dedup-skills

# Dedup skills Produce a **ledger** of two defect classes across every markdown file in the target tree (default `skills/`): **repetition** — the same rule stated verbatim or near-verbatim in two or more places, cross-file or within one file — and **self-conflict** — opposing directives (must vs never) on the same subject. Report first, always: this skill edits nothing; edits happen only after the user approves ledger rows. Two duplications are legitimate and are never findings: a short self-contained rule repeated where it is needed (duplication of one short rule beats a pointer chain), and replication carrying a sync-lineage note that names its counterpart. The finding is the *undocumented, long-form* repeat — and the conflict. ## Step 1 — Vendored set Read the tree's LICENSES/NOTICE/attribution file in full before scanning. Vendored or license-covered paths are annotate-only: findings inside them are reported with a vendored annotation, never proposed for deletion, and the attribution file itself is excluded from the scan (registry, not prompt content). Done when: every vendored path is enumerated, before any scan runs. ## Step 2 — Cluster pass Mechanical script over all markdown in the tree. Strip YAML frontmatter first — frontmatter is load-bearing routing data and never a dedup target, so it never enters the scan. Segment into paragraphs (skip fenced code), normalize tokens, and cluster near-identical spans so that **every member of a cluster verifies against its...