EPS Dashboard

Goal

Wire the epm:step-completed marker emission at all 17 EXIT sites in .claude/skills/issue/SKILL.md, plus the regression test that enforces full coverage. Completes §5 of Plan A from PR #321.

Parent: #320

Background

PR #321 shipped (commit da6bf6bc):

epm:step-completed marker schema in markers.md
_decide_entry_step helper with the load-bearing status:blocked → full replay short-circuit BEFORE marker read (closes Critic 2.B2)
17-row EXIT-site → exit_kind mapping table in §5 of the plan body

What was NOT shipped: the actual EDIT to each of the 17 EXIT sites in SKILL.md to post the right marker before exiting. Without this, _decide_entry_step reads markers that never exist, and re-entry always hits the full-replay path. The infrastructure works; the call sites that USE it are missing.

Plan §5's mapping table (the 17 EXIT sites)

SKILL.md line	Step	exit_kind
457	Step 2c "Defer / 3" plan-approval	parked
524	Step 4 advance to implementing	clean
566	Step 5b code-review FAIL revision_round>=3 → blocked	failure-exit
596	Step 6a HF gate manual approval needed	parked
599	Step 6a HF gate `HF_TOKEN` missing → blocked	failure-exit
643	Step 6b pod provision failure → stay	parked
674	Step 6c preflight FAIL on resumed pod	parked
696	Step 7 `epm:stale` post (>4h no progress)	parked
781	Step 8 upload-verifier FAIL	parked
866	Step 9b reviewer PASS → awaiting-promotion	parked
879	Step 9b reviewer FAIL	parked
1225	Step 0 label conflict abort	failure-exit
879 (alt)	Step 9c test-verdict FAIL count<3	parked
879 (alt2)	Step 9c test-verdict FAIL count>=3 → blocked	failure-exit
921	Step 10 step 0 completion-audit INCOMPLETE → blocked	failure-exit
1066	Step 10c pod-termination prompt completion	clean
1116	Step 10d worktree-merge prompt completion	clean

(Spot-check the line numbers against current SKILL.md HEAD before editing — line numbers drift with each commit.)

Acceptance criteria

All 17 EXIT sites in SKILL.md post  before exiting (or, for failure-exit cases, alongside the epm:failure marker)
tests/test_every_exit_site_posts_marker passes — count parity: number of EXIT patterns in SKILL.md == number of epm:step-completed post-sites in SKILL.md == 17
Empirical "measurably skips replay" verification: pick a half-done test issue (e.g. parked at status:plan-pending or status:awaiting-promotion); show that re-invocation with epm:step-completed markers in place takes measurably fewer tokens than re-invocation with the markers stripped. Document the token-count delta in the epm:results body. Acceptance: at least one parked re-entry path saves >2k tokens.

Compute

0 GPU-hours. ~0.5-1 working day. type:infra.

References

Parent: #320 (Plan A approved)
PR shipped: #321 (commit da6bf6bc §5 helper + router; EXIT-site wiring is what this follow-up adds)
Plan body cached at .claude/plans/issue-320-draft.md §5 (lines ~1116-1331; 17-row mapping at ~1047-1069)
Code-reviewer concerns on §5: https://github.com/superkaiba/explore-persona-space/issues/320#issuecomment-4402015283

[Parent: #320] §5 epm:step-completed EXIT-site wiring + regression test + empirical replay-savings check