Escaping the Nash Trap: Structural Estimation and Alignment of Strategic Reasoning in Large Language Models