feat(skills): add /issue_cleanup — issue-tracker reconciliation door

[Jammy2211]

Closes #174.

Adds /issue_cleanup, the missing door for reconciling the GitHub issue trackers — the issue-tracker counterpart to /repo_cleanup's git-debris sweep. Same proven shape: audit → bucketed dashboard → per-bucket human confirmation → execute → recap.

Why

Nothing owned this. /repo_cleanup sweeps branches, refs, stashes and worktrees and stops there; /community handles external users' issues awaiting a reply; /create_issue and /update_issue are single-issue primitives. The trackers had drifted to 82 open issues across 18 repos, of which 35 were closable. An ad-hoc sweep on 2026-07-28 cleared those (29 shipped-but-still-open + 6 obsolete 2018–2019 PyAutoCTI), leaving 47.

The reasoning is the deliverable

A naive "a complete/ record references this issue, so close it" rule is wrong five ways, each found by verification rather than inspection:

1. Body mentions are not claims — only header lines are evidence, and only for a known set of keys. Records use - notes: for long prose that cites issue URLs freely, so "any - word: line" re-admits exactly the prose the rule exists to exclude.
2. The header key carries the meaning — issue: (630 uses) completes; followup-issue: / library-followup-issue: / parent-issue: / upstream-issues-filed: / plan: (13 uses) mean the record spawned a still-open issue. Completing keys are an allowlist, so a new spawn-style key fails closed.
3. Annotations and status override — (open — …), (STAYS OPEN — …), and Status: issued. A record can sit in complete/ having merely filed its issue (ep-hierarchical-scale-collapse.md → PyAutoFit#1405, a live bug).
4. A phase-scoped claim does not complete an umbrella — four records each claim a Phase 5 item of PyAutoBrain#130; none establishes it is done.
5. Age is not evidence of obsolescence — probe whether the named API still exists. Six PyAutoCTI issues (2693–2785 days) closed on a zero-hit grep of autocti/; PyAutoHands#16/#17 are the same ~1337-day vintage but stayed open as still-valid unimplemented asks.

Closing requires two independent evidence legs (record header + merged PR) and always a human confirmation. The PR leg checks both the GitHub timeline and a PR named in the record body — timeline alone under-reports (12 of the 29 had only the latter).

Regression bar

Rules 1 and 4 were caught by the skill's own bar, not by inspection — which is why the bar ships with it. It pins the post-sweep state (47 open · A=0 · B=1 · C=7 · spawn-held=3 · D=6 · E=8 · F=22) and names the two traps explicitly: PyAutoGalaxy#417 must land in the backlog rather than a held bucket (or the header allowlist has regressed), and PyAutoBrain#130 must land in B rather than A (or umbrellas are being closed on partial evidence). Verified passing at the tip of this branch.

Also in this PR

• COMMANDS.md gains a maintenance doors tier documenting /repo_cleanup and /issue_cleanup as complements — neither was previously listed there.
• /wake_up gains a read-only issue-drift step, so drift surfaces daily instead of at 82-issue depth. The audit half is safe unattended; every close stays human-gated.

Notes for review

• Docs/skill-only change — no source, no tests to run. The behavioural check is the regression bar above.
• gh issue close is broken in this environment; the skill uses the comment + gh api -X PATCH pair, with state_reason completed vs not_planned.
• The two ~/.claude symlinks are intentionally not created here — bin/install.sh run from a worktree repoints ~/.claude at that worktree. Run it from the canonical checkout after merge.

🤖 Generated with Claude Code