daily-fix: reconcile-stop misses error-storming sessions

Workflow gap

The autonomous-session watcher's parked/terminal reconcile-stop pass stops sessions whose mapped task is terminal only when they look IDLE (no activity for > ~2 h). On 2026-08-05, task #2004 was merged + completed at 12:05:26Z by a watcher-respawned sibling, while the original session (transcript 0d005b55) was mid autocompact-thrash. That session kept cycling failed turns (thrash/too-long API errors + continuation restarts — constant "activity", zero durable work) on the already-completed task until 16:00:50Z, only exiting at its own 16:20 tick reading TERMINAL. The reconcile pass DID stop one idle duplicate at 14:24 ("auto-stopped 1 idle session(s)") but not the error-storming one — ~4 h of pure token burn on a completed task.

verified-at-filing (2026-08-06T07:1xZ): task.py view 2004 --json events → merged 12:02:25Z, Step 10d auto-complete 12:05:26Z, watcher session-reconcile-stop of ONE idle session at 14:24:29Z. The 12:23→16:00Z thrash-row tail is the miner's probed scan of transcript 0d005b55 (rows carry only SKILL.md re-reads + thrash errors after 12:05Z).

Proposed change

Extend the reconcile-stop predicate in scripts/autonomous_session_watch.py: a session mapped to a parked/terminal task with NO durable-work signal (no non-watcher marker post, no commit, no live pod/follow-up signal) is stoppable even when its transcript shows recent rows — failed-turn API-error rows and compact/continuation churn are not "activity". Keep the existing shields (live follow-up signal markers, keep-running tag, RUNNING pod) unchanged; the change is only that transcript-row recency alone must not read as liveness when every recent row is an error/compaction artifact.

Provenance

• fingerprint: 89ed1b8ddbe9

• workflow_fix_target: scripts/autonomous_session_watch.py

• origin: /daily 2026-08-05 problem sweep — miner 7 P2 (transcript 0d005b55 + #2004 events timeline, probed).