It would seem that the code "wit" below would run faster (by 2) since it does half the loops, but it runs longer!

No, the wit part of the benchmark runs 0.68 million times per second, while the non part runs 0.51 million times per second.

So it does run faster (by a factor of 35%).