Re-Volt native recovery, 8 November 2026

Home/End coverage is qualified, and native Home recovery is visibly demonstrated. Checkpoint/lap progress and an actual End flip remain unproved. This is a partial result, not a playability pass. Evidence is in scratch/runs/20261108-revolt-native-recovery/.

Qualified helper

Before allocating the single box, the owned input.cjs gained Home/VK36 and End/VK35. The actual browser adapter also accepts a bounded Up+Left or Up+Right pair through the same prearmed independent-release pulse; arbitrary combos remain rejected. The production DOM bridge, renderer input, host RPC and full module Guest Worker were exercised in a generated local DirectInput fixture. Trusted Home and End DOWN/UP produced actual buffered device records [(0xC7,128),(0xC7,0)] and [(0xCF,128),(0xCF,0)], respectively. Physical states cleared, both independent UP receipts passed, and the steering pairs also passed balanced trusted delivery and buffered-code checks.

See local-key-qualification.json, combo-key-qualification.json and helper-control-verification.json. Released readiness/identity controls were reused after hash comparison. The pulse's deadline, independent UP and UNKNOWN closure logic was unchanged. No production emulator source was changed.

The original PC manual reference is the released 20261107-revolt-control-reference/ capture of printed pages 4–5 (PDF page 3), SHA-256 95d66366b893df94174111ee7db37dff652fc7a4dd1ede303a1433f3f89fc311. It documents Home for return to track and End for flip. The Configure Controller wizard was never entered; all original defaults were retained.

Original race evidence

The registered original followed default PLAYER, RC Bandit and Toys in the Hood 1 into a reviewed race, with the actual legacy WebGL2 render Worker and default guest Workers. Guest and renderer both used module f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. The selected source commit is c7da83223fd559cb85bf92538102c792f94c1ff9; runtime-pins.json identifies 1,068 runtime/media/helper inputs. Original media is referenced by path and hash, not copied into evidence.

Personally reviewed capture Player result
review/forward-observed.png Player/camera displaced toward curb; 1 mph after release; late held capture shows 7 mph
review/end-two.png Upright, visible car at right curb, 0 mph; no observed overturn/flip
review/road-left-fresh.png Car tilted and partly below curb edge, 5 mph
review/home-from-curb.png After one-second Home, car fully on road and upright, 0 mph: visible native recovery
review/road-right.png Player turned on road, 8 mph
review/road-forward.png Forward pulse ended against tall curb, 0 mph
review/curb-left.png Short left/throttle did not establish curb clearance
review/reverse-curb.png After eight-second reverse, 2 mph but no demonstrated curb clearance

Every listed race image still shows lap 1/3. Opponent distances and advancing timers are not checkpoint evidence. Home recovery is established by identifiable curb/road geometry before and after the balanced native pulse; transient car ghosting alone was not counted as recovery.

Delivery and diagnostic boundary

Read-only observations connect physical input, published Worker bitmap and the original executable's keyboard buffer at 0x6C5F28. Home appeared as C7 in 45 samples, End as CF in 67, and forward as C8 in 56. Device slot 3 was acquired with its 256-byte data format. Both stored and active original mappings matched the documented defaults. A read-only car-list observation found the original player callback 0x418C50 and changing adjacent steering/throttle control bytes. No guest state, clock, configuration, bindings or private handlers were edited.

Static original disassembly identifies GetDeviceState at 0x439D80, default mapping setup at 0x4193B0, and native End/Home control flag generation at 0x418CB1/0x418CC6. The input mapping omission is therefore resolved. No first emulator disagreement was established beyond that boundary, so a generic physics/input fix would be speculative. The remaining concrete route boundary is native curb clearance and alignment near the start line; checkpoint/lap completion and an overturned-car End test remain open.

One End pulse was followed by black captures. A later passive refresh restored the race; render work continued without reported errors or lost context. This does not establish a persistent renderer failure. Two stale-scene requests were refused before dispatch with zero effects; fresh captures preceded subsequent input. Their attempted retrievals found an existing prior image and failed locally without replaying input. The retained logs distinguish these from an UNKNOWN outcome.

Custody and next work

The run retains exact source SHA, relevant binary diffs, owned helper patches, module, original-media identities and manual references. No source archive or media archive was made. Terminal evidence and lifecycle results are recorded in retrieval-verified.json, native-command-verification.json, normal-stop-confirmed.json and candidate-result.json.

Terminal verification passed: 30 exact ACKs (including two zero-effect stale refusals), 21 independently released balanced pulses, 24 trusted key DOWN/UP pairs, zero held keys/buttons, zero UNKNOWN and zero replay. All 1,068 selected runtime pins remained unchanged. The browser/launcher PIDs disappeared, both ports closed, and temporary browser profile and writer lock were removed. The initial custody assertion caught the launcher still draining immediately after Chrome exit; the later read-only check passed without repeating input.

All 336 terminal evidence references (273 unique hashes, 17,311,397 referenced bytes) were retrieved and SHA-256 verified before normal stop. Box bx_86zgy92g stopped at 2026-11-08 00:30:30.420 UTC, final API state archived, expiry null. Normal stop receipt stop_2e4c09c79e7ad42d completed at 00:30:30.438 UTC with forced:false. No archive operation was requested. The observable detached 25-minute watchdog had deadline 00:32:30.327 UTC; normal stop occurred first and both supervised lifecycle children exited.

Next work should use short native reverse/steering pulses with reviewed road alignment, then establish a real checkpoint/lap transition. An actual overturned position is required to qualify End's visible flip effect. A causal emulator fix requires a demonstrated disagreement after the now-proven keyboard path.