I'm the main author of SICStus Prolog.

You have rediscovered a trait of NP-hard problems: even if average runtimes are small, there are always outlier instances that take a huge amount of time.

What's surprising in your data are the low backtrack counts. It is as if the search makes no mistakes and all the time is being spent in constraint propagation.

Something that could contribute to the difference between SICStus and other solvers is your use of "A in 0..sup, B in 0..sup", which says that A and B have no upper bounds. Most other solvers either require a given upper bound, or impose some MAXINT bound.

I will dig deeper and get back with more comments.