A framework for hallucination mitigation in domain-specialized large language models with application to aviation maintenance decision support