Did you mean to have a picture in your post ? I don’t see any, you may want to repost it.

“Blindly”, I would suggest to check your ray directions exiting the source. Non sequential will trace any rays, independent of the sequence you want, so these rays may be valid, just not the ones you are interested in.

The photos somehow disappeared the first time, but they should be visible now, thanks! The ray direction does not appear to be the issue since the first pass of each mirror looks as expected. 

Do you use polynomials (e.g. asphere coefficients) ? That could result in different shapes if the signs are not set properly (sequential use the propagation direction as reference for +Z and non-sequential use the outwards surface normal, so sometimes they have opposite directions). Main curvature is easily checked, but corrective terms are harder to verify.

In the forward direction, the topmost blue rays (last segment?) is not exactly horizontal. You would need your starting ray to have the same angle, but the topmost red it looks perfectly horizontal.

Also, you may want to fletch the rays to make easier to debug.

I have a few parabolic surfaces. Adjusting the starting angle of the reverse ray gives me the results I would expect. I will take a look into the rays to determine the how far from collimated the forward direction is. Thank you for your help.