Quake III: controlled Major frag attempt, 2026-11-12

Evidence: scratch/runs/20261112-quake3-controlled-major-frag-1840/. The player-frag goal remains open. This attempt stopped at a specific observation blocker: a native Difficulty-menu transition was followed by a 30-second Page.captureScreenshot UNKNOWN. No arena, outgoing damage, killmessage or player-score increase was established in this run. The earlier Major encounters and paired arena checkpoint retain their previously stated scope.

Current original and admission

Source is main 05d58d56e6222bf3f9b3f8157818471388d951dd, relevant binary diff and owned helper patches, recorded in the run. Git supplies source recovery; there are no source-tree or repeated-media archives. Provisioning verified 862 pins and the actual emitted import branch before original launch. The registered quake3_demo uses the existing original pak0, SHA-256 e77abad2466f45a0a7ea018445528f9b95a0fe7789fa1abc1a7718bbf0754b08. The current standard module is 2,189,863 bytes, SHA-256 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17; the older notes' 1165… module is not this run's module.

Before allocation, the final emitted browser/binding/input/release helpers ran against actual local Chrome and original Boids. All 22 binding fields were populated. Twenty-five negative controls refused changed review identities, ambiguous module metadata and an actually replaced owner before dispatch. An eight-second key hold with separately released one-second steering, an eight-second mouse-button hold, and two completed full images during those holds passed. Independent releases also passed slow sampling, UNKNOWN key DOWN, partial key delivery and controller-loss controls. These establish physical delivery/cleanup properties, not guest polling or combat timing.

In the private original run, actual guest target 432897632037C844441BF63526CC10EC and renderer target 125388C2D04425B05AEADCE98DA6D7C7 both recover the current 04c534… module. The actual CPU instance token is 71f6509a-0bf2-482d-9ae5-7abe2cd64081:1; the renderer endpoint/instance epoch is b811dc5d-160e-41d4-9123-8dc8dab01085:1/b811dc5d-160e-41d4-9123-8dc8dab01085:2. GL endpoint 1 owns a real WebGL context, bound as b811dc5d-160e-41d4-9123-8dc8dab01085:3. Owner-scoped completed draws, vertices, presents and display write sequence are positive before menu input; software-front completions are zero. Neither a dropdown nor a legacy display transport supplies this producer identity. Context GPU vendor was not separately retrieved; the browser SystemInfo record is retained without substituting it.

Reviewed route and first blocker

The native route reached Single Player → Q3DM1 Arena Gate against Major → Fight → Difficulty. difficulty.png shows Bring It On selected. An Up selected Back, and Enter returned to Choose Level. The initial review text incorrectly called this an I Can Win selection; the full images and difficulty-review-correction.json explicitly correct it. No I Can Win choice or easier fight was earned. Another reviewed Up focused Fight.

Command 9 then sent native Enter. Its DOWN and independently prearmed UP were acknowledged, with trusted DOM edges and released button mask zero. The following full capture never returned within the 30-second CDP deadline, despite fromSurface:false. unknown-290.json and command-9-failure.json identify Page.captureScreenshot; difficulty-again.json preserves the preceding state observation. No completed replacement image exists and no dependent input or replay followed. Earlier completed captures also show substantial observation latency. This is a capture-path blocker, not proof of a guest trap, bad input consumption, rendering defect or native combat failure. No production repair is justified by this evidence.

Read-only recovery independently downloaded each of 113 paths twice, then checked local raw hashes: 108 unique hashes, ten full PNGs, nine stable 22-field capture bindings and seven balanced native key pulses / fourteen trusted edges. The validator passes; six focused native-input regressions and helper syntax / import checks pass. The final guest input queue remains UNKNOWN. No pointer-lock combat, weapon pickup, player frag or paired SoftwareGL follow-up was attempted.

Lifecycle and next work

An independent observable two-hour watchdog was armed before creation of private box bx_m824pzrv at 18:53:27Z, initially expiring at 20:53:26.562Z. Actual /home free space exceeded 1.5 GB at allocation and admitted actions; there was no RAM gate or reserve change. Chrome closed normally, exit 0 with separate EOFs. The runtime parent returned 1 with separate EOFs because of the screenshot failure. After unique evidence recovery and hash verification, normal delegated stop completed at 19:03:00.736Z: archived, expiry null, receipt stop_4c17f29ccb21827f, completed, forced false, error null. Actual manager and watchdog waits/EOFs both returned 0.

Next: qualify a bounded full-frame capture path that survives the original WebGL menu transition, then verify the actual native skill choice and acquire pointer lock for a short observed Major encounter. Preserve UNKNOWN outcomes without replay. The player killmessage/score increase, outgoing damage, paired SoftwareGL combat, FPS, parity, audio and full-27 scope remain open. This is a docs/evidence change; no production code or new locked build was needed.