TR2 original stopped on death before weapon observation — 2026-11-08

The single fresh original ended on the assigned death stop before Control. Lara was visibly down with an empty health bar at face-dog-fresh.png. Zero Control intervals were sent. The requested downstream weapon trace was not obtained; consumptionQualified, linked weapon status, ammo, target, LOS, locks, arm frames, shot result and target HP are unobserved. There was no restart, replay, additional gameplay after death, or production repair. This is an incomplete weapon investigation with a concrete gameplay blocker, not a qualified consumption result.

Evidence directory (R): scratch/runs/20261108-tr2-observed-weapon-1858/. Starting checkout SHA is eb064c3c26ced4cef7a6a8f222fb6d7470b6a28d; relevant.diff is empty. Git and small run-helper patches provide source recovery. Existing media/modules are referenced by exact path, size and hash; no source archive, media copy or peer WIP was retained.

Selection and fresh startup

The explicitly selected manifest is scratch/runs/20261108-tr2-weapon-consumption-1832/runtime-pins.json, SHA-256 ab9355dd50d0057c2541d910c6d35f0cf61060054086c2505633f6f730b76bb2, published by aadf4012e. Normal relocation keeps 615 providers/29 physical helpers and all 586 non-helper providers unchanged, including the weapon observer, shell guard 6de923cb8, and CPU d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc. The helper retains explicit weaponObserver:true only with actionObserver:true; neither flag was exercised because death came first. Custody history alone uses individually appended gzip members from startup. The installation deadline, record cap, actual-retired accounting and independent native release code are unchanged.

The cheap exact resolver, physical closure validation and actual driver preloader passed before allocation. The reviewed 95 references/7,418,690 bytes were independently rehashed unchanged. The prior 126 weapon/32 Action, direct/RPC, closure and locked-build proofs were reused without fixture churn. Historical manifests are immutable. No new production build is claimed.

Personally reviewed review/webgl/startup.png and startup-reviewed.png show the readable original Graphics property sheet before normal Enter: Primary Display Driver, Hardware 3D Acceleration selected, Full Screen, 640×480 High Color, and visible OK. Read-only fresh window inspection records main HWND65537 z2, dialog HWND65538 visible/unowned z10, and Graphics child HWND65545 under that dialog. The original input list was empty. One ordinary 100ms Enter accepted these settings. This actual shell is 945×889 and single-app=false; it does not independently reproduce the earlier single-app startup failure or establish new guard causality.

Fresh Chrome 151.0.7922.108/V8 15.1.206.16, OS, targets, SystemInfo and command lines are retained in authority/custody records. Owning D3D target 035DC1190F507A8AE75DBA2AD762A960, endpoint1, requests {api:legacy,backend:webgl} without failure. Its observed context is WebGL2, 640×480, and actual owning CPU/renderer modules both have the d8fac hash. The owner-returned ANGLE Intel UHD620/Mesa23.2.1 string disagrees with fresh SystemInfo SwiftShader/SwANGLE and appended SwiftShader command-line flags. Both surfaces are preserved; physical Intel GPU use remains unresolved. No backend, flag, viewport, clock, seed or binding experiment occurred.

Normal observe established ARM generation1, process:1000:endpoint:1, and a non-null live initial journal containing one raw record without failure, verified before any gameplay command. The first local confirmation attempt lacked its downloaded publication file; a read-only retrieval and corrected three-command check verified it without resending an observe or input action.

Reviewed approach and failure

Ordinary Space drew pistols. Reviewed turn/movement captures show upright Lara, drawn pistols and initially full health. align-left-two.png first shows a small dog down the alley; Lara faces the shuttered wall rather than the dog. The later align-right-two.png already shows the dog close beside her. The subsequent advance request's review text incorrectly calls it "distant". That was a review error: the retained pixels already showed an immediate threat, and the next movement did not achieve the required safe dog-ahead setup. No emulator or Control defect follows from that error.

advance-path-one.png shows Lara upright near the left wall with the dog beside her and health roughly three-fifths. A turn request then refused as stale before dispatch, zero effects. Fresh dog-ready.png still shows Lara upright, but the health bar is already empty; its review text incorrectly carried forward the preceding health estimate. A single released ordinary right turn followed. face-dog-fresh.png, captured at 19:00:26.300 UTC, shows Lara lying on the ground with an empty bar and the dog still standing. That is the death stop. The next and last command was ordinary quit.

No Control was sent, so there are no poll/Action/weapon observer records to interpret. Drawn pistol pixels do not establish numeric status4, ammo, arm readiness or targeting. No reviewed muzzle flash, bullet hit, kill or effective aim was established. The exact offline dependency map remains applicable but supplies no missing runtime values for this original. Internal WAT reentry and other threads remain unobserved; missing trace rows are not evidence of absent execution.

Evidence audit, compact custody and stop

python3 R/audit.py passes independent artifact reread, startup stack, fresh owner/module selection, ARM-before-gameplay, zero Control commands, balanced release receipts, final sampled physical keys clear, PNG CRCs, raw accounting, observed drain and final stop verification. There are 20 commands/20 ACKs, one zero-effect stale refusal, zero transport UNKNOWN records, 13 released gameplay/startup holds, and 19 valid PNGs. Audit failures from initial overly exact command-dictionary and wrong final-sample assumptions were corrected to inspect the actual transport timing field and after-release capture; no runtime input was repeated for these local audit corrections. Raw accounting is 393 = 60 qualified retained + 333 discarded (330 coalesced, three shutdown), without publication failure. This is not FPS or weapon-consumption qualification. Full27/parity/audio/FPS remain open.

Before delegated normal stop, 286 references/259 unique hashes/98,023,045 referenced bytes were retrieved and independently reread. Only 178 unique files/15,016,196 transport bytes needed downloading; matching reviewed images and existing modules use same-hash path references. Large JSON/logs are individually gzip-compressed. The custody log was gzip from its first record: 11,656,864 compressed bytes independently stream-decompress to 144,125,578 bytes, SHA-256 36864b831ecf3f8958e331afabdf1b7298a9335718db27616bf25091001ca241. There is no uncompressed local duplicate. retrieval-verified.json maps every remote original path/hash/size to its local bytes or gzip; RECOVERY.md describes streaming recovery. No old run was deleted. Own observed free-disk samples and download guards exceed 1.5GB; the shared board independently reports a transient host floor violation at 18:59:27 during the peer's local Chrome, followed by that worker's shutdown. Continuous host-floor compliance is therefore not claimed. The final local sample is about 1.52GB free.

Chrome has an actual exit0/signal-null receipt at 19:00:42.383 UTC. Drain verifies 30 observed remote PID/start identities absent, including adopted children/bootstrap/custodian; ports46709/42285 closed, profile/writers absent, held inputs released and remote lock free. Seven direct/adopted children have durable before-reap identity plus actual terminal wait receipts. Other Chrome-owned descendants have observed absence, not individually retained terminal waits. This selected custody uses a 50ms census, subreaper adoption and WNOWAIT registration; it is not exhaustive kernel history. CLONE_UNTRACED and nonleader exec remain unaudited. No ESRCH exit inference or stronger lineage claim is made.

One persisted UUID preceded one create. Three independent observable local lifecycle PID/start identities were armed before allocation with an immutable 25-minute cap. Box ID and initial expiry 19:20:39.030 UTC were persisted before runtime. All three lifecycle roles subsequently exited, supervisor empty. No local heavy Chrome/runtime lock was used by this task.

Box bx_h9dqjgej is ARCHIVED, expiry null. Normal stop stop_b0aaa2aaee36c2b8 completed 19:01:49.757 UTC, forced=false, error=null. The lifecycle's first archived receipt still showed pending; the separately retained final API inspection proves completion. No box remains billing idle. No old-box resume/delete, public deployment or subworker occurred.

The downstream weapon question remains open. A separately authorized fresh attempt would need to reach the reviewed setup and act before the close dog exhausts health, with accurate current-pixel review; this task performs no additional original, hold, repair or gameplay. Single-task worker stops.