Scaling Competence, Shrinking Reasoning: Cognitive Signatures in Language Model Learning