Recoil original demo: current-main browser gameplay, 2026-11-07

Reviewed original player-controlled forward movement, left turning and pulse-gun firing passed in scratch/runs/20261107-recoil-current-gameplay/. No emulator divergence was reproduced and no production source or media was changed. Historical acceptance and limits in recoil-demo.md, including main 274902b02 browser/touch gameplay, remain intact.

The tested source was main b56e2dc34062398bc975ecddbb80c48a57b9b2a3 plus relevant-source.patch and new-lib_wininit-boot.js.patch. The first patch is empty; the latter preserves the pre-existing untracked runtime dependency. runtime-pins.json records all 742 staged files, including the run-specific browser/input helpers. Git is the source archive; no source/media archive was made. All 107 original game files and the existing browser manifest match the 108 supplied pins in 20261107-recoil-next-inputs/fixture-pins.json; none was repinned. Original executable SHA-256: 82632b1b46153014af0caf1e0c9dfdebbb0d8482ba65b2735064f10972afa076.

The actual guest Worker module was recovered through CDP, retained as original/module-e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467.wasm, and matches the staged build. The environment was headless Linux desktop Chrome 151.0.7922.108, V8 15.1.206.16, 1024×768, cross-origin isolated, with Threads enabled. The direct registered route was /index.html?debug&threads&app=recoil_demo; the dropdown was not used. The game selected Software. The toolbar's WebGL selection does not establish an owning accelerated renderer: this run has no D3D render-worker endpoint. The owning software page renderer and canvas are observed separately.

The live application was process 1000, endpoint 1, slot 0; guest Worker target 1A6985B7F3C30C9FDCDA5B13569FECA1. Its document epoch was fa1c8875-83ce-48c2-8810-92849b4cdcfa; the software renderer object token was that epoch plus :5. Capture/binding JSON records actual process, endpoint, Worker, instance and module objects before and after each screenshot and before input. passive-monitor-v3.jsonl records the same software renderer object, document and endpoint throughout the gameplay holds, with canvas hashes and physical input states. served.jsonl and http-resolver-pins.json preserve HTTP/hash/isolation evidence for the fixture and system DLL/TLB closure. The four registered bundled MFC/CRT images were loaded; built-in OLE32/SHELL32 handling is explicitly reported in the original log, not described as native DLL loading. Fonts and stdole2.tlb were provisioned.

Generated real-Chromium controls ran before the original. They exercised live Worker/module/document and render-worker owner binding, zero-effect rejection of five owner replacements, stale/slow screenshot rejection, balanced native key holds and partial-failure release, passive guest/render progress, and sticky unknown/late-ACK transport closure. Ordinary DOM relative-motion and trusted CDP mouse down/up delivery also passed. Run-specific input admission uses a 120-second review interval, retaining the screenshot's original capture time and requiring its current review/hash and matching live owners. Observed admitted scene ages were 9.258–59.361 seconds, including 20.690 seconds for forward input: a 12-second interval would have rejected valid reviewed scenes. Fresh captures were inspected; an older review was never substituted.

Ordinary Software → OK → Game → Start reached the original menu. A click acquired pointer lock; bounded DOM mouse movements supplied relative counts through the production browser input bridge. Reviewed New Game and Start were activated with paired ordinary left-button holds. Read-only original cursor feedback at 0x564258/0x56425c corrected one overshoot. No guest memory, configuration, timers, callbacks or menu state were forced. Mission loading and the opening completed naturally without skip input.

Reviewed gameplay evidence, all relative to the run:

in-hold-verification.json records physical held-state samples and distinct canvas hashes: forward 16 samples, guest sequence 33238→33557; turn 15 samples, 35727→36046; first fire 15 samples, 38581→38872; matched fire 15 samples, 49569→49876. Each hold has a stable renderer/document identity and advancing guest execution with changing rendered pixels inside the held interval. These counters do not assert guest polling frequency, simulation steps or FPS.

Validation: locked tools/build.sh passed; generated owner/input/transport controls and mouse delivery controls passed; all 742 runtime pins passed again at cleanup. An initial generated control used an older helper without the capture-completion timestamp and failed its control assertion; it was corrected before original launch. A launcher environment value was initially numeric; Python rejected it before creating any browser child. Correcting that known pre-effect error did not replay the original. There was one original browser launch, and all 20 numbered original commands have matching successful ACKs.

All 308 remote artifacts (31,954,278 bytes) were recovered and SHA-256 verified before the delegated normal stop. Fifteen recorded process IDs were absent, both HTTP/CDP ports refused connections, the browser profile and writer lock were absent, and physical arrow/button states were released. Box bx_b3dw6rzy reached archived at 08:39:33.433Z through completed nonforced stop stop_212eeb29a8bdaee5; expiry/archive-after is null after stopping. The box's inherited descriptive name was mcm-bound-input-20261107; allocation UUID and actual ID, not that label, identify this single fresh Recoil box.

No physical-phone, WebKit, audio/audibility, logical/display FPS, accelerated rendering or native-game parity claim follows. Next work can independently qualify those environments or measurements; this run requires no gameplay fix.