Discriminative Constrained Optimization for Reinforcing Large Reasoning Models