Rule-bottleneck Reinforcement Learning for More Interpretable and Efficient Language Agents in Resource-Constrained Allocation Tasks *Code under development