Motocross Madness: steering-consumption preparation boundary, 2026-11-07

Steering remains unproved. No new box or original was launched. Preparation demonstrated a concrete admission defect in the retained run helper: a fresh review can authorize input after process, endpoint, Worker, module or epoch replacement. The task requires those identities to stay bound to the reviewed scene. The existing helper cannot safely carry the requested original forward. This is a harness boundary, not a demonstrated game or emulator defect.

Evidence: scratch/runs/20261107-mcm-steering-consumption/. Historical execution is separately recorded in mcm-refresh-driving-20261107.md, mcm-race-driving-20261107.md and mcm-turn-at-speed-20261107.md.

What the retained steering experiment measured

Personally inspected historical throttle-fresh, left-immediate, right-clear-three and forward-baseline captures. The rear RIDER remains near the center and upright. Quarry ground and horizon shift, but the speedometer needle is near the low end without an established numerical speed. The separately labeled green NEWTON rider crosses the view. Those opponents, the red direction marker and capture-to-capture terrain changes cannot establish player steering. These historical images are not a fresh baseline for another input.

The exact ordinary-keys.cjs sends native CDP keyDown for Up and then Left or Right, sleeps for an integer 50–2000 host milliseconds, and releases every attempted key in reverse order in finally. It neither waits for a guest poll nor samples guest progress while the keys overlap. The numbered success reply is written after the input helper, post-action wait and capture complete. It therefore establishes completed harness execution, not guest steering consumption. Prior generated trusted-event checks establish native delivery and balanced release; they do not close this measurement gap.

The recovered console contains exactly nine accepted driving groups, matched in order to commands and exact-ID replies. consumption-timeline.json preserves their raw dequeue timestamps and post-action counters. Four combined groups drain down/down/up/up within 1–2 ms despite requested 2000 ms holds:

Command Message dequeue span Post-action slice sequence Flips
throttle-right 2 ms 7267 31
left-pulse-two 2 ms 7308 39
left-pulse-three 1 ms 7381 51
right-clear-two 1 ms 7505 68

This newly extracted timing explains why the ACKs are inadequate. It does not prove zero DirectInput polls. host.js emits [input] after renderer.takeInput dequeues a Win32 message. DirectInput's GetDeviceState reads host_get_key_down_state; Worker RPC can use the published physical bitmap when no event is open, while an open event can use its key-state snapshot through lib/host-window.js. These are distinct observations. Console receipt timestamps also do not provide guest instruction timestamps. No emulator semantics mismatch has been demonstrated, and none was repaired.

The post-action counters establish eventual execution and rendering progress: the nine snapshots advance from slice 7174 / 18 flips / owner completion 181 to slice 7532 / 72 flips / completion 516. They do not locate that progress inside the physical hold or prove an at-speed simulation step. A local counterexample executes the exact hold helper with production renderer input handlers, observes physical overlap and balanced release, and receives helper success with zero guest frames and zero guest polls. No guest runs in this control; it demonstrates the helper's measurement contract, not original-game behavior.

Ordinary bindings and available diagnostics

The registered media's HELP/HELP.HTM keyboard table identifies Up = Accelerate, Down = Brake, Left = Steer Left and Right = Steer Right. Its existing path and SHA-256 are recorded in retained-fixture-identity.json; media were not modified or copied into this run. No guest physics, configuration or callbacks changed.

Production diagnostics already expose page _workerLastSlice sequence, blocks, EIP and RPC counters; guest D3D flip/submission counters; and owning render endpoint completion, command and error counters. A useful bounded observation must sample these while held, maintain the same owner/module/epoch, and distinguish guest progress from renderer backlog completion. Render fences are not simulation steps and slice completion alone is not keyboard polling.

The actual retained module exports get_tick_count, get_count, get_logical_frame_addr and get_logical_frame_count, confirmed by parsing its export table without instantiating or running it. Worker readExports can read live exports between commands. Logical-frame counts require a validated guest step address; their presence does not establish a configured MCM step marker. Address hit counters likewise need a validated poll/step location. None was established here. No logical-frame marker, clock, pacing or guest state was set.

Newly demonstrated admission boundary

boundary-controls.cjs invokes the exact retained scene-guard.cjs with a fresh reviewed identity and a live app differing in one field. All five cases dispatch: changed processId, endpoint, Worker target, module SHA-256 and epoch. The helper ignores every supplied identity field. The production browser caller is weaker still: its admission read projects only app name, running/stopped and guestWorker presence, so it cannot supply the missing comparison. The capture metadata holds label/hash/time without a bound owner or epoch.

This is a reproducible pre-original blocker under this task's identity rule. The old helper remains unchanged; these are failure-detecting preparation controls, not a repaired or qualified browser helper. No generated browser repair or new original is claimed. The next implementation must bind capture admission to the actual process, Worker target/endpoint, module and scene epoch, reject identity changes with zero effects, and cover the change in a generated browser before using any original. It must also record progress inside the bounded hold; if that progress or the reviewed at-speed baseline is absent, stop that test without further gameplay inputs or a steering claim.

The same local controls confirm stale input rejects with zero effects, and UNKNOWN, late ACK after closure, partial failure and stale-dependent input create no next command. Writer locks release. These checks preserve the existing closure boundary; they do not weaken guards or authorize replay.

Orphan terminal state, identity and completion

Read-only API inspection during this task independently confirms bx_tevwnwhg is archived, normal stop completed, forced:false, stopped 07:59:40.963Z, active expiry null. No resume, remote exec, stop, delete or new allocation occurred. The recovered terminal records show all six PIDs absent, six ports refusing, three profiles and writer locks absent, keys released, and all original commands 0–28 paired with exact-ID replies. The recovery's terminal artifact verification and ownership release were checked. Its failed first parent-exit assertion remains preserved in the recovery run, followed by the successful read-only terminal census. No orphan input was replayed.

Historical actual identity is guest Worker, owning endpoint 1 legacy WebGL, ANGLE Vulkan SwiftShader Device (Subzero), module e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. Original source commit was 958d2d11f0d91e3df22efc0befc929deb0921a0c with retained dirty patches and pins, not fresh-build provenance. The earlier race-driving Intel provider must not be substituted. This task executed no original renderer.

Validation: exact-helper counterexamples, production renderer overlap/release, sticky closure and stale controls, nine-group queue/console matching, retained artifact SHA-256 checks, module export inspection, personally inspected historical images, read-only lifecycle confirmation and documentation diffcheck. No emulator code or shared helper changed, so no locked build was required or claimed. Local free home remained above 2 GB. No source/media archive, public deployment, held-risk retry, FPS, audio, parity or full27 claim. No assigned box is running; ownership returns to the coordinator with this actionable harness boundary.