Require mutation checks when a subagent writes the tests

A subagent asked to write tests for code it or a peer just wrote will produce a green suite that often cannot fail. Make the deliverable include mutation checks: break each property on purpose, record which tests fail, then revert. In one run this caught that removing a route from the telemetry skip-list failed 2 tests, flipping pydantic extra=forbid to extra=ignore failed 9, and opening an event-name allowlist failed 2 — proving none of those assertions were vacuous. Cheap to demand, and it converts a claim of coverage into evidence. Also require the suite be run twice back to back, which catches fixtures that only pass against a freshly seeded database.