I suppose I usually sieve over algebraic specialQ out of habit; if asked to justify myself, the algebraic side for GNFS jobs is generally much larger than the rational side, and I believe it makes sense to use the specialQ to render effectively smaller the numbers which started off largest, but that's not an answer for why I use a pretty much universally in SNFS cases.
I haven't done the experiments to see how much duplication there is in a case with roughly equalsized rational and algebraic side if you sieve on both sides; it might be sensible as a way to push yields up on SNFS problems with particularly intractable polynomials.
