workflow-fix: route user 'pause <N>' to on_hold (durable, watcher-safe park)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #816 (emitting agent: /issue orchestrator, watcher-respawned session honoring a user pause).

Goal

Add a "user pause" routing rule to the workflow surface: on a user "pause " directive, the driving session parks the task at on_hold (already in the watcher's PARK set — durable, watcher-safe) with a resume-recipe note, instead of leaving status at an ACTIVE value with a prose-only hold marker the watcher cannot parse.

Workflow gap

• Bug observed: User said "pause 816" (2026-07-02). The driving session recorded the hold as a prose epm:progress marker (v72: "DO NOT resume, respawn, or auto-dispatch ... watcher: do not crash-recover") and left the task at status=running, then stopped. ~2h later the watcher's orphan-respawn pass — which keys on "active status + no live registered session + no fresh markers" and cannot parse prose holds — respawned an autonomous /issue 816 --auto session AGAINST the explicit user hold (consuming respawn attempt 1/2 for the day and a fresh session's tokens). The respawned session had to detect the pause note itself and park the task at on_hold by hand.
• Why it is a workflow gap: No workflow surface prescribes how a session should implement a user "pause ". The correct durable mechanism already exists — on_hold is in the watcher's PARK set (autonomous_session_watch.py line ~478) and the status enum documents it as "explicitly set aside ... excluded from auto-dispatch" — but nothing routes the "pause" verb to it, so sessions improvise prose markers that every automated pass is blind to. (#903's paused-takeover sentinel does NOT cover this: it is a short-TTL (default 6h, FAIL OPEN) session-takeover shield, not an indefinite user hold — even a sentinel-writing pause would be respawned after 6h.)
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

+ CLAUDE.md (§ Routing experiment intent, next to the abort affordance) and/or
+ .claude/skills/issue/SKILL.md (§ abort affordance):
+   **User pause** ("pause <N>", "put <N> on hold"): implement the hold
+   DURABLY, not as prose —
+     1. uv run python scripts/task.py set-status <N> on_hold \
+          --note "USER PAUSE <verbatim>; resume: set-status <N> running +
+                  spawn_session.py spawn-issue --issue <N> --auto"
+        (on_hold is in the watcher PARK set: no orphan-respawn, no
+         auto-dispatch; markers re-enter the /issue loop at the same round
+         on resume). If a same-issue follow-up round holds
+         `followups_running`, note the status-hold interaction explicitly.
+     2. Stop/park the driving session; stop any pod per Step 8-bis.
+   **Resume** (user greenlight only): set-status <N> running, then
+   spawn-issue --issue <N> --auto.
+ .claude/rules/background-automation.md: one line noting on_hold is the
+   user-pause mechanism (PARK set) vs the #903 paused-takeover sentinel
+   (short-TTL takeover shield) — the two are not interchangeable.

Scope / surfaces

• Primary target: CLAUDE.md, .claude/skills/issue/SKILL.md, .claude/rules/background-automation.md
• Grep the workflow surface for the pattern before editing (grep -rln 'pause' .claude/ CLAUDE.md scripts/spawn_session.py) and update every relevant hit; list them in the plan. Do NOT touch the #903 paused-takeover sentinel logic (different mechanism, working as designed).

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; if CLAUDE.md changes it stays consistent with the rule files.
• The rule must be documentation/routing only in v1 — no new marker kinds, no status-enum change (on_hold already exists), no watcher code change needed.
• 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: CLAUDE.md, .claude/skills/issue/SKILL.md, .claude/rules/background-automation.md
• fingerprint: be4ec04c96eb

Origin: task #816 user pause (2026-07-02 "pause 816", epm:progress v72) vs watcher orphan-respawn (epm:progress v73, 2026-07-03T08:33:29Z, attempt 1/2). The respawned session detected the conflict and parked #816 at on_hold manually (epm:status-changed running -> on_hold, 2026-07-03).