workflow-fix: triage window boundary opens at post time, hiding enumerate-to-post seam markers

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #2054 (emitting agent: issue-orchestrator). Promised in #2054's triage-correction note (epm:progress v108); root-cause diagnosis in #2054's v98 marker.

Goal

Record the enumeration-boundary timestamp in the external-markers triaged: line so the next triage window opens from the enumeration point, closing the enumerate-to-post seam.

Workflow gap

• Bug observed: markers landing between the triage enumerator run and the breadcrumb post fall behind the new boundary and are permanently invisible to the pre-dispatch triage duty
• Why it is a workflow gap: task_workflow.triage_candidates_since_last_dispatch opens the candidate window strictly AFTER the most recent event carrying the triage line (or a launch-kind marker), but the triage line is posted at breadcrumb-post time T1 while enumeration happened at T0 < T1 — any external marker landing in (T0, T1) was not enumerated at T0 and is behind the boundary at every later call, so no session ever triages it. Incident (#2054, 2026-08-05): the user-directive marker v91 ("make sure training is at least 5000 for each setting") landed 53 s before the r11 breadcrumb post and was invisible to rounds r11–r14; caught only by a manual events re-read (v98 forensics), costing a 4-round directive miss and a triage-correction round (v108).
• Confidence (emitter): high
• verified-at-filing: grep -n "TRIAGE_LINE_PREFIX" src/explore_persona_space/task_workflow.py → 9 hits in 1 file (def at :2762; the boundary matcher at :2839-2846 — boundary = idx on kind in launch_kinds or TRIAGE_LINE_PREFIX in note, window = events strictly after boundary; sibling pre-record window logic at :2981, :3061-3088); plus grep -rn "external-markers triaged" .claude/skills/issue/SKILL.md → 5 hits (triage-line format spec at :6948-6974) (2026-08-06). The seam is structural in the boundary matcher: no token in the triage line carries the enumeration point, so the window boundary can only be the post position.

Proposed change (candidate diff sketch — refine in planning)

# task_workflow.py — triage-line format gains an optional boundary token:
#   "external-markers triaged: <N> applied / <M> deferred (boundary=<ts of last enumerated event>)"
# triage_candidates_since_last_dispatch:
- if event.get("kind", "") in launch_kinds or TRIAGE_LINE_PREFIX in note:
-     boundary = idx
-     break
+ if event.get("kind", "") in launch_kinds or TRIAGE_LINE_PREFIX in note:
+     boundary = idx
+     recorded = _parse_triage_boundary_ts(note)   # None on legacy lines
+     if recorded is not None:
+         boundary = index of last event with ts <= recorded  # reopen the seam
+     break
# SKILL.md triage-line spec (~L6972): document the boundary=<ts> token; the
# enumerating session stamps the ts of the LAST event it enumerated (or the
# enumeration wall-clock) into its triage line at post time.
# Legacy lines without the token keep today's behavior (fail-toward-today, never wider misses).

Scope / surfaces

• Primary target: src/explore_persona_space/task_workflow.py
• Secondary: .claude/skills/issue/SKILL.md (triage-line format spec + the duty text that composes the line)
• Grep the workflow surface for the pattern before editing (grep -rn 'external-markers triaged' .claude/ CLAUDE.md scripts/ src/explore_persona_space/task_workflow.py) and update every composer/parser hit; list them in the plan. Tests: extend the existing triage-window tests (grep triage_candidates_since_last_dispatch under tests/).

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
• Backward compatibility is load-bearing: thousands of legacy triage lines exist across events.jsonl files; a legacy line (no boundary token) MUST parse to today's behavior exactly. The window must never SHRINK vs today (fail-toward-triage: over-enumeration is safe, under-enumeration is the bug).
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: src/explore_persona_space/task_workflow.py
• fingerprint: dfb17c925991

target_file: src/explore_persona_space/task_workflow.py, .claude/skills/issue/SKILL.md bug_observed: markers landing between the triage enumerator run and the breadcrumb post fall behind the new boundary and are permanently invisible to the pre-dispatch triage duty why_workflow_gap: the window boundary in triage_candidates_since_last_dispatch is the triage-line POST position, not the ENUMERATION position, so the enumerate-to-post seam (53 s in the #2054 v91 incident) permanently hides markers proposed_change: record the enumeration-boundary timestamp in the triage line so the next triage window opens from the enumeration point, closing the enumerate-to-post seam diff_sketch: |

• triage line gains "(boundary=)" token stamped by the enumerating session
• triage_candidates_since_last_dispatch parses the token and reopens the window
• from the recorded enumeration point; legacy lines keep today's behavior confidence: high related_task: #2054