Replay circadian adaptation over hardest-case training: objective, accuracy, latency, decision boundary, and model internals.