worktree_audit.py stale-sweep reaps an active inline-round worktree (pod-side work + parked parent look idle)

Surfaced by: the #2223 intervention-strength-sweep inline-override round (interactive chat), 2026-08-17.

Gap: the daily stale-worktree sweep (scripts/worktree_audit.py --apply, run via cron_worktree_audit) reaped .claude/worktrees/issue-2223-casestudy WHILE an active inline round owned it. The round looked idle to the audit's liveness heuristic because: (a) the round's compute ran on pod-2223-strsweep (no VM-side process to detect); (b) the parent task #2223 sat at awaiting_promotion (a terminal-ish status, not an ACTIVE lifecycle status); (c) the worktree tree was momentarily clean (work committed to the branch + pushed). So the "live-process / active-issue / dirty kept" gates all read false and the worktree was removed mid-round. No data was lost (the 72 cells were safe on the pod + committed to origin/issue-2223-casestudy), but the round had to RECREATE the worktree mid-deliverable — and the manual recreation used a custom sparse cone that omitted .claude/, which then crashed the workflow-lint-phase-done-reserved pre-commit hook (workflow_lint.py reads .claude/config/agent_spec_size_caps.txt at import → FileNotFoundError) on the next commit, costing a second recovery.

Why this is a standing gap, not a one-off: an inline-override / same-issue follow-up round routinely (i) runs its compute on a pod (no VM process), (ii) operates on a parent at a parked/terminal status (awaiting_promotion, on_hold, completed), and (iii) keeps its worktree tree clean between commits. That triple is exactly the shape the audit's idle heuristic treats as reapable — so ANY active inline round on a parked parent is exposed to having its worktree reaped out from under it.

Proposed fix: worktree_audit.py should treat a worktree as ACTIVE (skip reap) when ANY of:

• the parent task has a fresh follow-up signal marker (epm:run-launched / epm:followup-scope / epm:free-analysis-followup-run) NEWER than the latest done-transition, OR the keep-running tag is set (the SAME predicate the watcher's pod-safety pass already uses to shield pods — reuse it for worktrees);
• the worktree branch has commits not on origin/<branch> OR uncommitted/untracked round work (git -C <wt> status --porcelain), even when the tree is momentarily clean between commits, prefer a short activity grace window keyed on last branch-commit mtime. Human-named worktrees are already never touched; this only widens the "keep" set for auto-generated issue worktrees whose round is provably in flight.

Confidence: high (reproduced live: the worktree was reaped and had to be recreated by hand this round).

Target file: scripts/worktree_audit.py (+ the shared follow-up-signal predicate in scripts/autonomous_session_watch.py / tick_triage.py it should reuse).

Regression pin to add: a fixture where a worktree's parent task carries a fresh epm:run-launched newer than its latest done-transition (or keep-running) → worktree_audit KEEPS it even when the tree is clean and no VM process matches.