My tests were describing the code, not checking it

AI-generated tests can pass while missing real-world edge cases, leading to incorrect rounding. Manual mutation testing uncovered coverage gaps, prompting a fix