iteration-close — distill, shed, prove takeover, rehearse bootstrap, seed next

A product that iterates forever accumulates two silent debts, and this skill pays both down at every iteration boundary — safely (nothing is lost) but ruthlessly (nothing dead is kept):

• Dead files linger and rot into two-truths — a deprecated tree sits around for weeks; a runbook keeps claiming a component is "not built" long after it shipped.
• Knowledge that lives only in a chat window dies with it — whole working pipelines have died in scratchpads; hard-won decisions die with the session. Keep the lesson, not the artifact.

It is generic: it works on any repo by reading that repo's own conventions, and where a convention is absent it creates the minimum rather than assuming or erroring. "A fresh window or new machine can continue" is a HOPE until it is TESTED; being current (pull) is always safe, but publishing (push) is a deliberate choice. Respond in the language the user is working in.

Two ways to run