Yeah, a lot of the original work on how GAs would optimize NP-complete problems is pure hand-waving and wishful thinking. Building-block theory, schema processing, it all sounds nice, but tends to result in disappointment. The thing GAs are actually good at is population-based search -- maintaining a bunch of alternative ideas at once. Either your suggestion of using GAs as a local optimizer for the non-satisfiability part of a mixed problem, or the inverse, using the GA to suggest starting points and using a SAT solver for the local search, is much more likely to produce interesting results. It's an idea I keep coming back to, although I still haven't struck gold with it.