How does a 167M-parameter protein foundation model outperform a 650M general-purpose model on the tasks that matter?