> if you take that approach when testing all the thousands of components together, which is simply just too complex to exhaustively model[1], you'll never get anywhere.

This is exactly why ideas like test-driven development don't work well as a general approach.

Most realistic systems exhibit non-linear interactions where correctness is not compositional. Local correctness does not compose upward in any meaningful sense. Top-down design (working backward from the customer) allows for you to perform what is effectively one big global search. Bottom-up design (TDD) requires many local searches that all have to fit together perfectly at the very end. With units & composition, the consequences of component A's interactions with component B may not be considered until nearly the end of the project. If you are testing an integrated vertical, you will discover these interactions much earlier.