TR2 single observed Control — 2026-11-08

One native Control hold ran; the Action observation is unqualified. Lara had visibly drawn pistols, full health and a dog ahead. The published observer exhausted its 4096-boundary allowance at boundary 4097, after 32.105 ms, with no witnessed poll or Action records. It restored the original instance and the normal driver stopped immediately. No repeat, restart, replay, hold extension or second original occurred. This is a concrete observation-coverage blocker, not evidence of a Control defect or absence of guest polling. No substantive code repair is claimed.

Evidence: scratch/runs/20261108-tr2-observed-control-1717/ (R). Starting Git SHA 3898779fb3565d92d2a9ebb40b99aa1e1ac84e13; R/relevant.diff is empty. Production code is unchanged. Exact Git SHA, small owned helper/note patches, unique logs/captures/modules and existing media path/hash references support recovery; no source/media archive or peer WIP was copied.

Reviewed sequence and result

Only Graphics Enter and ordinary observe preceded confirmed ARM generation 1, owner process:1000:endpoint:1, with a non-null live failure-free initial raw journal. There was no gameplay input before confirmation. A redundant local download refused because the identical ARM file had already been fetched; the successful remote confirmation and independent hash match are retained. This local collision was not a runtime UNKNOWN or input replay.

A separately reviewed native Space hold of 1500 ms and passive settling established drawn pistols. right-four.png clearly shows both pistols raised beside Lara's head and full health. Short reviewed turns and forward holds followed. advance-six.png shows Lara upright, pistols drawn, full health, and a small black/brown dog ahead near the far opening. Target lock and effective aim were not established. Two stale admissions refused before dispatch with zero effects; subsequent requests used freshly captured, reviewed scenes.

Exactly one command used key:Control, holdMs:2000, actionObserver:true. Its native interval was 17:22:19.234–17:22:21.235 UTC, 2001 ms recorded with the unchanged independent release timer. Physical VK17 and VK162 were sampled down wholly inside the interval; VK163 was clear. No extension was requested. The observer report is input-observed-control-once-action-observer.json: status:boundary-limit, boundaries:4097, records:[], pollObserved:false, pending:false, removed:true, qualified:false. Correlation records physical Control but zero in-interval poll/Action pairs. HRESULT, DIK1d/9d, owning bitmap, binding/conflict and paired EBX40 therefore have no actual witnessed rows in this run.

Personally reviewed after-observed-control-held.png shows Lara upright with full health and raised pistols, and the dog closer ahead. No clear muzzle flash, impact, hit, kill or target lock is established. This is a held, not after-release, picture. The driver closes on unqualified observation before the normal after-hold capture; no after-release pose/consequence was captured. Death is not shown in the reviewed frames. The final key bitmap was not independently sampled before close; release receipts and browser/process drain are the retained release evidence. The task stops at this unqualified result, without manufacturing further combat evidence.

Some short command-review descriptions misstated camera perspective or wall side: right-four is frontal, right-five rear/side, and advance-one is beside the right brick wall. The personally reviewed corrections and image hashes are in R/reviewed-consequence.json; the original request text remains retained.

Preserved authority and validation

The explicit selected generation is scratch/runs/20261108-tr2-action-observer-1652/runtime-pins.json, SHA-256 1b931b9a4f9b1266651575a9d79ec6e76cd09c44bdbf5d9c7655595194f8d00b, from published 16919b487. Normal helper relocation retains 615 providers / 29 physical helpers. Non-helper providers remain selected bytes; historical 614 generation remains unchanged. Only run paths and obsolete 5 GB disk checks changed in owned helpers, following the owner's 1.5 GB rule. No binding, input policy, backend, guest clock, seed or private guest writes changed.

Before allocation the exact resolver validated all providers/helpers, syntax, 226 accepted control references, 23 existing media references and 49 repair references. The actual selected driver preloader resolved 615 providers and exited without Chromium. The unchanged 25-observer, direct/RPC input-path, 13-helper, HTTP resolver and full-build evidence was independently rehashed and reused, not rerun as a separate proof project.

All five parsed JavaScript sources match their selected pins. Actual CPU and owning native WebGL2 module hashes both equal d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc; backend remains WebGL2 on ANGLE Intel UHD Graphics 620. Existing pixel authority is reused unchanged. Internal WAT reentry and other threads are unobserved: no witnessed poll does not imply no guest execution or poll.

python3 R/audit.py passes: ARM-before-gameplay, joined staging barrier, module/parsed-source pins, exactly one observed Control command and one actual hold, physical sample, fail-closed observer removal, release receipts, 27 PNG CRC checks, artifact rereads, lifecycle and drain. There are 28 commands, 27 ACKs including two zero-effect stale refusals, and one terminal unqualified failure, with zero transport UNKNOWN records. Raw accounting is 544 = 68 qualified retained + 476 discarded (471 coalesced, five shutdown), with canonical read sources retained and no journal failure. These counts are not FPS. Full27, parity, audio and FPS remain open.

Closed lifecycle

One UUID idempotency key was persisted before the single create. The independently observed 25-minute watchdog and supervisor/normal/watchdog PID/start identities were recorded before allocation. Box bx_22873stk, initial expiry 2026-11-08 17:40:40.582 UTC, was recorded before runtime. Browser, cleanup and overall deadlines were immutable. Remote custody held its runtime lock from startup through drain; no local heavy runtime or local runtime lock was used.

All 616 uploads joined before verified staging publication; queue and parent consumption match manifest 10892ce2305cafdd8299f53043fffdd8c9fa8bfaade4111dbc0d8bf31dc3b819. Before stop, 378 artifact references / 345 unique hashes / 392,195,084 referenced bytes were retrieved and independently reread. Drain passed for 27 remote PID/start identities, including observed/adopted descendants, bootstrap and custodian; endpoints, ports, profile, writers, held-input receipts and remote lock drained, and all 615 providers were unchanged. Chrome exited code 0. Custody is a 50 ms census plus subreaper adoption and durable WNOWAIT registration, not exhaustive kernel fork history or identities already reaped by parents. Local lifecycle roles exited zero, their identities are absent, and supervisor ended empty.

Final state: archived, expiry null. Delegated normal stop stop_c4fce7b3ae9dc8ae completed 2026-11-08 17:23:44.625 UTC, forced=false, error=null. No task box remains billing. No old-box resume/delete, public deployment, external message or subworker was used.

Next requires separately assigned offline causal work on the measured boundary exhaustion before another original. Do not infer a Control defect, increase limits speculatively, or repeat this closed original. The requested qualified poll/Action and after-release consequence remain unproven.