MW3 ordinary Digit5 qualification and staging boundary — 2026-11-07

The retained driver now accepts ordinary top-row 5 as Digit5, virtual key 53. A read-only observer demonstrated actual successful KEYBOARD state returns, including pressed and released Digit5, in the generated production guest/render Worker fixture. Original cockpit movement remains UNKNOWN. The single authorized fresh box stopped at a source-pin mismatch during provisioning, before the original browser or any original input started.

Evidence is retained in scratch/runs/20261107-mw3-keyboard-throttle/. Only this note is committed; the task-specific driver, observer, recovery scripts, patches, module and evidence remain in the ignored run directory. There is no production change or emulator repair and no locked build applies.

Documented ordinary controls

The installed local MW3 media has no README. Recovery located the pre-release demo at Internet Archive. The 36,738,513-byte MW3Demo.exe has SHA-256 97feaf6f835a69a73d2182ee58294579b0dcc815cc50e018c1b0e6bf5ddc9828. Its selected Program Files/readme.doc, 647,168 bytes, identifies itself as MechWarrior 3 Demo, 18 March 1999. catdoc extraction documents top-row 5 for 50% throttle, Left Arrow for turning left, and Spacebar for firing. demo-binding-proof.json retains document/text hashes and exact bindings. The installer and intermediate cabinet were temporary and removed. Only the selected documentation and small range probes remain, without game archives. The installer lists a 1,844,224-byte mech3demo.exe, matching the installed file's length; executable byte equality with that installer was not tested.

An earlier small range recovery from Sarna's download labeled “demo” yielded a README identifying itself as a Version 1.2 updater. It was explicitly rejected as demo-binding evidence and retained as sarna-README-not-demo.TXT with readme-recovery.json. The predecessor's binding assertion and a general controls guide were not substituted for the demo documentation.

Exact local driver and keyboard observer closure

The predecessor's mouse-specific buffer probe was replaced in this run only. At the verified GetDeviceState completion, the observer reads the returned request from the guest stack: return address at ESP−16, device at ESP−12, size at ESP−8 and output buffer at ESP−4. It reads the wrapper's slot and the compiler-generated DX_OBJECTS placement, then distinguishes type 7 / device kind 1 (keyboard) from mouse kind 2. It captures the requested 256-byte buffer and format size rather than reading the predecessor's fixed mouse address. Its breakpoint condition returns false and writes diagnostic JavaScript buffers only. The module/body/location proof remains exact: module e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467, 470-byte handler body, completion 887941. Breakpoints and diagnostic globals were removed after observation.

The generated page issues x86 public-COM calls through published vtables to create, format, acquire and poll a keyboard using the production module and GuestWorker, with the production render Worker attached. Trusted Chromium Digit5 events feed the generated fixture's host key bitmap through the production RPC publisher. Its zero-capacity buffered-keyboard adapter returns zero; this fixture tests immediate state, not buffered keyboard transitions. It does not establish MW3's own input path or gameplay semantics. Its first setup failed on an absent generated window-info host response; that failed browser was closed normally and retained in local-setup-failure/. The final fixture uses the supported default background cooperative level.

The exact task driver and observer then recorded 652 HRESULT-zero keyboard returns, no drops or diagnostic errors. The device was 149037112, buffer 16968580, caller 16968896, format/request size 256. DIK slot 6 changed from zero to 0x80 during the 1,500 ms pulse and returned to zero after UP. keyboard-consumption-summary.json retains the exact correlation. UP was independently armed before DOWN; both actual native events were trusted. Exact transport ACKs 0–3 covered arm, pulse, disarm and quit.

Focused controls rejected wrong Worker instance/endpoint, modified module bytes and wrong completion location before breakpoint installation. Keyboard validation rejected wrong object/device kind, request size, format size, HRESULT, device pointer, buffer and caller, and refused mouse-only evidence. All 18 ownership/module/epoch fields rejected replacement; the retained production freshness gate rejected a 120,001 ms expired capture with zero input effects. A stalled DOWN ACK independently attempted UP and produced sticky UNKNOWN. Closed-child transport refused further commands. The published native-motion focused regression passed. No accepted input matrix was repeated; native-motion and capture guards were unchanged.

The proved blank acquisition anchor (350,430) and full route envelope were retained and passed a conservative simulation with the unchanged native helper. That simulation is physical headroom evidence, not original navigation.

First original boundary: qualified source changed before upload

One persisted UUID Idempotency-Key request allocated bx_sdf6q4ce, original expiry 17:03:27.951Z. The qualified runtime pinned lib/host-audio.js at 210,449 bytes, SHA-256 70795825d7d11ecb026e22c2875ef4eaf4fba60758e5da4fcc99ab624c4f4513. Before upload, that shared file changed to 210,748 bytes, SHA-256 bb89e4e95190b7db98390e5d67f1c99fdcdf7e7e77f0388aa0126cd256869365. Its other owner had announced an active MIDI repair. The retained stager's SHA/length assertion refused it. No source file belonging to that owner was edited, no pin was relaxed, and the original was not booted.

first-blocker.json proves the qualified file is exactly recoverable from Git 10bf01e429f7c0347de3dae4ad11002bf248dc7c. The changed hunks are preserved in blocker-host-audio.patch. Remote inventory confirmed 307 matching runtime pins, the sole missing pin lib/host-audio.js, no mismatched uploaded pins, no original directory, no runtime process, and no task-owned Node/Chrome process. After the provisioning failure, the shell staged the small boot script; it did not execute it. No cockpit, throttle, turn or fire attempt occurred. This is a staging prerequisite, not an MW3 or audio defect claim.

An initial read-only custody assertion mistakenly matched its own Python command text as a Node process. Its exact failed result is retained. A new read-only query classified executable argv[0], excluded itself and established the custody facts above; it issued no guest or input command.

Recovery and lifecycle

Git is the source archive: source SHA, relevant binary patches, new-helper patches, actual loaded local module and existing media path/length/SHA records are retained. This does not claim a rebuild at the source SHA. Unique remote custody evidence was retrieved and SHA/length verified before normal stop; uploaded source and media were not copied back into an archive.

Two explicitly delegated OS subprocesses were armed before allocation: a normal-stop delegate and an independent lifecycle watchdog. The watchdog's immutable 1,500-second deadline was 16:28:28.099Z, derived from the persisted allocation receipt 16:03:28.099963Z. The API later reports creation at 16:03:27.930Z; the cap is receipt-relative, not a claim of exact creation-time alignment. Both processes recorded half-second heartbeats. The watchdog was verified alive immediately before stop and after confirmed normal stop, then exited normally. No lifecycle-process failure was observed.

Delegated stop stop_62f9bd189a298057 completed at 16:05:13.221Z, forced:false; the box is archived, current archiveAfter:null, retained data expiry unknown. No lifecycle ownership remains. Local driver and Chrome exited normally with ports 43851/35701 closed, profile/writer lock absent and Digit5 released. The failed setup browser also exited normally. Temporary recovery files, lifecycle processes and messageboard watcher were closed. Sampled free space stayed above 2 GB; a continuous disk minimum is not claimed.

The next separately scoped attempt can recover the exact qualified source from the recorded Git SHA or requalify a stable composed runtime after the audio owner's release, then obtain separate authorization for a fresh original. No second allocation or original retry is part of this task.