Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox