“Test-time compute and pretraining compute are not directly exchangeable; test-time compute is less effective for hard problems and cannot compensate for large gaps in a model's base capability.”