cleanuplisted

# Cleanup — Staleness Sweep Code has tests. Prose has nothing. A README that describes last month's architecture, an agent instruction file naming a renamed script, a shared-memory fragment asserting a credential that was revoked — none of these fail a build. They are read, believed, and acted on. **Stale-but-trusted is worse than empty**: an empty store makes you go and look, a confident wrong one does not. This skill sweeps the surfaces that carry durable claims and reconciles them against what is actually true now. ## When to run - **Before a retro** — so the retro reasons about reality, not about the record. - **Before sprint planning** — briefs written against stale ground truth produce stale work, and that cost is paid at execution time. - **After a run of merges** — several PRs in a short window is exactly when prose falls behind code. - **When a stale claim has just bitten you** — sweep the neighbours; the same author, session or assumption usually produced more than one. Not a routine tidy. If nothing has changed, this finds nothing and should be cheap to conclude. ## What counts as a durable-claim surface Anything read as authoritative by a future reader — human or agent — that no test verifies: 1. **Entry-point docs** — README and equivalents. First contact; wrong here is expensive. 2. **Agent instruction files** — loaded into context automatically, so errors propagate silently into work rather than being noticed on read. 3. **Design and strate