That's the problem with LLMs. Sometimes they follow instructions too precisely, others they sneak in a slow and buggy IEEE-754 parser instead of using `stof`.