MW3 original buffered input and visible gameplay — 2026-11-07

One fresh original run reached the reviewed menu, default setup, operation map and cockpit. Ordinary Digit5, Left Arrow and Space each produced a consumed press/release pair in MW3's normal GetDeviceData caller 0x4f76c8. Reviewed screenshots show displacement, a heading change, and laser cooldown/heat effects. A beam, impact, damage, kill, audio or sustained playability is not established. No emulator defect or production-source repair is claimed.

Evidence is in scratch/runs/20261107-mw3-upload-gameplay/. Preparation used main 1c42d2302333e901c9528e03e4d1ff4800159dd0, including published allocator commit ef2709a76ff57f7b4cccf5e60286f4fd9b01a7c7. The canonical executing module was 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121.

Staging repair and retained controls

The run-specific provision.py records every submitted path with its pinned size/hash and uploaded, failed or skipped outcome. The first failure inhibits queued work; all futures, including successful in-flight peers, are drained and recorded before returning failure. Exception records retain type and HTTP status without signed URLs or credentials. Upload failures cannot reach boot.

The focused regression simulates PUT 502 with concurrent successful peers: 20 submitted, 20 drained, 3 uploaded, 1 failed, 16 skipped. It also checks the all-success case, signed-URL exclusion, executable remote hash-verification code and rejection of a changed staged file. Final results are in upload-regression-final.json.

Actual staging uploaded 314/314 paths without retries or failures, then verified every remote size and SHA-256 before boot. This comprises 312 pinned runtime/helper inputs plus the manifest and boot helper. The first read-only verification command failed with a confirmed Python indentation error, before executing any verification or runtime code. The corrected read-only command passed; both command results are preserved. The reusable generator now compiles its generated verification program before submission, with a local execution regression. The failed inventory was not an input UNKNOWN.

The predecessor's passed input, owner/module/epoch, 120-second freshness, native-motion, binding, independent-UP and sticky-UNKNOWN controls were retained instead of repeating broad qualification. retained-controls.json hash-references those results. The static executable Node closure check passed for 66 files, and all 312 pins passed. No private renderer handler, guest-memory write, altered binding/profile, clock change or synthetic guest replaced the original.

The locked build passed with WINE_WAT_NAMES=1. Named and canonical code bodies were compared byte-for-byte. The installed GetDeviceData body and all proof fields except the complete module hash match the qualified predecessor: function 5962, body [888196,889324), entry 888199, keyboard completion 888442. The observer still recovers executing CDP module bytes, verifies the complete hash/body and exact executable locations, pairs real requests/completions and distinguishes normal caller from startup flush/drain. GetDeviceState was not used as the keyboard observer.

The explicit allocator adoption is retained as incorporated-gl-allocator.patch. Relative to the predecessor's pinned runtime, only the canonical module hash changed. Published fece02eca frontend completeness behavior remains included. MW3 used the legacy D3D WebGL backend; this run does not attribute its gameplay result to the GL CPU-storage allocator repair.

Reviewed original route and effects

The browser was Chrome 151.0.7922.108 / V8 15.1.206.16 on the separate box, with production guest/render Workers, cross-origin isolation, native viewport, 640×480 guest canvas and WebGL2 on Intel UHD 620 / Mesa 23.2.1. The owning legacy render endpoint reported ready, error 0 and zero software draw fallback. The actual browser-recovered canonical module is retained under retrieved/original/module-f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121.wasm.

The reviewed main menu led through ordinary pointer-lock acquisition, bounded trusted native mouse motion, pilot a via ordinary a/Enter, and unchanged Bushwacker / Operation 1A / Battle Progression / Wave 1 defaults. Software cursor and tracked physical coordinates were kept separate. At Launch the software cursor was (575,432), physical point (554,584); the highlighted lower-right deployment control matched the established Launch control. A readable tooltip is not claimed. The first post-click capture still showed the map; a passive wait reached the cockpit without another click.

Each screenshot below was retrieved with a size/hash check and personally reviewed. Full PNGs and capture receipts are in reviewed/.

Capture Visible observation
cockpit.png Cockpit framing, terrain, HUD, OP Able 449 m, clock 00:00:01; ER Laser selected; ammunition 100/100/24/24/20.
armed.png OP Able 449 m, clock 00:00:02; stationary baseline before Digit5.
after-throttle.png Clock 00:00:03, OP Able still 449 m; this immediate capture alone did not prove movement.
after-left.png Clock 00:00:05, OP Able 441 m; terrain alignment and compass strip begin changing.
after-fire.png Clock 00:00:06, OP Able 428 m; heading strip shifted right, consistent with left turn; ER Laser indicator red and heat bar raised.
unobserved.png After observer removal and passive wait, clock 00:00:07, OP Able 410 m; changed terrain alignment, retained changed heading, raised heat/red laser indicator.

The 449→441→428→410 m objective-distance progression and changed terrain support real displacement. The shifted compass strip supports a left turn; no numerical turn angle is asserted. The Space consumption followed by laser indicator/heat changes supports firing behavior. A transient beam/impact was not captured. Unchanged ballistic/missile ammunition is expected for the selected energy weapon and is not ammunition-consumption evidence. This instrumented run is not an FPS or uninstrumented timing measurement.

Three command review strings contain stale/transcribed clock or distance values: cockpit was 449 m/00:00:01, armed 449 m/00:00:02, and after-left 441 m/00:00:05. Those original receipts are preserved. The PNG pixels and this final review correct them; the earlier strings are not independent measurement evidence.

Actual original consumption

The observer retained 39 paired successful normal completions, no drops, pairing errors or diagnostic errors. Passive acquired standard-keyboard polls preceded input: type 7, device type 1, format 256, stride 16, request 128, flags 0 and HRESULT 0. original-consumption-verified.json validates the normal records and these ordered consuming pairs:

Ordinary input Hold DIK Data Sequences
Digit5, VK 53 150 ms 6 128, 0 5, 6
Left Arrow, VK 37 2,000 ms 203 128, 0 7, 8
Space, VK 32 2,000 ms 57 128, 0 9, 10

The first two pairs arrived together within their respective successful normal completion; Space's two edges arrived in separate successful normal completions. Every DOWN and independently prearmed UP received its native ACK. There was no input UNKNOWN, replay, alternate throttle or repeat pulse. Both breakpoints and diagnostic globals were restored before the final passive capture. The closure and original records verify the current module rather than borrowing the predecessor's fixture-consumption result.

Custody and shutdown

Normal shutdown was explicitly delegated to a separate OS process. The independent 1,500-second watchdog was armed and alive before allocation; deadline 21:56:42.260Z. The sole UUID-idempotent box was bx_t39baaa2, created 21:31:42Z, initial expiry 22:31:42.110Z. No second box was allocated. /home had 9.3 GB free at preparation; allocation required >5 GB.

After ordinary quit, Chromium exited code 0, signal null. Both runtime PIDs were absent, both owned ports closed, profile and writer lock absent, and native events left no held key/button. All 312 remote pins were unchanged. 234 remote artifacts / 9,975,293 bytes were retrieved and size/SHA-256 verified before requesting stop, reusing already retrieved exact screenshot copies rather than downloading duplicates. These are unique run outputs and the executing module, not source/media archives.

Normal stop stop_e22ba4964d6880a6 completed 21:38:42.898Z, forced:false. Final state is archived, archiveAfter:null; retained-data expiry is unknown. The watchdog was verified alive immediately before and after normal stop; both lifecycle processes then exited normally. No runtime/lifecycle ownership remains with this worker.

Recovery uses the exact Git SHA, relevant tracked/new-source binary patches, owned helper patch, manifest, executing module and 36 existing media/system paths with sizes/hashes. Media were referenced in place, not copied into run evidence. No source-tree archive or public deployment was made. Further work, if separately requested, can investigate rendering quality or capture a beam and impact; this bounded attempt needs no gameplay-input replay.

Final verification passed the focused uploader/stage-hash regression, static Node closure, original consuming-pair/native-UP/restoration checks, all remote retrieval hashes and documentation whitespace. The evidence checker reports zero containment, missing-file or schema errors for this run's 52 references; unrelated global Alice and Tunnel/Twist errors remain in the saved output.