Another cool thing, that was counter intuitive to me is that more than just rejecting multipath, spread spectrum can actually make use of multipath to improve the link (although I doubt it is done much in practice).

Conceptually, what you do is correlate on the incoming signal like normal, this first lock is the dominant path. You take the component that correlated and subtract it from the incoming signal and correlate again on what remains. This lock is the second multipath. You can do this to recover as many multipaths as you want or have dynamic range for. Finally, the relative offsets in the correlations tell you how to re-align all the multipaths in time. Once they are aligned you can sum them together and get more SNR than if there had been no multipath at all.