sweep discards a validating result.json unread on any non-completed status; zero-hook-event runs pass run/sweep untouched

[wsquared]

Observed (0.10.0)

In sweep.py's triage handling:

if result.status != "completed":
    plan, errors = None, [session_failure_reason("triage", result)]
else:
    plan, errors = validate_triage(result.result_json, open_now)

When the session status is anything but completed, result.result_json is never read. After max_triage_attempts the run escalates with "triage output failed validation: ..." — blaming output that was never validated. The migration path has the same shape (validate_migration runs only on completed). #194's env_fault pause softens the transport-failure case only; a lost event channel is not classified as an env fault.

Incident

Run 20260816-113627-d0b2 (macOS, claude adapter). The hook relay had never been registered in that project, so every session's Stop event was lost and each session read as timeout with session_id: null. Both triage attempts wrote a result.json that passes validate_triage (strict and cache mode alike) — attempt 2 finished its actual work in 59 seconds — and both were discarded on session status alone. The escalation blamed the output; the output was fine. signals.py's own docstring predicts the failure mode: losing Stop events means every session stalls to session_timeout_min — "the loudest possible regression, delivered silently".

Proposals (either or both)

1. Validate the result artifact on non-completed status too. If it validates, either use it — a timeout with a valid, complete artifact is a completed turn whose Stop event was lost — or at minimum attach "result artifact present and passes validation" to the escalation so the operator debugs the event channel rather than the agent. SessionResult.stop_seen (Session read-back adopts another story's spec: a review that produced nothing is scored done #261) is already the right discriminator: status == "timeout" and not stop_seen with a validating artifact is the lost-event-channel signature. Routing it like Dev session lost to an API/transport failure is charged as a story timeout, exhausting the attempt budget #194's env_fault pause (pause, don't charge attempts) would fit the existing shape.

2. Make run/sweep enforce the zero-hook-events condition. cmd_validate fails on unregistered, missing, unreadable, or stale relays, but nothing stops a run whose sessions produce zero events: it burns session_timeout_min per session and escalates with the wrong blame. Hard-failing (or pausing) after the first session that ends with stop_seen == False and an empty run events dir would surface the miswiring at session one instead of N timeouts later.

Happy to attach the journal/state.json from the incident run.