Text this: Large Language Models lack essential metacognition for reliable medical reasoning