Re-Volt current paired driving, 2026-11-12

Evidence-only run: scratch/runs/20261112-revolt-current-paired-driving-checkpoint-0911/. This extends the current paired active-control scene coverage by one title. It does not prove a checkpoint increment, completed lap, finish, save, full 27-title completion, backend parity, audio, hardware rendering, or qualified FPS. Historical November 9 road-checkpoint and November 11 native-recovery results remain scoped as originally recorded.

The runtime source pin is 8a834a3dc7affc8c6e53245beab56e87a32b81e8 (the requested documentation successors 281713c5f / 4737c2e50 do not change its runtime). Both actual guest worker module captures hash to 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17. The registered local-only re_volt_demo installation was used through test/binaries/candidates/re-volt-demo/.wine-assembly-browser.json, SHA-256 c87bcdc1af685f052fb72ad6ad3492739670d45aecbc2672a337f36b49bd5ddc; original installed/revolt.exe is 58ba4835e6607eae62be4e023ef989172cb519c78f34811acba547e9e0e3581a. All source/media closure hashes are in runtime-pins.json. No new download, repack, public deployment, guest-state write, route/config/physics modification, or save was performed. WIP was excluded by pins; relevant local source differences are retained as a patch.

Provisioning actually joined successfully with parent wait and EOF. Before either browser started, a missing transitive tools/gltron-logical-fps.cjs dependency was found and supplied; the corrected complete require closure and 1,030 final hash checks passed, bound to this box and manifest. Initial manifests and the supplemental closure process are retained. This is separate from the earlier NFS3 HTTP-400/pre-ACK suspicion, which this run does not resolve.

Each browser installed recursive early worker observation before resuming targets. Full scene captures bind the actual guest, module object, process, endpoint, canvas and owning renderer worker. Software's owning endpoint reports backend software, no WebGL adapter, and completed software raster work. WebGL's owning endpoint reports webgl, a present adapter, zero software adapter draw work, and actual WebGL2 contexts. Chrome's native SystemInfo reports ANGLE Vulkan SwiftShader/SwANGLE on both sessions. The JavaScript context's Intel UHD 620 string is spoofed by the browser environment and is not hardware evidence. No FPS is inferred from draws, flips, copies, or game-clock values.

Session Native activity Bound scenes / balanced pulses Limit
Software Single Race, RC Bandit, Toys in the Hood 1; visible acceleration to 21 mph and later 26 mph, independent steering, drain entry, native wrong-way X and Home recovery 23 / 18 Street/fence routing and wrong-way drain entry; lap remains 1/3, last lap zero
WebGL Time Trial, same car/track; 16 mph from start, separate right/left corrections, cleared alley fence, drove south along first street past authored route-node neighborhoods 2/3 21 / 19 Curb-constrained heading; no checkpoint/lap proof before terminal helper failure

Every dependent input followed individual full-frame review. Key release was armed before DOWN, and all 37 recorded dispatched pulses have acknowledged DOWN and UP. Software Home twice produced a black game area while the app/endpoint remained live; separate passive captures restored the rendered race, consistent with the November 11 recovery scope. These observations are retained and do not establish a persistent emulator crash.

The changed approach used opponent-free Time Trial on WebGL and read-only original .fan / .taz route geometry to guide ordinary keys around the early alley fence. Read-only native positions corroborated movement. The inherited car+0x340 matrix was found to be a local model transform, not the changing physics-body heading. Body pointer car+0x394, body matrix +0x64 and position +0x14 supplied a useful observation; this informed normal steering only. By to-bend, native car position was approximately (-9368,104,16642), body forward approximately (-0.003,-0.066,-0.998), and the full frame showed an upright car along the curb at 10 mph. Native position, authored route proximity, zone classification, wrong-way status and recovery are not checkpoint proof. Read-only disassembly identified native zone classification at 405d90 writing car+0x2808/+0x280c/+0x2810/+0x2814; it did not establish the ordered race acceptance transition. No such transition is claimed.

The precise terminal blocker was a caller error at 09:34:18Z: curb-clear used action:"keys" with Up/Right. This helper accepts combined keys under action:"key". Its action assertion failed before dispatch, and browser/control custody became terminal. The request was not replayed, no later gameplay inspection/input was issued, and the browser closed. This is a run-helper failure, not evidence of an emulator defect. A future fresh run should validate the action schema locally before launch and use the already-supported action:"key", keys:[...] form, then continue the first street/bend with short individually reviewed steering. Real checkpoint acceptance still needs an identified native progression field or a native lap/last-lap change; repeating positions/recovery would not close it.

Local zero-effect stale regression passed: pre-dispatch STALE_SCENE_BEFORE_DISPATCH gets an exact zero-effect refusal, explicit fresh capture then a distinct input is admitted, and a post-effect failure remains terminal. Five existing focused input/renderer tests passed; see focused-tests.json and logs. No production defect was demonstrated, so no production fix, build or commit was made.

Recovery used 414 unique gzip blobs for 519 remote references, compressed and hashed on the box before transfer. Every compressed SHA-256 and decompressed size/SHA-256 was rechecked locally. The canonical module is deduplicated across both sessions. compressed-evidence-manifest.json maps original names to blobs; check-evidence.py reproduces the backend/module, balanced-pulse, scene-hash and stop checks. Initial and failed attempts remain in the run. A lifecycle-only finalizer first looked for the obsolete closed.json filename and failed before writes; the corrected finalizer reads this helper's actual cleanup.json. Both attempts are retained. No source or media archive was created.

Box bx_6dncfgk4 was created at 09:12:01Z with initial expiry 10:06:38Z, after an observable independent immutable 55-minute watchdog was armed. Evidence was verified before normal delegated stop at 09:36:26.969Z. Final state is archived, expiry NULL; receipt stop_36f2cdd1e551c4fc is completed, error NULL, forced false. Remote software owner actually waited with code 0 and both EOFs; WebGL owner actually waited with code 1 and both EOFs after the helper assertion, while Chrome exited 0. Both browser and Chrome PIDs are absent and profiles removed. The local long-running launcher separately exited 143; its cause is unknown and is not conflated with remote runtime waits. Its manager was absent; the independent watchdog observed the stop, wrote its terminal confirmation and is absent. An actual parent wait for that reparented local watchdog is unavailable and is not claimed.

The coordinator independently verified all519 references/414 compressed and raw payload hashes, actual scene/module/backend bindings, native SwiftShader reporting and full software/WebGL driving images before publishing this note. No code build was needed for this documentation-only publication. The existing Re-Volt overview and historical notes retain their original scope.