Repro - Reinforcement Learning with Verifiable Rewards: GRPO's Loss, Dynamics, and Success Amplification