skill-enhance

# Skill Enhance **Core responsibility:** Take one existing Lattice skill from current state to modern-capable-model grade through a fixed pipeline: baseline → dependency map (background agent) → precision rewrite with guaranteed behavior preservation → combined three-skill QA → independent verification. The skill being enhanced must end up smaller-or-equal in redundancy, fully grammatical, convention-complete, and provably non-regressive. Organizing objective: **deterministic execution** — every sentence has exactly one reading, every reachable state exactly one action, every cross-file reference literally true. Acceptance test: two different capable models reading any sentence would act identically on it. **Input:** One skill path or name (any tier). Molecules are the primary case. **Output:** Edited files + close-out report. **STOP: NEVER commit, push, stash, checkout, reset, or run `build-codex-plugin.sh`. Working tree only — surface install/mirror commands for the user instead.** **How to verify this skill did its job:** - Pre-capture git diff exists; every preserved behavior in it is confirmed present after rewrite - QA trio re-run returns no critical gaps, CLEAN or tightened-only results, and PASS - Independent verifier reports zero real defects - Scenario walk leaves no branch with two possible readings ## Step 1: Scope and baseline 1. Read PROJECT.md (single source of truth for conventions) and the target SKILL.md in full. 2. Classify: tier (atom / molecule / r...