Reinforcement Learning for Large Language Model Selective Evidence Adoption from Contaminated Retrieval Results