Posterior Optimization with Clipped Objective for Bridging Efficiency and Stability in Generative Policy Learning