That only works if the algorithm isn't already understood and so you don't know what optimal is. If you tdd sort there is no way to get divide the list to small chunks, insert sort each small chunk, then merge sort them. (20 years ago this is what gnu sort did, I assume that this algorithm has changed now that cache is a bigger factor in performance). Even knowing that the above is your end goal TDD to get there is wrong because you don't know how the best algorithm will change when something else comes up as important.

Tests are good, but they need to be universal for any implementation, which means you often cannot tests the internal details that prove you didn't use bogo sort (picking a pathological example to make the point)