Reap completed autonomous /issue sessions within ~1h (not 12h) — stop the ghost-session pile

Overview / Motivation

Autonomous /issue --auto sessions do not self-terminate when their task reaches a terminal/park state (completed / awaiting_promotion / archived / blocked). The watcher has a reconcile pass that auto-stops issue-MAPPED sessions whose task is terminal at >~2h idle — but when a task completes, the session→issue mapping is LOST, so the now-unmapped session falls out of the 2h terminal-reconcile and into the generic 12h idle-unmapped bucket. Result: completed sessions linger as "running" ghosts for up to 12 hours (~17 such ghosts observed 2026-06-28, idle 6–12h, all mapping to completed tasks; same class as the 2026-06-12 25-session VM-lag pile). Manual sweep was needed to clear them, and the user reads the pile as "the daemon is bugged" (it is not — daemon + watcher are healthy).

Goal

Completed-task autonomous sessions get reaped within ~1h, not ~12h, without relying on a manual sweep.

Scope / surfaces (pick the cleanest of these in planning)

1. Preserve last-mapped-issue on the session so the watcher's terminal reconcile (>2h, terminal task) still applies after a task completes — instead of the session dropping to mapped=False and into the 12h bucket. (scripts/spawn_session.py session registry + autonomous_session_watch.py reconcile pass.)
2. OR autonomous_session_watch.py: in the idle-unmapped pass, if a session's LAST-known mapped task is terminal, reap at a SHORT threshold (~30–60 min) rather than the generic 12h unmapped threshold.
3. OR /issue SKILL.md: an autonomous (EPM_AUTONOMOUS_SESSION=1) session self-exits when it reaches a terminal/park state (Step 10d completion / awaiting_promotion park) — the cleanest, removes the ghost at the source.

Option 3 (self-exit at terminal) is preferred if feasible; 1/2 are the watcher-side safety net.

Constraints / invariants

• Never stop a session with a live TTY (user-attached), a running pod, a keep-running tag, or a non-terminal task. The existing exclusions stay.
• A live same-issue follow-up round (followups_running / fresh epm:run-launched) must NOT be reaped.
• Workflow-surface only. scripts/workflow_lint.py + ruff pass.

Acceptance

• A completed autonomous /issue --auto session is gone (process exited / stopped) within ~1h of the task reaching terminal, with no manual sweep.
• A still-active session, a user-attached (TTY) session, and a live-follow-up session are all left untouched.

Provenance

PM-session investigation 2026-06-28: user asked "are they actually running? I think the happy daemon is bugged." Daemon + watcher were healthy; the pile was completed sessions not self-terminating + the mapping-loss dropping them from the 2h terminal-reconcile into the 12h unmapped bucket. Manual sweep cleared ~17 ghosts.