AI-generated code looks just as clean as human code in review. That is exactly the problem: review was built to catch reasoning, not syntax.