Generating Natural-Language Surgical Feedback: From Structured Representation to Domain-Grounded Evaluation