This is indeed a nasty, unexpected behavior in the minimizer. I like Jeff’s guess that you are in such an awful region of conformational space to start with that everything is bad, but it’s still surprising. What do before and after actually look like? If you start with protein eclipsed into protein, then you’re going to end in the same wrong state.
If you want to solve this domain assembly problem generally, the FloppyTail executeable can take a stab at it. Domain assembly (er, well, envelope prediction) will be fully supported in 3.3 and it sorta-works in 3.2.