Chasing the Tail: Effective Rubric-based Reward Modeling for Large Language Model Post-Training