And I think human written tests at that. If the LLM is blind to the failure mode X, does it know to reliably write a test to evaluate the behavior of X?

wow. there are actually no tests here.