Re-Volt current browser driving admission, 2026-11-18

Run: scratch/runs/20261118-revolt-current-browser-driving-checkpoint/. Outcome: harness error; manual driving, checkpoint/lap and ordinary Quit unverified. No production repair or original replay. No runtime defect was established. The old legacy-software capture timeout remains a closed scope.

The immutable published source pin is 0d6f51781a6d9daa28f265c07a1c5bd44228c0d3; compiled/served module SHA-256 is 1504acf2ea5cac3ec3b8c627f4edc64c564e62ad37c06928f288bd8ae83b4f37. The original's executing module digest is unknown: its live Worker module recovery did not occur. Do not substitute the served-module hash for that missing observation. The unchanged registered re_volt_demo, its real Mss32.dll and manifest media were used; 463 media/search-candidate pins were checked. The original EXE remains 58ba4835e6607eae62be4e023ef989172cb519c78f34811acba547e9e0e3581a. Foreign shared-tree overlays were excluded from the served Git graph and preserved as foreign work. No guest/config/registry/media edits were made.

Before the original, an authored native Device7 fixture exercised actual WAT D3DIM GPS draw/fence/readback into the published software executor, checking a textured RGB565 red pixel. A separate normal production-browser GL-GPS fixture showed a yellow triangle on blue, then red after ordinary keyboard input. It verified actual guest/render Worker module bytes, native HWND generation 1, thread 1, focus 65537, no capture/modal, independent exact paired-arrow releases and ordinary Escape guest exit 0. All five browser fixture scopes have actual child/enclosing waits and both pipe EOFs. The native fixture's tool completion is retained separately; its two pipe EOFs were not independently recorded. These controls do not qualify every D3DIM publication route.

Failed authored scopes were retained and closed. Calibration2/3 pinned the output canvas before its ordinary startup transition. The harness was changed before the original to wait for the presenter's window to appear in an actual completed compositor record. Calibration4 then exposed a controller ordering error: its release clock began before a final awaited identity check. UP fired at 1794992809476 ms; ArrowRight DOWN issued at 1794992809479 ms, leaving physical ArrowRight DOWN. The correction finishes identity work before starting the hold/release clock. Fresh calibration5 passed paired release and ordinary exit. None of these is an engine repair.

The fresh original started at 09:09:35 UTC. Limits were declared beforehand: 900 seconds startup, 1800 seconds total, final 180 seconds reserved for Quit, 20 seconds cleanup; per-request and whole root-capture bounds were 3000 ms. Identity work was separate from finite gesture clocks. No game input was sent. The saved whole private-display observations at 09:11:26 and 09:14:03 show different natural Demo Mode street/camera views. They do not prove manual driving, speed, steering, a checkpoint or lap.

A successful read-only page diagnostic at 09:12:36 reports one live d3dim-gps endpoint, ID 1, epoch 1, WebGL1 device and actual ANGLE/Mesa renderer string, with no bridge error/closing state. Its page presenter has zero presents and no owning window, while the completed ordinary display record contains the live native HWND 65537. The device's cached initial stats report only LAYOUT records; they are not proof of zero later renderer draws. CPU readback/publication is a candidate explanation, not an established cause. No retired lib/d3dim-gpu.js was served.

The exact admission blocker is the harness's unconditional requirement for a GPS-presenter-owned frame. That requirement remained false even while ordinary native display publications and changing Demo Mode images were observable. It was qualified with GL-GPS, not a D3DIM native-display/readback case. No guard was relaxed or bypassed to send input to this original.

The subsequent read-only module diagnostic failed at 09:16:08.300 UTC. Target.getTargets returned three Worker targets with empty URLs and a parent ID matching the exact page. The diagnostic's origin/URL filter excluded all three, then asserted that a main Worker had been inspected. No Worker was attached and no executing module was recovered. This is a discovery failure, not proof that the Workers were absent or stopped.

There is also a closure-policy failure: the separate startup controller did not share this diagnostic's sticky UNKNOWN gate. It issued 54 subsequent passive Runtime.evaluate requests, IDs 1462–1515, until host closure. The first was 09:16:08.466 and last 09:16:23.003 UTC. Zero original Input requests occurred. The controller was stopped through the OS, the owned browser terminated, then the controller resumed to observe connection rejection and run cleanup. No later endpoint or replay was used after that host closure. This run therefore does not pass the requested no-request-after-UNKNOWN property.

The next prerequisite is an authored normal-browser D3DIM fixture that covers native/CPU-backed completed display publication, exact-page child Worker discovery including empty URL metadata, executing module recovery, and one shared permanent closure gate across startup and diagnostics. Required backend, owner, current pixels and freshness checks must remain active. No further original is authorized by this report alone.

Chrome exited 0 through host cleanup; this is not ordinary game Quit. All six browser scopes' browser/display/supervisor/watchdog/enclosing waits have both pipe EOFs and supervisor no-leak receipts. Media hashes, source diff, selected whole PNGs, request/body/reply boundaries, recovery recipes and the closure violation are indexed in the run; result.json is written last. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt null. Held titles and foreign UT/AoW2/render work were preserved; no delegated agents, Codex workers, Python, Wine or deployment were used.