I’m all that familiar with the minimization internals, but I also would have guessed that the “dfpmin” version shouldn’t ever result in increases to energy.
That said, you may want to play around with different minimization algorithms (see http://www.rosettacommons.org/manuals/archive/rosetta3.2.1_user_guide/minimization_overview.html for an overview), and see how those work in your case. I would particularly suggest trying “dfpmin_armijo_nonmonotone”. It should be slightly better at escaping bad regions (although your case may be pathological enough to show problems even there.)
In addition to the other suggestions, you may be able to escape the bad regions by using some sort of Monte Carlo proceedure – instead of just minimizing, do some (small) active perturbations, with or without minimization in between, and hopefully you might wiggle yourself out of whatever strange conformations random angle selection has got you into.