EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up surfaced by the Claude code-reviewer on task #825 (round base-separator-control r1).

Goal

Derive max+1 for real pod-side epm:results sentinel versions at drain time in poll_pipeline._post_drained_sentinel (or add a launch-time version-threading convention), so multi-round tasks stop landing results markers below the existing max.

Workflow gap

Bug observed: Real pod-side epm:results sentinels on multi-round tasks structurally land version: 1 below the existing max (5 such rows on #825 alone); the drain path threads explicit sentinel versions verbatim (poll_pipeline.py ~:2068), and pod-side dispatchers hardcode "version": 1 because they cannot read events.jsonl (the no-pod-side-task.py rule).
Why it is a workflow gap: the below-max collision class (#480) is re-created by every multi-round dispatcher; consumers currently survive by reading recency, but the highest-version-per-kind convention (markers.md) is silently violated for epm:results.
Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

(none — synthesized from prose follow-up. Options the planner should weigh: (a) drain-side: _post_drained_sentinel derives max+1 for REAL results sentinels when the sentinel's declared version is ≤ existing max (smoke/dryrun sentinels excluded); (b) launch-time convention: the orchestrator threads the next version into the dispatcher env and dispatchers template it. Update .claude/rules/pod-side-reporting.md alongside whichever lands.)

Scope / surfaces

Primary target: scripts/poll_pipeline.py
Companion: .claude/rules/pod-side-reporting.md

Constraints / invariants

Workflow-surface only. Never let a smoke/dryrun sentinel bump real marker versions; keep drain idempotency (a re-drained identical sentinel must not double-post).

Provenance

workflow_fix_target: scripts/poll_pipeline.py
fingerprint: d572e0ed2106

Verbatim surfaced prose (Claude code-reviewer, #825 base-separator-control r1): "scripts/poll_pipeline.py _post_drained_sentinel + .claude/rules/pod-side-reporting.md: real pod-side epm:results sentinels on multi-round tasks structurally land version: 1 below existing max (5 rows on #825 alone) — consider drain-side max+1 derivation for real results sentinels or a launch-time version-threading convention. confidence: medium."