f7690 reaches the keyboard poll only when its game input gates permit. 4f76c5 calls vtable+28; the return address is 4f76c8, stride 16, count 128, flags 0, buffer pointer at 5fc910.

4f7610 ages game key states at 5fc924 + DIK*8; 4f7700 processes returned records. A fresh press becomes state 1, a held state ages to 2, and release ORs bit 4. 4f793a dispatches registered key callbacks when bit 1 is set. Consequently, a press and release in the same later drain can still dispatch a preset callback. A physical-up snapshot, a zero-count later poll, or a released-neutral frame does not disprove an earlier press.

Binding object 600af0 stores packed primary/secondary DIKs; 4f8120 / 4f8140 extract them. 4f84d0 registers callbacks, 4f8560 dispatches the mapped action, and 4f9510 -> 4f8590 queries action state. 41f8bb queries action 12, tests state bits 0/1, and supplies positive 0.025 throttle increment; action 13 supplies negative 0.025. 41fb8f onward applies the controller's mode, neutral and decay logic. Thus a two-second Up pulse is not equivalent to latching a full-throttle preset, nor does its final neutral image establish the value during the hold.

41e9e0 registers throttle preset actions with 41e9b0, which forwards through controller 5cc380 virtual+18 to 41f170. Top-row 0's action 1e reaches 41f490, submits signed 1.0 through virtual+20, and updates mode flags. These callbacks also have live controller/mech gates. Their execution, active binding values and controller state were not captured in 0050. No claim that neutral logic actually caused the original result follows from this static trace.

Retained original evidence

ArrowUp trusted DOWN 1794354735781, UP 1794354737783 (2002 ms). Digit0 DOWN 1794354805750, UP 1794354805850 (100 ms). The exact commands requested holdMs 2000/waitMs 3000 and holdMs 100/waitMs 5000 respectively. The latter full-frame capture began at 1794354811411, over 5.5 seconds after release. The final cockpit speed observation therefore does not bound speed during the pulse.

Full approach, preset and later observation frames retain Firefly, terrain and cockpit; no positive chassis speed is established. Firefly range changes alone cannot attribute player translation. Setup is the default Commander Bushwacker, Operation 4A Battle Progression wave 1, no lancemates, Firefly primary. No original fire was attempted. The run's disk-floor termination does not establish a throttle root cause.

Bounded observer for a separately authorized game

tools/mw3-throttle-observer.cjs wraps only the owning Worker's host API-log imports before WebAssembly instantiation. The existing read-only decoder tools/mw3-keyboard-completion.js validates actual compiled stdcall returns. It records only keyboard GetDeviceData calls from 4f76c8, preserves the first invalid raw return, bounds records/depth/duration, restores the known prior API-log setting, and refuses incomplete or replaced observations. It sends no input and writes no guest memory. API logging can add overhead; this is a diagnostic, not a timing-performance comparison.

Install in the verified owner/import construction context (bundle the two CommonJS helpers for browser use if needed):

const observer = attach(hostImports, {
  getExports: () => owningInstance.exports, memory, readCompletion,
  objectsBase: BASE.DX_OBJECTS, objectsSize: SIZE.DX_OBJECTS,
}); // before constructing owningInstance
// Verify exact module SHA, Worker/epoch and known API-log setting first.
observer.start({durationMs: 20000, apiLogBefore: 0});
// Existing authorized pulse runs unchanged with independently prearmed UP.
// Capture passive full frames during Up, then after UP and after Digit0.
const snapshot = observer.stop(); // save even if correlate refuses
const result = correlate(snapshot, {
  code: 'ArrowUp', downAt, upAt, downSequence, upSequence,
});
observer.restore();

Use trusted DOM edge timestamps on the same epoch clock, not CDP ACK times. For each trusted edge, also record the live page renderer's _diKeyboardSequence immediately after its normal input handler runs, bound to that renderer/epoch. Supply those exact DOWN/UP sequences to correlate; timestamps alone cannot distinguish an older queued pulse. Do not substitute queue counts or fabricate sequence values. These extra edge observations are required for correlation, not evidence that a guest callback ran. One bounded window can cover both already-authorized pulses; correlate Digit0 separately using its own interval. Installation after instantiation cannot replace captured imports. Do not replay an accepted/unknown command. Keep exact owner/module identity, raw records, full-frame timestamps and independent UP evidence. A missing return can mean a game input gate or polling delay, not automatically emulator loss. A successful matched consuming return proves only that the game received records. consumedDuringHold reports whether the press return preceded native UP; peeks do not qualify. If consuming returns exist but speed remains zero in held frames, the next missing boundary is the live binding/controller callback and mech state, rather than host delivery. Do not rewrite guest controls or extend the hold to hide it.

Local validation

node tools/mw3-throttle-observer.test.cjs instantiates the full built module and invokes the real compiled x86 COM dispatcher, acquisition and keyboard handler against the production renderer/host. Both VK38/DIKc8 and VK48/DIK0b yield exact in-hold consumed press/release pairs. A second Digit0 pulse drained after UP produces both edges but is correctly classified as late; repeated peeks cannot qualify. Missing/stale native sequences, incomplete traces, cross-device pairs and reuse are refused, and stopped captured wrappers remain inert.

This is a synthetic guest fixture, not a game run or chassis-speed test. Initial unexpected audio-epoch import failures are retained separately from the corrected successful fixture. The task creates no box and runs no original game. No emulator or guest change is warranted by the evidence.