Q3 corrected prelaunch and visible menu progress — 2026-11-08

One fresh original reached the Quake III main menu and the Q3DM1: Arena Gate level-selection screen. The reviewed screenshots are retained in scratch/runs/20261108-q3-corrected-prelaunch-gameplay-2337/ as startup-review.png, singleplayer-review.png, and enter-menu-review.png; their original downloads were stored as gzip first. Movement, firing, torch, in-world playability and death status are UNVERIFIED. No second original, replay, game/backend/configuration/cap/clock change or agent launch occurred.

Offline controls found and reproduced a substantive diagnostic defect: q3-execution-boundary.cjs subtracted entry/exit marker clocks even when their execution-context IDs differed. It now requires one positive integer context ID, finite nonnegative ordered timestamps, and a nonduplicated entry/exit order. Invalid pairs retain their observations and report a concrete unavailable reason and null duration. A CDP response remains independent. The composition's exact provider pin was updated. Published code commit: 7227918eb (actual production-expression wrapper negatives, composition compile and locked build PASS). The new controls use local protocol marker delivery, not a browser, game or GPU fixture. Accepted 2248/2305/2332 controls are reused by PATH/SHA references, without identical native reruns.

The actual published tools/q3-native-prelaunch.py --check-only was executed locally before allocation. Its first full graph correctly refused duplicate providers: local /binaries is a symlink to /test/binaries, whereas earlier remote uploads had made both aliases separate files. The corrected run input pins the 913 canonical providers once; the retained static server already rewrites /binaries/ to /test/binaries/. The barrier itself is unchanged. The corrected actual local check bound the runtime imports and passed without an original claim. The remote destination was different and fresh; its final full provider barrier preceded the sole claim and native parent. Neither the old launch-code.py nor the invalid module-name import was executed.

SuperTux2330's explicit archived/drained release at 23:37:31Z preceded allocation. A durable UUID, immutable 25-minute deadline, finite TTL and observable independent watchdog were recorded before creating bx_bfn5frew. The browser retained its 480-second ceiling and the existing supervisor, composition-then-execution preload order, 20-second work allowance, and 30-second close cap. The executing CPU and render-worker module SHA is 4b45bda53a90b665896cf3b85d9a157b7c3703563c88b6666e4ee2e4b346bcc3. The provider source base is recorded in the manifest plus small owned patches; no peer WIP, source archive, repeated media or module copy was staged.

Actual console evidence records Q3_DIAGNOSTICS_ARMED with beforeHostImports:true. Actual render-worker observations report WebGL 1.0, ANGLE with Intel/Mesa, an owning api:gl/backend:webgl endpoint, and zero software batches in the captured snapshots. These are observations, not GPU completion, device-performance or successful guest-input inferences. The GL partial explicitly remains UNKNOWN: its bounded ring overwrote records. PAGE, RENDER, WORKER and CDP observations remain separate; marker intervals are only within a validated execution context. No cross-clock subtraction is used.

Three bounded reviewed inputs were submitted: one 500ms menu click, a 500ms Enter that led to the captured level-selection screen, and a 500ms Enter on the selected level. The click has trusted DOM DOWN/UP and CDP completion but no claimed guest effect. Both key holds have independent UP attempts, actual UP completion and settled release promises in the unique run's compressed physical-completion telemetry. These host receipts do not prove DirectInput polls or simulation steps. Commands 0 and 1 received actual ACKs.

Command 2 (level_select) hit its unchanged work deadline at 23:44:57.824Z and received no ACK. At that instant the Node trace has Page.captureScreenshot request 345 pending for 16,726ms. Separate passive page evaluations 346/347 had returned; the last PAGE/WORKER/RENDER owner reads before the screenshot also returned. The screenshot later reached its own CDP deadline and remained UNKNOWN. The partial level-select-state.json is not a screenshot or successful command. No further gameplay input or capture was requested. This distinguishes observed boundaries without assigning GPU/driver/guest cause.

Automatic first-failure close produced an actual Browser.close ACK, Chrome exit0 without signals, normal:true and drained:true in 13,648ms. The native command, runtime parent and outer launch exited1; this failure is preserved. Final custody had 332 owned tasks, 332 actual reaps and 3,316 journal events, with each before-reap PID/start witness checked. Actual parent wait and stdout EOF, closed ports45445/37763, absent profile and writer lock, empty runtime FD holders and a free runtime lock were verified before normal stop. Untraced clone flags remain unaudited and nonleader completeness UNKNOWN; these final observations do not repair or qualify historical runs.

All 81 remote regular files (78 unique contents) were inventoried and recovered or matched to an existing retained gzip artifact before stop. Stored/original SHA-256, sizes and gzip CRC were independently reread. An initial recovery assertion correctly rejected using an equivalent decompressed gzip with different stored bytes; exact remote gzip members were then retained and checked. Initial pre-ACK screenshot fetches returned404 and are preserved as orchestration errors; they sent no guest/CDP request and did not replay a command. The completed enter-menu capture was subsequently retrieved and reviewed.

bx_bfn5frew is archived, stopped at 23:46:39.017Z. Stop stop_4a533be78017c274 completed at 23:46:39.039Z, forced:false,error:null. Final archiveAfter:null; expiresAt:NOT RETURNED. Initial archiveAfter was 2026-11-09T00:06:02.128Z, TTL1499s. The independent watchdog and its same-shell EXIT wait completed with status0 and EOF; the outer launch and every owned command child were actually awaited. Board-watch wait and exhaustive local inventory are sealed in the run. Free disk samples exceeded 1,500,000,000 bytes; no continuous-floor or RAM claim is made.

Next: investigate the actual screenshot request's missing completion using these retained independent boundaries before a separately authorized fresh gameplay attempt. This single task stops with the diagnostic implementation published, visible menu/level-selection progress, an explicit gameplay limit and an archived box receipt.