Will Clinger, who wrote the portable implementation, is on record as saying that round-to-even is better than round-away-from-zero.  I'd like comment, preferably with arguments, from others.  My own inclination is to make flround do what C round() does, namely round away from zero, as part of my original intention for the SRFI, but I don't want to make flround un-useful for Scheme programmers.  On the third (possibly gripping, possibly not) hand, RnRS round is still available for round-to-even.