A blind spot for large language models: Supradiegetic linguistic information