In a call like (do (for [x (do (println "realized") nil)] x) nil), no elements of the for comprehension are ever requested, and so it is not actually necessary to evaluate the inner do-block. However, this expression causes "realized" to be printed, because the first sequence-expression in for is evaluated even if no items are ever requested from the output lazy-seq.
It's not documented whether this is intended or unintentional, but I was surprised by this behavior, and a brief unscientific survey on #clojure suggests that other users, even "old hands" who've been using clojure for years, don't expect this either.
I've attached a patch that wraps the problematic expression in a lazy-seq call. This is not quite ideal, because it means that the first iteration is "lazied" twice, as in ((fn step [s] (lazy-seq ...)) (lazy-seq xs)), but a change to make this not happen would be much broader in scope, and this seemed the least dangerous.