Even when that does happen... and I'd argue that it often does not, if you've done the work to make your interface pretty, and write a precise, accurate test suite, you've usually done the work to make the actual functions neat and orderly...
Even when you end up with a black box full of entropy, it's segregated from the rest of the system. You can feel free to change the rest of the system around it and know that the black box will keep doing it's job, as long as you keep using the well-written API within spec.
But, that's not all you get... because you took the time to write the tests, you can refactor this entropy box to your heart's content... until it starts looking more approachable. You've got your tests around the API, right? Then you don't even really need to change them to do the refactor work. The only reason to change the tests would be if you want the function to do something else (or you missed something, of course).
Take small steps... refactor out a couple lines at a time... run the tests with each iteration. Based on previous experience, you're going to end up with a fairly clear and concise implementation... not to mention performant.
Absolutely! Get something working and then keep it working while making small steps.
On the other hand, the example that comes to mind is Ron Jeffries' TDD sudoku solver.
I've seen several systems where the magic black box is doing things hilariously wrong---as long as it works on the test cases and the production results are sufficiently difficult to verify, it'll be accepted as gospel.
Even when you end up with a black box full of entropy, it's segregated from the rest of the system. You can feel free to change the rest of the system around it and know that the black box will keep doing it's job, as long as you keep using the well-written API within spec.
But, that's not all you get... because you took the time to write the tests, you can refactor this entropy box to your heart's content... until it starts looking more approachable. You've got your tests around the API, right? Then you don't even really need to change them to do the refactor work. The only reason to change the tests would be if you want the function to do something else (or you missed something, of course).
Take small steps... refactor out a couple lines at a time... run the tests with each iteration. Based on previous experience, you're going to end up with a fairly clear and concise implementation... not to mention performant.