Action cache: replay resolved intent→action with ~zero LLM tokens

The expensive step in a run isn't performing an action — it's the model deciding *which* element to act on.

**Idea:** persist the resolved intent→action (ref strategy, op) in `store` keyed by intent + URL pattern, and replay it deterministically next time, falling back to the model + re-caching on drift. This is the biggest single lever for making repeat tasks fast and cheap. See `docs/how-it-works.md` for the model. Design discussion welcome before implementation.