1 & 2: This is setting the bar rather high, no? I hate GWAS, but am surrounded by people who do them all the time. To show any kind of function for a SNP is hard, to do it in humans is very, very hard for something like this involving nervous system function, and to prove the difference in function is "adaptive" virtually impossible. The furthest people go, and fairly rarely, is to show an eQTL and spin some unconvincing tale about how the difference in expression might have some effect.

But array-based methods do have one advantage which is that the 450K for example is heavily concentrated in genes and promoters. So the variant, which indeed does not span just one nucleotide, does have at least putative functions, which are simply the known functions of the gene, although yes, you cannot say whether the variant itself is helpful, harmful, or null on the functions of the gene. It is at least a step up from people who come to me (bioinformatics) and ask me to use my divining rod to guess what their SNP 100Kbp from the nearest gene might do.

3. Putting it in an abstract is fairly desperate, I agree. But broadly speaking the way you know you aren't getting total nonsense is if a lot of related terms cluster towards the top. Which looks to be what they got. It looks a lot cleaner than most GOEA results I see.

Perhaps you could educate me on what would be proof of positive selection. Would it be at least convincing evidence, if not proof, if you showed that phenotype P is associated with variant V, and variant V is associated with positive traits X,Y,and Z after controlling for P (so you can rule out direct links between P and X/Y/Z which bypass V)? From what I have seen, investigations into functions of individual SNPs are almost invariably a waste of time ending in handwavy stories so I wonder if there is a better approach.