I'll go first.
I'm 100% in the set-operations camp. But what happens when the set logic on the entire desired input domain leads to a such a large retrieval that the query slows down significantly, comes to a crawl, or basically takes infinite time?
That's one case where I'll use a itty-bitty cursor (or a while loop) of perhaps most dozens of rows (as opposed to the millions I'm targeting). Thus, I'm still working in (partitioned sub) sets, but my retrieval runs faster.
Of course, an even faster solution would be to call the partioned input domains in parallel from outside, but that introduces an interaction will an external system, and when "good enough" speed can be achieved by looping in serial, just may not be worth it (epecially during development).