There are two epsilons.
epsilon1 determines two things:
1. whether a constraint is active in a QP sub-problem (numerically determining zeros)
2. whether a new solution in the next iteration changes much; if not the algorithm stops
epsilon2 is the epsilon in QPPrimalActiveSetMinimizer, which determines
1. whether a constraint is active (numerically determining zeros)
2. more importantly, it checks whether an automated computed initial solution is feasible. If this initial solution computed by SuanShu is not feasible, it will declare the problem infeasible.
I suspect that the last point makes your problem fails. Does increasing epsilon2 help your performance?
Again, if you send us your sample code (that compiles), we are happy to look into it.