Return Fire: blank review contradicts retained capture bytes

Run: scratch/runs/20261109-return-fire-jeep-blank-transition-cause-2055/. Inspected checkout: 02e95626dcbc46b7cab11618eec1b904f844f7b9. Historical runtime: a5f4d260838790605c1c0d62acc046884eeaf7fd, module 3e491c0840bb7771328738a0d160b4bbaa3d7ab72074cbd38a7b41e5d7e8ec18.

The reported disappearance after Right100ms is not present in the retained PNG. This corrects the visual conclusion in 20261109-return-fire-bunker-entry-jeep-2045.md, without changing its terminal safety decision or historical artifacts. No new native execution was needed.

Discriminating evidence

The prior run's review/jeep-info.png, review/jeep-screen-mode.png, and deduplicated retrieved/original/jeep-screen-mode.png are byte-identical: 25,334 bytes, SHA-256 7b27f488f0624cbde33f68f3938a24606eef4208c01d755d88e3730143388ae3. The remote command5 inventory and both capture records already report that same hash. Thus this is not merely a local path mix-up or a changed downloaded file.

Independent PNG decoding with CRC checks yields the same RGBA digest for all three files: 636aefff98a657423518f12fb8686b4f87bdbb15ba9994ee59adb49fd3517912. The image is 945×889, has 91 colors, and contains 91,564 non-teal pixels within the game-window rectangle x156–483, y97–382. It contains the title bar, menu, jeep information and lower display. It is not a uniform teal desktop. audit.cjs and audit.json retain the reproducible numerical comparison; retained-command5.png is the sole copied capture, with no pixel transformation.

Separate model-review tool calls reproduced contradictory presentations of these same bytes: some showed the window, others only teal. The view_image returned data URL was independently compared to the original file's base64 and matched exactly. Its MIME was application/octet-stream. Forwarding those bytes as image/png once displayed the window, but later explicit PNG and original-detail calls still appeared blank. MIME alone is not a proven cause or repair. A proposed MIME adapter was tested and removed before publication; the failed local assertion and subsequent observations remain in run evidence.

This locates the concrete discrepancy after retained capture publication and decoding, at image presentation/model review. The underlying presentation mechanism is outside this repository and remains unknown. No causal emulator input, window, surface or compositor defect is established by this incident.

Input and lifecycle checks

Read the actual retained browser-run.cjs, bound-input.cjs, ordinary-key consumer and collector, alongside tracked composition and renderer input code. Command5 uses native ArrowRight, code ArrowRight, VK39. The hold receipt records exactly 100 ms; a wholly in-hold passive sample reports key39=32768, and the capture identity reports key39=0 afterward. The renderer maps Right to Set-1 scan code 0x4D with extended-key handling. There is no wrong-key evidence.

The actual capture consumer calls Page.captureScreenshot, base64-decodes the response, hashes it, writes those bytes once, then publishes binding/capture metadata. Process, page renderer, canvas, worker instance and module identities match before/after. The guest remains running without a trap. Retained state does not contain a window-table/surface-event history, so it cannot establish every transient guest transition. The exact identical frame establishes only that the captured output still contains the jeep screen. A movement-mode toggle, guest keyboard polling, later persistence and successful gameplay are unproved.

The next useful action is to address the image-review presentation discrepancy using this exact PNG/hash, or resume separately assigned gameplay with reliable review. Repeating this native sequence to obtain another blank-limit note would not discriminate a new emulator boundary.

Validation and runtime ownership

The retained-byte audit passes. Tracked composition, collector, native-key and lifecycle focused checks and the locked build were rerun; results are in this run's logs. No shared pinned provider or production source changed. Evidence contains exact Git SHA and a relevant working-tree patch; no source/media archive.

No box was created, leased, resumed, captured, stopped or otherwise controlled for this investigation, so there is no new UUID, expiry or stop receipt. Historical evidence references box bx_p7b6wwek, UUID a3a3df08-526c-40bb-8832-06d1cacfe525, already archived; initial expiry 2026-11-09T21:08:14.092Z, final archiveAfter:null, separate expiresAt not returned. Historical receipt stop_19ec2c8bedbc4a5d completed unforced at 2026-11-09T20:47:00.318Z. That run retains actual watchdog wait0 and separate stdout/stderr EOF. These are historical receipts, not newly performed actions.