improve-muse

# improve-muse — the hardening cycle One invocation = one hardening slice, end-to-end: find the most valuable problem in what Muse ALREADY IS → fix/harden/delete → verify → commit AND push. Sibling `grow-muse` builds what doesn't exist yet. **Boundary (one item, one owner):** BROKEN / debt / dead weight → here — and broken includes failing its function for the user WITHOUT erroring (unreadable text, dead affordance, wrong or misleading output). MISSING capability → grow-muse. Working-but-poor (usable and correct, just clunky — slow flows, confusing labels) → grow-muse. Cross-finds get one `- [open] for=grow-muse :: ...` record in backlog.md, never built here. A loop calling only this skill hardens forever — pair it with a grow-muse loop to move both axes. **Standing authorizations (Jinan 2026-06-27, this skill only):** push on green verify (never red); auto-pick with no human gate (a genuine ⏳ human-decision fork is skipped with a note, never guessed); scope bigger — a coherent unit, not the narrowest edit. ## The cycle 1. **ORIENT** — `pnpm self-eval`: non-zero ⇒ the regression IS the slice, stop finding. Else `git log --oneline -8`; is Ollama up? 2. **FIND — probe live reality first; ledgers are the fallback. Take the FIRST rung that yields:** 1. self-eval regression. 2. `node scripts/scout-signals.mjs` — recurring real failures in `.muse/runs/`. 3. **Live pain probe (≤5 min):** `muse doctor --json`; hit the core surfaces (one `ask`, `/health`, we...