TR2 dog encounter — 2026-11-08

Lara died in the single fresh original. Native weapon draw was visibly established before further alley travel, and the dog was identified ahead before it reached Lara. Three reviewed native Ctrl holds did not establish successful combat. The final personally reviewed capture shows Lara collapsed with an empty health bar and the dog standing above her. Dog survival, far-opening threshold/crossing, jump and landing remain unproven. Gameplay stopped on death; no restart, replay or second original was attempted. Full27, parity, audio and FPS remain open.

Evidence: scratch/runs/20261108-tr2-dog-encounter-1610/ (R). Starting Git SHA 9761d8b9ae1ef98b6927fbff41f696fb13f10704; relevant binary diff is empty. No production source changed and no substantive integration repair is claimed.

Controls, observation and consequence

Before allocation, the exact current checker verified 614 providers / 29 physical helpers, including the unchanged shared 16-key map. The actual ordinary key module was cheaply exercised with mock sends: Space, Control, arrows and Alt produce matching native keydown/keyup pairs. This is mapping evidence, not gameplay qualification. The 226 accepted control references, 23 existing media references and 49 repair references were independently reread and reused unchanged; unchanged fixture suites and the full build were not rerun.

The original PC manual, printed pages 2 and 7, identifies Space as Draw Weapon and Ctrl as Action, used to fire drawn weapons. No input-policy expansion or guest write was used. Independent releases were preserved; no compound gameplay input was used.

Only startup Graphics Enter and ordinary observe were queued when observe-confirmed-before-gameplay.json confirmed ARM generation 1, owner process:1000:endpoint:1, and an initial non-null, live, failure-free raw journal with zero records. No gameplay command preceded that confirmation. The cheap extracted ordinary driver check also passed before allocation.

Reviewed sequence:

Final capture: R/review/webgl/combat-assess.png, SHA-256 c18dce1a1585ba4870034b5070218a53b7243f563ffd41e2e00a90a2b189b99f. This attempt does not justify repeating stationary Ctrl holds as a proven dog strategy.

Preserved authority and validation

All five actual parsed JavaScript sources match their accepted pins. Actual CPU and owning native WebGL2 module hashes both equal d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc; owning backend is WebGL2 on ANGLE Intel UHD Graphics 620. Accepted pixel-authority evidence was reused unchanged, without a new pixel logpoint.

Final raw accounting: 837 = 145 qualified retained + 692 discarded (681 coalesced before publication, 11 normal shutdown). Every qualified row retains its canonical read source; journal failure is null. These are observation counts, not FPS. Historical journals remain untouched.

python3 R/audit.py passes: pre-gameplay ARM, manifest barrier, all five parsed sources, actual modules, publication accounting, 49 ACKs / five stale refusals with zero effects / zero runtime UNKNOWN, 32 holds, 64 trusted balanced native events, all 16 keys released, 44 PNG chunk-CRC checks, artifact rereads and lifecycle/drain checks. A few local attempted image reads after stale refusals found no output; no request was replayed to manufacture that evidence. Run-specific helper syntax checks and the cheap ordinary mapping/observation checks passed. Accepted unchanged tests were reused.

Closed lifecycle and recovery

An independently observed 25-minute watchdog was prearmed before allocation; supervisor/normal/watchdog PID/start identities were recorded while alive. UUID idempotency key was persisted before the one create. Box bx_pg79mqj7, initial expiry 2026-11-08 16:33:56.558 UTC, was recorded before runtime. Overall/browser/cleanup deadlines were immutable. Remote custodian held its runtime lock from startup through owned drain. No local heavy runtime or local runtime lock was used.

All 615 uploads joined before content-barrier publication; queue and parent consumption proofs match manifest SHA-256 6cdabd6ea8ab30a97c345b521bf3b6b9df4e7fd7749a4c7ada38fa2a29bf80fe. Before requesting stop, 608 artifact references / 555 unique hashes / 817,251,223 bytes were retrieved and independently reread. Drain passed for 31 recorded remote PID/start identities, including observed/adopted descendants, custodian and bootstrap. Endpoints, both ports, profile, writers, held input and remote lock drained; all 614 provider hashes remained unchanged. Custody covers a 50 ms descendant census plus subreaper adoption and durable WNOWAIT registration, not exhaustive kernel fork history or intermediates already reaped by parents. Diagnostic subprocesses were synchronous and parent-awaited. All three local lifecycle identities are absent; both child exits are zero and supervisor ended empty.

Box state is archived, expiry null. Normal stop stop_aad1229993aa52d1 completed 2026-11-08 16:21:57.751 UTC, forced=false, error=null. No task box remains billing.

Recovery retains exact Git SHA, empty relevant binary diff, small owned helper/note patches, unique logs/captures/modules and existing media path/hash references. BOAT.TR2 remains at its accepted existing path/hash e710716fcaf039fb3462e26e33f0ead8eb45bd01c8bbfa036c455a1d19823d9e. No source/media archive, peer WIP, public deployment, old-box resume/delete or subworker was used.

Next requires a separately assigned task: investigate why the observed drawn-weapon/native Action sequence failed to acquire or stop this dog before another original. This closed original must not be replayed. Stop after death; no success or repair claim.