Current foundation models outperform trainees on text questions but struggle with multimodal ophthalmic imaging ...