TR2 staged native jump and reviewed turn — 2026-11-08

The fresh original exercised the repaired standalone Alt executor and produced a new personally reviewed native left-turn consequence. final-turned.png shows Lara side-on where jump-current.png showed her facing away, with the lattice displaced right in the changed alley view. This establishes a visible turn, not crossing the slope or reaching a new opening/plaza. Alt delivered trusted native down/up and ACKed, but neither its held image nor the settled images establish an airborne pose or successful jump/landing. Those goals remain open.

Evidence is scratch/runs/20261108-tr2-jump-staged/ (R). No production code changed. The accepted native module and predecessor pixel authority remain unchanged.

Fresh physical closure and startup barrier

Preparation verified the corrected predecessor successor-runtime-pins.json and normalized its 29 physical helpers into the fresh directory using tools/runtime_helper_closure.py. Before allocation, the exact provider/read/upload binding included launch-locked.py, both repaired key helpers and their actual tools/native-key-controls.cjs dependency. JavaScript/CommonJS/Python syntax and physical hashes passed. The launcher explicitly set LOCAL_PRODUCTION_FIXTURE=''. The original 613 pins were preserved with fresh helper paths/hashes; the published tools/native-startup-barrier.py was appended as provider 614. All 226 accepted references, 23 existing media references and 49 repair artifacts were hash-verified. Launch disk exceeded 7.46 GB.

Provisioning awaited all 615 individual uploads: 614 providers plus the manifest. Only then did the remote barrier content-verify 57,913,808 provider bytes and exclusively publish staging-ready.json. Immediately before queue creation, the queue required that receipt. Immediately before consuming the one launch request, the custodian required it again in-process. All three recovered proofs agree on manifest SHA-256 99274f81373d3261f3d63b44191199ffdf4dc47b431384b22722f1c4256c9ccf. There was one UUID-idempotent allocation and one original, with no queue replay.

The remote custodian held its own runtime lock and subreaper custody before provisioning through final descendant drain. No local heavy runtime or local runtime lock was used. An observable detached independent 25-minute watchdog was armed before allocation; overall/browser/cleanup deadlines were immutable.

Reviewed native inputs and authority

I personally inspected the current startup PNG: the Graphics tab and OK button were visible. The ordinary Enter request then ACKed and native-start.png showed the alley. All five actual CDP-parsed JavaScript providers matched their accepted pins. The actual CPU and owning renderer modules matched d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc; the owning context was native WebGL2 on ANGLE Intel UHD 620. Every admitted input retained the existing source/owner/scene identity and freshness guards. The accepted retained-native pixel proof was reused without another costly pixel logpoint.

One ordinary Up hold requested 2000 ms and independently released at 2000 ms. The later fresh image showed the slope nearer; no repeated no-progress Up was sent. After reviewing that image, a standalone Alt hold requested 1500 ms. Its independent timer receipt measures 1501 ms; the trusted DOM events are exactly 1500 ms apart, AltLeft, with native key-up recorded. Its immediate held capture and subsequent settled captures do not prove a jump animation. No Alt+Up, simultaneous-key expansion, firing, guest write or scene manipulation was used.

Three separately reviewed left turns requested 1500, 2000 and 2000 ms, independently released at 1500, 2000 and 2001 ms. The final settled view shows a definite changed Lara orientation. That new native turn is the visible consequence from this run; an area beyond the slope is not claimed. final-held-input-state.json records all 16 admitted keys at zero, with all twelve trusted native events balanced.

All 20 commands ACKed, including three explicit stale-scene refusals with zero effects; each refusal was followed by a fresh capture and personal review. No runtime UNKNOWN occurred. One observation request omitted waitMs and was refused before queue publication; the corrected bounded request used a distinct ID. The missing-image diagnostic after a stale refusal and the first drain assertion are retained. The first drain still found the normally exiting controller; a bounded passive observation confirmed its exit before a distinct fresh drain passed.

Final exact accounting is 670 raw = 53 qualified retained + 617 discarded, comprising 602 coalesced records and fifteen retired on normal shutdown. Qualification, source token and publication evidence remain in the captured JSON. These are observation counts, not FPS.

Validation, concurrent checker update and lifecycle limits

python3 R/audit-staged.py passes: provider/barrier generation, actual parsed sources/modules, command ACKs/refusals, independent releases, trusted balanced events, all-key release, qualified-source accounting, eighteen PNG chunk CRCs, retrieval hashes, remote drain and local lifecycle. Accepted 10 closure-tool + 12 integration + 11 key + 11 barrier tests, prior locked builds and renderer/guard controls were reused by hash, without fixture reruns. No new production change required another build.

A peer published stricter import-occurrence checking in 0099373b4 while this original was running. The local checker initially used here is recoverable exactly from starting commit 117fb7d1d262408cacc00b0362a4c5644cc37c8e, SHA-256 3e5106726c2780292f29803f62b58f545263d4ea544f20fca1a5970704ef754f. The final audit binds that historical source rather than pretending the current mutable tool is unchanged. The new checker SHA-256 99cd56de12b3b2d19a8edf1f3b2ea53a1198a38632846217cce871752104a449 now refuses this driver's unsupported inactive generated.cjs import before another allocation. closure-tool-concurrent-update.json retains the refusal. This run's actual fixture environment was empty and no fixture handler executed. A successor must make its normal-driver closure pass the stricter tool; this closed lifecycle is not launch-ready recovery authorization.

Before normal stop, 251 artifact references / 227 distinct hashes / 289,052,691 bytes were retrieved and independently reread. Drain verifies all 25 recorded remote PID/start identities, bootstrap/custodian and adopted descendants absent; controller and Chrome gone, native Chrome exit zero, both ports closed, profile and writer absent, all admitted holds released, all 614 runtime provider hashes unchanged and the remote lock free. All three local lifecycle PIDs are absent, both supervised children exited zero and the supervisor ended empty.

Coverage remains the explicit 50 ms descendant census plus subreaper adoption/reaping, with parent-awaited synchronous diagnostics, not exhaustive kernel task history. Initial local lifecycle start ticks were not recorded. A supplementary global FD scan was requested after the automated normal stop had already completed and was refused machine_not_running; that failure is retained, no box was resumed, and no supplementary global FD census is claimed.

Box bx_ba4h4axs is archived, expiry null. Normal stop stop_e0219e6ba648c78f completed 15:07:43.475 UTC, forced=false, error=null. No task box remains billing.

Exact Git SHA, empty relevant binary source diff, small owned input/helper patches, pinned blobs and existing media path/hash references provide recovery. No source/media archive, peer WIP, public deployment, old-box resume/delete or subworker was used. Next: fix the normal driver's physical closure under the stricter checker before any separately assigned continuation; jumping and travel beyond the slope still need visible proof. Full27, parity, audio and FPS remain open.