Learning Shrinks the Hard Tail: Training-Dependent Inference Scaling in a Solvable Linear Model