workflow-fix: pod-safety alert for unlaunched orphan pods on active tasks

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1481 (emitting agent: /issue orchestrator).

Goal

The watcher's pod-safety pass flags a RUNNING managed pod on an ACTIVE-status task with no epm:run-launched newer than pod creation after a grace window (alert-only).

Workflow gap

• Bug observed: a detached wait-for-capacity provision outlived its requesting turn and delivered an unlaunched orphan 8xH100 pod that billed $32/hr for ~2h, invisible to every watcher pass.
• Why it is a workflow gap: pod_lifecycle provision deliberately detaches (setsid, #573) and can deliver a pod AFTER the requesting orchestrator turn is gone; nothing then claims or launches on the pod. The watcher's pod-safety pass auto-stops/alerts only on DONE/parked tasks — a RUNNING pod on an ACTIVE task (#1481 at running) is presumed healthy even with zero workload ever launched (no epm:run-launched since creation, 0% GPU). On 2026-07-18 pod-1481 idled ~2h ≈ $64 and was found only by a manual burn probe; it also blocked the 17:52Z RunPod fallback provision (name conflict).
• Confidence (emitter): medium
• verified-at-filing: grep -n "setsid" scripts/pod_lifecycle.py → provision detaches into its own session (the "#573" detach banner printed in the 17:52Z stderr tail); grep -n "POD_SAFETY_AUTO_STOP\|pod_safety" scripts/autonomous_session_watch.py → the pass keys on done/parked statuses only — no unlaunched-orphan predicate for ACTIVE-status tasks (2026-07-18)

Proposed change (candidate diff sketch — refine in planning)

+ # pod-safety: unlaunched-orphan flag (alert-only, never auto-stop):
+ # a RUNNING managed pod whose owning task is at an ACTIVE status but has
+ # NO epm:run-launched marker newer than the pod's created_at, after a
+ # >=60-min grace (covers provision/bootstrap + experimenter dispatch),
+ # fires a deduped Telegram alert naming the pod + $/hr + the reuse-or-
+ # terminate commands. Never stops (a launch may be seconds away).

Scope / surfaces

• Primary target: scripts/autonomous_session_watch.py
• Grep the workflow surface for the pattern before editing and update every hit; list them in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• Alert-only for ACTIVE-status tasks — the pass must NEVER auto-stop a pod on an active task (the existing conservative contract stands).
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: scripts/autonomous_session_watch.py
• fingerprint: dc6de410ab93

Surfaced prose (verbatim): a detached wait-for-capacity provision outlived its requesting turn and delivered an unlaunched orphan 8xH100 pod that billed $32/hr for ~2h invisible to every watcher pass (pod-safety only covers done/parked tasks).