feat(code-quality-plugin): add parallel-sweep rename map and auditor briefing to bulk-sweep-classify

[laurigates]

What

Extends the existing code-quality-plugin:bulk-sweep-classify skill with two sections. Additive — no existing content was rewritten, and the skill's own regression test (scripts/tests/test-bulk-sweep-classify.sh, 11 checks) still passes.

1. Parallelizing a Sweep — Resolve the Rename Map Before Dispatch

Once a sweep fans out across parallel agents, a failure mode appears that the four-category discipline alone doesn't cover: the agents disagree with each other about borderline renames. Each decision is locally defensible, and the result is a codebase that compiles file-by-file but not as a whole.

The fix is to resolve every cross-file naming decision before dispatch into a single explicit rename map (old → new, plus an explicit do-NOT-rename list drawn from the skill's category 2–4 buckets), and brief that map verbatim into every agent's prompt. That moves agreement out of the agents and into a shared artifact — so file grouping then only has to guarantee disjointness, a much weaker and easier property than getting independent agents to converge.

The section also names the success signal, because it's counter-intuitive: an agent refusing a rename is the contract working. A run where nobody declines anything usually means the map wasn't actually constraining them.

2. Brief Adversarial Auditors With the Artifact's Purpose, Not Just the Transform

An auditor given only the transform contract ("rename X to Y across these files") has no way to distinguish a deliberate change from scope creep — so it flags the change's own reason for existing. Every auditor prompt needs both the transform contract and a statement of what the artifact is FOR and what is deliberately in scope.

Why it's reusable beyond one repo

Neither section is language-, framework-, or repo-specific. The first is a property of fanning any cross-file decision out to independent agents: consensus is expensive to obtain from agents and free to obtain from a pre-resolved artifact. The second is a property of any adversarial review whose brief omits intent — the auditor's only available frame becomes "is this in the diff contract?", so purpose-driven changes read as creep.

Both sit naturally in this skill because it already owns the "the regex sees text, the transform needs semantics" discipline; the rename map is just that classification resolved once and shared, rather than re-derived per agent.

Where it was used

A 112-file Trends → Foresight rename in ForumViriumHelsinki/thelma (PR #1263), which also renamed a Postgres enum TrendType → ForesightType. The skill's existing four-category discipline is what kept that sweep from corrupting an enum value; these two things it did not yet cover.

Evidence cited in the sections:

Observation	What it demonstrates
112 files across parallel agents, one central rename map	Disjoint file grouping was sufficient once agreement was pre-resolved
Agents declined to rename LinkableEntityType — contains the target word, not in the map	The do-NOT-rename list held against an agent "helpfully" extending scope
All three adversarial auditors flagged the PR's central purpose as scope creep; 10 of 25 findings were false positives from that one omission	Cost of briefing auditors with the transform contract alone

Also in this PR

• Cross-links added to Related: agent-patterns-plugin:parallel-agent-dispatch (dispatch mechanics — the rename map is the sweep-specific payload it carries) and agent-patterns-plugin:adversarial-review (the review pass itself).
• code-quality-plugin/README.md catalog line extended to mention the new guidance (the pre-commit README-currency nudge fired; addressed in the same commit).
• Frontmatter modified: bumped to 2026-08-08, matching the repo's convention.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DvcEm3GrLgboEaKA24evDG

[github-actions]

Plugin Compliance Review

Plugin	plugin.json	Frontmatter	Body	Marketplace	Release Config	Bash Patterns	Descriptions	When-to-Use	Size	Overall
code-quality-plugin	✅	✅	✅	✅	✅	✅	⚠️	✅	⚠️	⚠️

Recommendations

• ⚠️ code-quality-plugin/bulk-sweep-classify: description is 236 chars — over the 200-char target band (see .claude/rules/skill-quality.md)
• ⚠️ code-quality-plugin/bulk-sweep-classify: SKILL.md is 11901 chars (~2975 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ code-quality-plugin/code-docs-quality: SKILL.md is 13061 chars (~3265 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ code-quality-plugin/code-hidden-failures: SKILL.md is 11901 chars (~2975 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ code-quality-plugin/dry-consolidation: SKILL.md is 11848 chars (~2962 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)