feat(loop): auto-cleanup worker worktree + branch after merge (#91)

[robercano-ghbot]

Summary

Closes #91.

Worker worktrees + their local branches accumulated forever after a PR merged — git worktree list had many stale prunable entries, several pointing at a pre-rename path that no longer exists on disk. This adds auto-cleanup right after merge-ready.sh merges a PR.

• merge-ready.sh: pulls headRefName from the PR JSON, and after a successful merge shells out to the new worktree-cleanup.sh helper, tagging its JSON-lines output with the PR number.
• worktree-cleanup.sh (new): <base-branch> [<branch>...] helper, factored out so it's independently unit-testable offline. Safety rails — ALL required, or the branch is skipped with a logged reason and left alone:
  • worktree path matches the worker naming convention (.claude/worktrees/agent-* or .claude/worktrees/issue-*) — never the main worktree or anything else
  • the worktree tree is clean (git status --porcelain empty)
  • the branch is fully merged into base (git branch --merged <base>)
  • never --force; never git branch -D — a refusal from git worktree remove / git branch -d (dirty/locked/unmerged) is logged as a skip, not overridden
• worktree-cleanup.test.sh (new): offline smoke test against a throwaway temp git repo with real git worktree add worktrees (no network, no gh). Covers: merged+clean+matching → removed; dirty → preserved; unmerged → preserved; non-matching path (main worktree, and an arbitrary dir name) → never touched; no-worktree branch → no crash. Picked up automatically by .claude/self/checks.sh's *.test.sh glob.

One-off migration (issue #91 also asked for this, run manually, not part of this diff): ran git worktree prune -v on the reCode repo to clear dangling registrations under the old pre-rename ai-project-orchestrator/.claude/worktrees/* path (directories no longer exist on disk). No in-flight reCode workers were touched.

Cockpit bonus (issue #91, optional): skipped — renderWorktrees() in cockpit.sh currently sources its data from a plain readdirSync of .claude/worktrees, not git worktree list --porcelain, so flagging prunable entries would mean switching its whole data model (and touching cockpit.test.sh's assertions), not a trivial addition.

Test plan

• GATES_FILE=.claude/self/gates.json bash .claude/scripts/gate.sh build — pass
• GATES_FILE=.claude/self/gates.json bash .claude/scripts/gate.sh lint — pass
• GATES_FILE=.claude/self/gates.json bash .claude/scripts/gate.sh test — pass (includes new worktree-cleanup.test.sh, 22/22 checks, plus smoke-fanout.sh)
• bash .claude/scripts/worktree-cleanup.test.sh runnable standalone

🤖 Generated with Claude Code

[robercano-ghbot]

Fixed the merged-ness ordering bug flagged by review:

worktree-cleanup.sh now fetches origin/$base and checks ancestry against it (falling back to the old local git branch --merged check if origin is unreachable), instead of relying solely on local $base — which isn't fast-forwarded until merge-ready.sh's post-loop local_sync block, i.e. AFTER cleanup runs. git branch -d is still tried first; if it refuses because local $base hasn't caught up, a direct ref delete is used, but only once merge-base --is-ancestor has independently verified the merge against origin/$base — never -D/--force.

Added scenario 6 to worktree-cleanup.test.sh, which reproduces the exact production ordering (merge pushed to a real bare "origin" from a separate clone, local main deliberately left stale) — confirmed it fails against the pre-fix script and passes after. All 29 checks pass (was 23).

Gates: build/lint/test all green (GATES_FILE=.claude/self/gates.json bash .claude/scripts/gate.sh build|lint|test).

Commit: 933f202