Beating o3/o4-mini with Codebase-specific Reinforcement Learning