A Blind Spot for Large Language Models: Supradiegetic Linguistic Information