Final capture freshness pipeline investigation

Evidence: scratch/runs/20261120-hexen2-final-capture-freshness-pipeline-repair/.

No causal repair was established. The proposed status relaxation remains withdrawn. The inherited strict adapter and shared runtime were unchanged. There is no old-fail/new-pass native result or class/difficulty/level credit.

The previous genuine first DOWN is reconstructed in prior-pipeline-analysis.json from retained timestamps, without refreshing any sample. Its prepare-time producer was wall1795188663333, monotonic1795188663333.0051. Prepare returned at3355; final whole capture started3356 and completed3489; retention/focus completed3526; socket send3527; native cancellation3684 (sample age351ms, token0, no application, deadlines unexpired); trusted DOM DOWN3686; the later real producer3694. Absolute wall prefixes are179518866 for these shortened times. All15 prepare predicates passed; the later stale sample failure preserved the fourteen structural checks.

The prior long-task observer reports a 234 ms page task from1795188663453.01 through3687.01, spanning socket send and cancellation. The matching long-animation-frame entry has no script attribution. Thus capture/retention contributed 170 ms after capture start, and socket-to-cancellation contributed 157 ms. The specific blocking function, any newer sample before cancellation, and a cached-sample defect remain unknown. Reordering or relaxing status cannot be presented as a proved fix.

Before browser creation this task retained complete executable recipes and hashes. Read-only wrappers were prepared around actual bridge observations, host imports and Wine/renderer methods; they preserve original arguments, results, exceptions and clocks. They never generate USER polls, refresh timestamps, force frames or modify admission. Their detailed first-DOWN instrumentation was not reached.

The fresh original loaded d3c95bc2 in both actual CPU/render Workers and verified all 18 media inputs. Entire 800x600 host and focused-menu images were personally reviewed; trusted host focus and native binding passed. At2026-11-20T15:44:07.710Z the existing continuous disk-floor check closed the scope with1,499,983,872 free bytes on /home. Closure preceded prearm, strict prepare and DOWN. The explicit Enter300 command was left unprocessed; its file was written without a closed-scope check, potentially after closure, but caused zero endpoint requests or input. No replay occurred. Class/difficulty/level remained unrun.

The finite native supervisor joined all 22 recorded processes with leaked=false. Browser, display, host-release and native-release helpers have actual exit/EOF receipts; no native edge was armed. Zero parent endpoint requests followed closure. Owned profiles, display authority/socket/lock and empty temporary directories were removed, including the exited browser's two singleton links. The launcher root and Node cache were preserved. No Box was created; ID, expiry and stop receipt are null. Independent actual disk readings recovered above 1.5 GB after runtime cleanup; this does not reopen the scope.

Focused strict Quake 51 and Hex 42 JavaScript controls passed, including stale/frozen/replaced/retired-owner controls and existing expiry/UP handling. These validate unchanged guards, not a native repair. No redundant full build was run because this task changed only this note and ignored run recipes/evidence. Publication outcome is retained in the run. Shared bridge/host/WAT and other owners' files were preserved.

Precise completion blocker: the fresh trace was interrupted by the actual disk floor before the first native action. The narrow remaining investigation is attribution of the retained 234 ms page task across final capture/socket and genuine check_input publication, using the retained read-only probes after the owner authorizes a fresh scope. No shared runtime cause has yet been isolated for assignment as a code fix.