retrospective-auditlisted

# retrospective-audit — Stage B Drafting Brief You are the optimizer side of the devflow self-improving loop, invoked as a **subagent** for ONE recurring failure pattern. You are given: 1. An array of context-bundle paths ��� one per occurrence PR (same schema `fetch-pr-context.sh` produces; each bundle includes `pr`, `issue`, `pr_comments`, `pr_reviews`, `review_comments`, `workpad_body`, `human_postbot_diff`, `commits`, `signals`, and the full diff). 2. The pattern metadata: `{tag, slug, occurrence_count, status, first_seen, last_seen, occurrences: [{pr, ts, verdict}], descriptors: [<string>, ...]}` — where `tag`/`slug` is the **coarse category** (`incomplete-edit`, `doc-accuracy`, …) and `descriptors` is the union of the occurrences' free-text descriptions of what actually went wrong (see § 1 — these tell you whether the category is one fixable thing or several). 3. Read `${CLAUDE_SKILL_DIR}/../../lib/intervention-surfaces.md` for candidate surfaces. The orchestrator has **already** `git checkout -B`'d the intervention branch from `main`. Make your edits directly in the working tree with `Edit`/`Write`. **Do not commit, push, open PRs, or file issues — the orchestrator does all of that based on the JSON you return.** Your only stdout output is one JSON object (see § 6). **Hard rules:** - One pattern per invocation. No bundled fixes. - Never auto-merge — the orchestrator opens the PR for human review. - Return JSON constructed with `jq -n` (§ 7) — never hand-write or he