address-feedback: extract run-state persistence and merge-gate checklist scripts

[cbeaulieu-gt]

Summary

address-feedback/SKILL.md describes two mechanical, non-judgment
components entirely in prose: run-state persistence (read/write a JSON
file, enforce caps) and the merge-gate checklist (a chain of gh api
boolean checks). Both are candidates for extraction into scripts.

Context

Part of #33 (milestone: Deterministic script extraction). Depends loosely
on the gh-pr-review-address script work (#33) since this skill
composes that one's Step 2/2.5/3/4 unchanged — do this one second.

Candidates

1. Run-state persistence (§ Run-state persistence, .tmp/address-feedback-pr<N>.json).
   Load/validate/write the state file, increment tick_count, compare
   run_id/owner_repo/target for continuation-vs-new-run, enforce
   tick_count > MAX_TICKS / act_rounds >= MAX_ACT_ROUNDS caps. This is
   plain file I/O and comparison logic — a scripts/address-feedback-state.py
   with load/save/check-caps subcommands removes an entire class of
   "did the LLM read the JSON correctly" risk.
2. Merge-gate checklist (§ Safety guards → Merge gates). Six
   independent boolean checks against gh api/gh pr view output
   (PR still open, effective permission, CI green on the merge commit
   not just HEAD, no CHANGES_REQUESTED, mergeStateStatus in allowed
   set, no DO NOT MERGE banner, merge-queue enqueue-vs-merged
   distinction). None of these require judgment — a script that emits
   {gate: name, passed: bool, detail: str} per gate, plus an overall
   all_passed: bool, replaces several paragraphs of "check X, check Y,
   AND them together" prose with one call.

Technical Notes

• The WAIT / ACT / EXIT decision itself, the quiescence determination,
  and the judgment-call-park override stay on the LLM — those genuinely
  read bot review content and prior state to decide what to do next.
• ScheduleWakeup calls stay LLM-issued (this is a /loop dynamic-mode
  skill, not compatible with a script owning the wakeup call).
• Reuse _gh_common.py where the merge-gate script needs gh api fetches
  identical to ones already fetched in gh-pr-review-address.

Out of Scope

• Any change to the loop-driver control flow (Tick steps 0–4) — those
  stay LLM-orchestrated.
• The autonomy-override park logic (dup-check + issue creation) — that's
  judgment (is this a duplicate?) even though the mechanics of creating
  the issue are already scripted elsewhere (gh-create-issue skill).

Acceptance Criteria

• Run-state load/save/cap-check logic is script-driven, not
  prose-described JSON manipulation
• Merge-gate checklist is a single script call emitting per-gate
  pass/fail + detail
• SKILL.md updated to call the script(s) instead of describing the
  checks inline
• Test coverage under scripts/tests/

🤖 Generated by Claude Code on behalf of @cbeaulieu-gt