A new multi-agent reinforcement learning framework, OGR-MARL, is proposed to solve the heterogeneous USV cooperative pursuit problem in constrained port waterways. This framework integrates shared evader belief, role-conditioned option targets, adaptive rule penalties, and residual policy learning, allowing different MARL algorithms to learn corrective actions on top of rule-guided behaviors. The experimental results show that OGR-MASAC achieves a 75.0% capture rate and demonstrates good generalization in complex port scenarios. Blogger's Review: OGR-MARL provides a new perspective and method for solving complex problems in multi-agent reinforcement learning, with strong practical value and research significance.