Heuristically evaluates each iteration via sleepwell-helper evaluate; persists rating/observation/course_correct in state.last_eval and injects into the next prompt.
