Q3 original with screenshot completion telemetry — 2026-11-09

One fresh original reached the main menu and Q3DM1 Arena Gate level selection. Actual movement, firing, in-world rendering and death remain UNVERIFIED. The original ended at the first command UNKNOWN; there was no replay, extra capture, second original, cap/options/ACK/input change or speculative GPU/emulator cause.

This run reproduced and repaired an actual diagnostic defect. Successful screenshot300 returned in 52.595394ms, but later completion snapshots called its record age 39,546.316ms, then 41,338.187ms, elapsedMs. This falsely made a completed fast capture look slow while screenshot343 was pending. The helper now records the first response/send-failure/CDP-deadline settlement and keeps that elapsed interval fixed; ageMs separately describes record age. A response after CDP deadline preserves the deadline interval and sticky UNKNOWN. The owning composition pins the new helper. The marker-clock source from 7227918eb is byte unchanged, SHA256 1858eea7efb8875e44088ffe9f5ea434c3b22915444113204fb51673e4f5ac0d. This corrects duration attribution; it does not fix screenshot343's missing reply.

Evidence is in scratch/runs/20261109-q3-reviewed-provider-gameplay-0009/. The exact retained completion log is selected with original/stored SHA and sizes in actual-duration-selection.json.gz; actual-duration-replay.json.gz verifies the correction on those actual events. Focused controls exercise the composed send/receive functions over a real localhost WebSocket and actual numbered consumer, including the unchanged real20s deadline, wrong sessions, late replies, no ACK/replay, lifecycle events and bounded storage. The added duration regression covers successful/error replies, pending capture, send throw and late reply after deadline. All seven controls passed with socket/server closed, pending0/timers0 and actual wait0/EOF. The repaired canonical full local check and final locked build passed. The first locked build refused a peer's active producer-hash WIP; that owner refreshed its reviewed pin before the final build. These offline protocol controls provide no synthetic gameplay proof.

SuperTux2356's explicit archived/drained RUNTIME_RELEASE at00:08:23 preceded allocation. The final canonical914-provider local prelaunch passed against Git cc2cb5a31f0a65c49364394e2d8c7e10ce2f6dc5 with no relevant tracked WIP. A fresh UUID, independent observable25min watchdog and finite1499s TTL were prearmed. The identical full remote provider barrier preceded the sole original claim. Composition remained before execution preload. The current Q2-published CPU module was pinned and actually observed in the browser: a320d4c85156644fd12b6134d66b11784f3100b9b135756206c087cdacd43295, 2185596bytes. This is a current-module attempt, not the prior2337 module. Accepted unchanged controls, including0003's nine prelaunch controls, were reused by verified PATH/SHA. Git SHA plus small relevant patches retain source identity; there are no source archives, repeated media/module copies or peer edits/deletion.

Four captures completed and were personally reviewed: startup, singleplayer, enter-menu and fight. Commands0–2 have actual ACKs. Native click/Enter evidence does not establish DirectInput polls or simulation steps. The FIGHT click's capture remained at level selection, with the cursor near SKIRMISH; its host ACK was not called guest acceptance. One subsequent reviewed500ms Enter had an independent UP and settled physical-release promise, then attempted its ordinary capture. Command3 activate_level hit first UNKNOWN at00:21:52.030Z, no ACK.

The new completion log records screenshot343 at00:21:33.817Z, session 9794E37A7491B85C6E74869A4031BE8B, original {format:"png"}, send returned about0.601ms later. There is no raw response for that request through its 20,003ms CDP deadline or transport close; no late response or routing/session mismatch was observed for it. Existing page evaluations344–347 returned while the screenshot was pending. PAGE/guest/renderer owner reads342/337/339 returned before submission; there were no such owner replies during it. The render owner reported apiGL/backendWebGL, failed:false, control:null. No frame/context/target replacement event was recorded between submission and first UNKNOWN; absence does not prove continuity. Target detach/destruction followed Browser.close. This narrows the observed boundary to a submitted capture with no raw response; it does not distinguish compositor/readback internals or prove their cause.

Automatic first-failure close received Browser.close ACK; Chrome exited0 without signals, normal:true/drained:true in16876ms under the unchanged30s cap. Runtime, native command and outer launch status1 are preserved. All240 owned tasks had actual reaps and before-wait PID/start witnesses across2226 journal events. Actual parent wait/EOF, ports33557/44001 closed, absent profile/writer lock, empty runtime FD holders and a drained runtime lock were checked. Untraced clone flags and nonleader completeness remain UNKNOWN; historical local143 is unchanged.

All92 remote regular files were recovered or identified as exact pinned inputs; 88 unique retained artifacts total2016806 stored bytes. Independent rereads check stored/original SHA256, sizes and GNUgzip CRC before delegated normal stop. Box bx_xh3swzjg is archived, stopped00:22:36.249Z. Stop stop_30cf9529bd252586 completed00:22:36.266Z, forced:false/error:null. Initial expiry/archiveAfter was00:43:02.249Z; final archiveAfter:null and no separate expiresAt returned. The watchdog had actual same-shell wait0/EOF; allocation/watch parent also wait0/EOF. Explicit archived/drained RUNTIME_RELEASE was posted00:22:55. No runtime remains held by this task.

Next: investigate the missing raw screenshot343 response using the retained actual session/lifecycle records. The duration repair is tested offline and does not authorize another unchanged original or extra screenshot retry.