Redline causal mouse experiment: startup refusal before observation

Evidence: scratch/runs/20261109-redline-browser-mouse-cause-2310/. The authorized single native fixture ended at the unchanged Chrome publication gate, before any CDP connection, image or mouse command. This is a concrete startup blocker, not a mouse timing repair or vehicle progress.

The retained command-23 records were independently reread and hashed, with compressed inputs fully decompressed. They still establish press send to raw reply of 165.579 ms, independent release at 151.418 ms, and the press raw reply 14.161 ms after release invocation. Delayed local promise continuation is excluded; transport and browser service remain inseparable. A short socket-send call does not prove browser receipt.

Offline controls used the actual current composed execute/send/receive, deployed mouse/menu helper and numbered transport. Success, delayed press, delayed release, terminal interruption, sticky refusal, bounded records and no later protocol/capture passed. Delayed press remained refused rather than being accommodated. Exact provision/media assembly/barrier preflight passed without source/media copies. The generated parent/controller/caller preflight passed without Chrome, including actual waits and separate pipe EOF. The locked repository build passed. Production code and timing guards were unchanged.

The small task fixture controller changes the generated controller's ordinary comma key request to one ordinary center click with the existing 150 ms hold. It retains the current model-review gate, numbered transport, caller, browser, shutdown and custody. Before native launch, offline inspection corrected its manifest argument: the trusted digest now comes explicitly from the preparing parent, rather than being read from the runtime seal. The corrected generated launch was exercised offline again. Earlier and final small inputs are retained as patches. Both preparation manifests were independently reread; the native manifest has 907 providers and SHA-256 6473f96581ee525fb0c6f5f62383b4bde00beccd5668926ff717ae95cc5d05b6.

The causal question stated before launch was whether trusted DOM mousedown handling itself occurs after the independently scheduled release, allowing for the browser clock bracket, or whether the observed delay lies outside DOM handling. The small PE fixture cannot assume Redline's workload. A conditional original diagnostic was prepared, but never launched after the native refusal.

At 23:13:37.611Z the unchanged 100 x 100 ms DevToolsActivePort publication gate failed, elapsed 10,044 ms, last read ENOENT. Close admission was recorded at 23:13:37.612Z. The controller recorded the same terminal failure at 23:13:37.634Z. The caller's failure path immediately closed protocol admission; an additional explicit SIGTERM cancellation was sent at 23:13:57.299Z while cleanup remained active. No later protocol, input, capture, retry, quit or deadline extension was requested. The raw wire has zero requests and no raw drops. There is no browser clock sample, DOM marker, mouse-service command record or image. Missing markers provide no evidence about mouse transport or browser handling. No visible vehicle was established and no entry was attempted.

Cleanup retained a late DevTools publication, after terminal refusal, and an actual browser SIGKILL receipt with streamEOF false and streamsTruncated true. The late port is not readiness or permission to retry. The controller watchdog exited 1; its failed disarm assertion is retained separately from the first publication failure. Browser EOF and complete native trace are not claimed. Custody nevertheless physically reaped all 108 owned tasks, with no survivors. The ordinary parent actually exited 1 with separate stdout/stderr EOF. Both server and DevTools ports were independently closed and the remaining profile was removed only after physical drain.

Before the sole delegated normal stop, 73 remote artifacts / 71 distinct hashes were retrieved and independently reread for SHA-256 and size. Full gzip decompression retained Chrome stderr as explicitly invalid/truncated gzip; it is not called verified EOF. No source/media archive was created: Git and small relevant/owned-input patches provide source recovery.

Box bx_efhfrhjf, UUID 9247c46e-ba71-46ce-97ee-11924fde1491, initially expired at 2026-11-09T23:36:53.572Z. Its prearmed observable independent watchdog had an immutable deadline of 23:36:53.912Z. Exactly one delegated normal unforced stop completed at 23:14:47.392Z, receipt stop_6adaa7b037768ea8. An independent final API read confirms archived, archiveAfter null, no stop error. The independent watchdog was actually waited, exit 0, with separate stdout/stderr EOF and unchanged configuration. No owned box remains live.

Next work needs a separately authorized startup-service investigation before another causal mouse run. There is no demonstrated mouse cause to repair and no basis for relaxing a guard. Any later vehicle attempt must establish the vehicle from a fresh separately reviewed image; the older cabinet/ramp images are not vehicle evidence.