EPS Dashboard

Overview / Motivation

Auto-filed by the /daily Step C parked-candidate sweep (2026-07-28) from a formal candidate block parked on task #1750 (ts 2026-07-28T12:02:59Z, fp 51cea2984dae; surfaced by the plan critic, Methodology lens, during #1750's review). #1750 landed the prose duty (pod-side descope posts a marker + rewrites the handle); this filing is the mechanical detector for when that duty is skipped.

Goal

Add a WARN-only per-tick poller check that flags a dispatch handle sidecar (.claude/cache/issue-<N>-handle.json) whose sentinel_path / workload_pid / attempt id point at a DEAD prior attempt while a live relaunch is observed.

Workflow gap

Bug observed: a handle sidecar pointing at a dead prior attempt is consumed silently by the poller every tick — completion is never observed and nothing warns (incident #1689 r15b: the handle pointed at the old attempt's .completion-sentinel.json while the live relaunch wrote a new path).
Why it is a workflow gap: the pid-file contract has WARN-only detectors (the #1156 pid_file_stale_vs_marker / #1650 family); the handle sidecar has none, so a skipped handle-rewrite duty is invisible until someone audits by hand.
Confidence (emitter): medium
verified-at-filing: grep -n 'handle_stale_vs_live\|handle_older_than_relaunch' scripts/backend_poll.py → 0 hits (absence claim; the detector does not exist — 2026-07-29 UTC). Call-hop note (clause g): the named precedent detector pid_file_stale_vs_marker lives in scripts/poll_pipeline.py (lines 1507/5680/6027/6111), NOT backend_poll.py — the planner should confirm which poller actually consumes issue-<N>-handle.json (CLAUDE.md says the bg-Bash poller reads the handle back; both backend_poll.py and poll_pipeline.py are candidate sites) and place the detector at the consuming site; both sites are recorded here. Landed-fix history check: git log --oneline --since='7 days ago' -- scripts/backend_poll.py → no handle-staleness commit.

Proposed change (candidate diff sketch — refine in planning)

+ in the poller tick assembly (at the handle-consuming site):
+   if handle.extra.workload_pid dead AND observed live workload pid differs -> WARN handle_stale_vs_live
+   if newest epm:run-launched marker ts > handle sidecar mtime -> WARN handle_older_than_relaunch

Tick-JSON flag + WARN line, verdict unchanged (mirror of the #1156 pid_file_stale_vs_marker detector).

Scope / surfaces

Primary target: scripts/backend_poll.py (confirm consuming site; scripts/poll_pipeline.py holds the precedent detector and may be the correct or additional site)

Constraints / invariants

WARN-only — never changes the poll verdict.
Workflow-surface only; ruff passes; recursion guard applies to the spawned session.

Provenance

sha-verify (filing-time, #1467): 51cea2984dae cited in commit context does NOT resolve as a commit in this repo at filing time — treat as a transcript/session reference, not a commit.
workflow_fix_target: scripts/backend_poll.py
fingerprint: 51cea2984dae

target_file: scripts/backend_poll.py bug_observed: A dispatch handle sidecar (.claude/cache/issue--handle.json) whose sentinel_path/workload_pid point at a DEAD prior attempt is consumed silently by the poller every tick — completion is never observed and nothing warns (incident #1689 r15b: handle pointed at the old attempt's .completion-sentinel.json while the live relaunch wrote a new path). why_workflow_gap: pod-side-reporting.md gains a prose duty to rewrite the handle on descope relaunch (task #1750), but there is no mechanical detector when the duty is skipped — the pid-file contract has WARN-only detectors (#1156/#1650); the handle has none. proposed_change: Add a WARN-only per-tick check in the poller comparing the handle's extra.expected_artifacts.sentinel_path / extra.workload_pid / extra.runpod_attempt_id against the live run's observed attempt/pid; set a tick-JSON flag (e.g. handle_stale_vs_live) + a WARN line, verdict unchanged (mirror of the #1156 pid_file_stale_vs_marker detector). confidence: medium related_task: #1750