Hm. That is definitely odd behavior from the minimizer. I wonder if it is because you are in a very bad region of the energy function. That is, if the random phi/psi selection creates badly overlapping clashes where your two domains largely overlap (which you would see by a very high fa_rep score, as well as in the structure), then it may be hard for the minimizer to resolve, due to the very high energy and gradients. Because the minimizer is sampling discrete points in the domain (in your case phi-psi of residues 161-168), it may get confused because the energy function is not smooth enough.
If this is the case, a workaround might be to reject configurations before minimization if they are above a certain fa_rep threshold which might indicate unreasonable levels of clashes.