Back then when I tried to write a 3d solver for assemblies I failed for that very reason: standard numeric solvers are unable to cope with the fact that rotational and transitional degrees of freedom have different sensitivities, a unit change in the parameter leads to to massively different changes in the sketch dependent on the type of parameter. So the sensitivity for translation to translation is always 1, angle to angle also always 1, but angle to translation may be way higher. That is a unsolvable problem, as reducing the angle sensitivities may solve angle to translation problem, but than created a angle to angle sensitivity problem, as this gets way too small or big dependet on scaling. Another issue is that one solver variable is part of many output dofs, hence you would need different scaling for each single output dof but have only one scaling factor for the variable. There was a very nice easy to read paper about that, but can't find it anymore...
I always thought the 2d solver did not have this problem, as all variables in the sketch have been translations (only point x,y coordinates as free variables), obviously that changed (I looked last 5 years ago
). But that could be a possible solution, maybe you can rewrite some of the angle variables in a way it only uses translational dofs?
As a note, all free 3d assembly solvers have this problem. e.g. for solvespace you can easily construct problems that makes it fail all the time just by choosing the sizes of parts.