Redline command/capture-trace boundary repair

Evidence: scratch/runs/20261109-redline-command-trace-deadline-repair-2230/. Code on main: b454987afd4272a335e2ebd3af8ab2f0dee0b2f8.

The generated consumer previously awaited diagnostic worker traces after the required image proof and publication, before returning to the numbered ACK consumer. The retained 2210 resume image completed at 22:19:26.843Z; command 9 failed its work deadline at 22:19:31.797Z. Both guest-worker snapshots survive. The next worker in the actual sequential consumer is D3D render worker F53ED06D5E3EB7AC68AA136AC7C3C719; its resume snapshot is absent, and the pending diagnostic later rejected on runtime closure. This locates the third await by consumer order. The raw wire ring had exhausted its capacity, so the exact stalled CDP request ID and browser service delay remain unknown. These facts do not establish a guest-death, blank-frame or startup cause.

The current-publication route now collects diagnostic traces concurrently with a maximum 1,000 ms wait and a 100 ms command reserve. It permits only one outstanding diagnostic read per worker and records incomplete reads explicitly. Late results cannot publish trace artifacts or issue another protocol request. The existing CDP guard still owns each pending read. Required state, image, identity and current-lock proof remain awaited. Late images and terminal work are discarded, and the command checks its original work deadline before ACK. No work, CDP, admission, mouse-release, custody or cleanup guard was extended. Shared providers remain unchanged.

The focused test executes the actual composed capture/execute functions with the deployed numbered transport. The old sequential trace tail reproduces a completed-image/no-ACK deadline failure. The repaired held-trace control returns an exact image ACK with an explicitly incomplete diagnostic receipt. Success, late image, required-capture refusal, terminal interruption, sticky later-command refusal, duplicate outstanding-read suppression and late-result publication controls pass. Existing current-lock planning and terminal-close/review tests pass, along with four durable-journal tests, six bounded-custody tests, exact provision/media/barrier preflight, actual generated parent/controller/caller preflight, conditional-original preflight and flock /tmp/wine-build.lock bash tools/build.sh.

Both phase manifests and all task inputs were sealed and independently reread before the sole allocation. The native manifest contains 906 providers and has SHA-256 584e82ddbf7739a0c6531024811a6c412552d03085c6c07e3cbe530efa9a958d. Source recovery uses exact Git commits and small patches, including a verified reference to the preceding run's reconstructed non-Git inputs. No source or media archive was created. Durable CONT/EXIT handling and physical reaping were retained unchanged.

The one changed native fixture passed. Chrome published in 505 ms / 5 polls of the unchanged 100 x 100 ms gate. Startup and comma images were separately reviewed before dependent input/close. Ordinary quit ACK and parent exit 0 have separate stdout/stderr EOF; custody reaped all 177 tasks.

Only after that pass, the one conditional original published in 705 ms / 7 polls. Each returned image was separately viewed before dependent input. The launcher click used this image's saved canvas geometry. Reviewed ordinary Enter selections reached New Game and Stadium City; passive captures followed the changing introduction into the on-foot health100/ammo200 HUD. Ordinary arrow controls produced reviewed garage views and visible car front/side geometry. The retained short Right control includes 83 positive guest VK39 reads. Short turns sometimes left the scene nearly unchanged, while longer turns crossed the garage; no movement or entry was inferred from an ACK alone.

There are 23 successful exact numbered ACKs, each with a hash-matched complete image. Seven of their diagnostic receipts explicitly report incomplete traces. Every successful image completed at least 5,128 ms before its command work deadline. The first terminal failure is command 23, focus-world, at 22:50:22.409Z: mouse delivery exceeded hold bound. The reviewed current image was unlocked; its saved canvas geometry was checked before an ordinary center click. The unchanged mouse helper independently invoked release at its hold boundary, then refused because the press had not returned before release was invoked. Its three mouse request records each retain one reply, but raw wire capacity was exhausted; the exact service delay is unknown. No command work deadline or capture-trace failure was reported for this command. No subsequent capture ran, no failed-command image/ACK exists, and no later numbered input, capture, retry or quit was requested. Side-door entry and driving remain unverified. Further work must discriminate this mouse service boundary before a fresh entry attempt; this runtime must not be resumed or replayed.

Original custody reaped all 348 tasks. Its parent actually exited 1 with separate stdout/stderr EOF. Chrome cleanup reports SIGKILL, truncated stderr and streamEOF false; browser EOF is not claimed. The surrounding original tool session returned 143, separately from the durable parent wait. Both lanes' server/DevTools ports were independently closed, no profile processes survived, and the remaining original profile was removed after physical drain.

Before shutdown, 573 remote artifacts / 491 distinct hashes / 21,177,418 bytes were retrieved and independently reread for SHA-256 and size. PNG CRCs and gzip CRC/EOF were checked. Original Chrome stderr is retained as explicitly invalid gzip/EOF cleanup evidence. Unique logs, captures and modules remain in the run.

The sole box was bx_u92q6cbf, UUID f5f28deb-fc68-4048-ac65-1a334b1db1c6. Initial expiry was 2026-11-09T22:57:14.734Z; the immutable independent watchdog deadline was 22:57:14.782Z. Exactly one delegated normal unforced stop completed at 22:52:34.627Z, receipt stop_09f7bff61bc5c8af. Independent API reread reports archived with archiveAfter null. The watchdog was actually waited, exit 0, with separate stdout/stderr EOF, unchanged configuration hash and no surviving PID. Its allocator tool also returned 0. No owned box remains running. Free /home remained above 1.5 GB.