workflow: promote bug-class sibling-sweep from reconciler memory into primary code-reviewer + implementer (stop review whack-a-mole)

Overview / Motivation

Autonomous /issue code-review loops exhibit a "whack-a-mole": the experiment-implementer fixes the flagged instance, and each subsequent code-review round surfaces a NEW real bug — usually a SIBLING in the same subsystem the previous round just touched — so the loop burns many rounds (or hits the cap and strategy-pivots) without converging. Instance: #779 produced ≥6 real blockers clustered in the raw-completions I/O subsystem, plus r_B extraction and the primary metric, one-per-round.

Root cause (diagnosed): the sibling-scan / bug-class-sweep discipline (enumerate EVERY instance of a bug class per round — a 7-step recipe backed by 8 incidents: #498/#601/#505/ #514/#489/#407) lives ONLY in .claude/agent-memory/reconciler/feedback_claude_misses_same_file_siblings.md. The reconciler fires only when the Claude and Codex reviewers ALREADY disagree (PASS-vs-FAIL). The primary code-reviewer loads code-reviewer/* memory, NOT reconciler/*, and has only per-instance gates — no always-on step requiring it to sweep the subsystem for siblings. So on any finding it reports the one cited instance and the siblings surface in later rounds. Compounding it, experiment-implementer.md's revision-round rule says "targeted edits — do NOT rewrite unrelated code," which actively discourages fixing the class.

This is ORTHOGONAL to #784 (which changes the round-cap value + strip logic). This task is about SCOPING each round to the bug CLASS, not the cap.

Goal

Make the code-review loop surface AND fix bug CLASSES per round (all sibling instances in one round) instead of one instance at a time — promoting the reconciler's battle-tested sibling-scan recipe into the primary reviewer + implementer protocols.

Change set (interdependent — MUST land together; a partial merge leaves the workflow self-contradictory)

1. .claude/agents/code-reviewer.md — new mandatory "Step 3.7: Bug-class sibling sweep." For EVERY Critical/Major finding, the cited line is an INSTANCE; the contract is the bug CLASS. Before verdict, grep the whole file + the sibling function/rubric/resampler family + sibling scripts sharing the finding's data contract + parallel layers (figure-vs-analyze) for the same pattern (e.g. parsed.get("score",0), per-persona vs global custom_id index, raw_completions write-without-upload, except Exception, sentinel reads). Report ALL sibling hits under one "### Bug-class sweep: " heading with file:LINE each; a load-bearing sibling is its own Critical, a secondary one a standing rec. Add an enforcing Rule 14: "Every finding is a bug CLASS, not a line. FAIL enumerating all load-bearing siblings; a single-instance fix that leaves siblings is the whack-a-mole failure mode." (Promotes the 7-step recipe from reconciler memory feedback_claude_misses_same_file_siblings.md.)

2. .claude/agents/experiment-implementer.md — revision-round class-hardening carve-out. The current "Make targeted edits — do NOT rewrite unrelated code on a revision round" rule contradicts the class-sweep. Add a carve-out: when a FAIL item names a bug CLASS (or the reviewer's Step 3.7 enumerated siblings), fix EVERY named sibling instance this round, and before returning run a one-line self-sweep grep of the just-fixed pattern across the touched subsystem, reporting it under (c) How to verify. (Implementer-side mirror of the reviewer step — required to resolve the contradiction above.)

3. .claude/skills/issue/SKILL.md (Step 5c / round-N+1 brief composition) — carry the sibling family into the punch list. When a reviewer verdict contains a "### Bug-class sweep" enumeration, thread the FULL sibling list (not just the top finding) into the round-N+1 implementer brief so its punch list is class-scoped.

4. .claude/agent-memory/reconciler/feedback_claude_misses_same_file_siblings.md — one-line pointer noting the recipe is now promoted to the primary code-reviewer.md Step 3.7 (keeps the memory as the reconciler's fallback and in sync with the enforced step).

Scope / surfaces to check

• Grep the reviewer-rubric surface so the sweep step propagates to BOTH twins: the Codex twin .claude/agents/codex-code-reviewer.md inlines "the same rubric the Claude reviewer uses" — if it inlines the rubric verbatim rather than referencing code-reviewer.md, add Step 3.7 there too (verify by reading how the codex twin composes its prompt).
• Check .claude/workflow.yaml for any review-step enumeration that must list the new Step 3.7.

Constraints / invariants

• Workflow-surface only. Grep before editing; keep code-reviewer.md ↔ codex-code-reviewer.md ↔ SKILL.md consistent.
• Orthogonal to #784 — do NOT touch the round-cap value or the strip logic (that's #784's task).
• scripts/workflow_lint.py --check-asks passes; ruff on any touched scripts passes.
• Not architectural (adds a review step + directive; no enum/schema/subcommand rename) → auto-approve
  • Step-10d auto-merge per the workflow-fix default.

Provenance

Diagnosed by a workflow-improvement subagent commissioned from the #779 whack-a-mole. The four changes are the subagent's single high-confidence candidate (change 1) plus its three prose follow-ups (changes 2-4), consolidated into ONE task because they are interdependent (the reviewer sweep and the implementer carve-out contradict each other unless both land).