As far as I understand (correct me if I am wrong), the outer part of the coaxial is the ground of the transmission line (zero voltage), and the inner part carries the RF signal (varying voltage).
Stefan, I tried to outline that this understanding is not quite correct. The zero voltage, or symmetric voltage, implies that you measure center conductor and ground conductor voltage relative to something. To model things properly, it is important to understand what that "something" is, and if it exists in the model.
By the way, since there is a physical gap between the two arms, the ungrounded internal port you proposed may not be suitable in this case.
Trust me ... I suggested this method based on more than 10 years EM modelling experienece with Sonnet ;-)
The second method in your drawing, where one port is connected to a different location, is not correct. The signal and ground of a port must be physically close, otherwise some arbitrary return path between signal and ground is part of the results. This is true for all solvers, and some solvers have explicit warnings about this (unphysical results) in their documenation.
The first method from your drawing can be implemented, if you want to use boxwall ports. You can use ports +/- for differental or +/GND for single ended. In this case, the box ground is the global reference. But the question is ... what is the meaning of that ground reference? See above.
I still believe that my suggestion in port 2 is the best and most accurate model. But you can try different methods and see what the difference is (or if there is no difference).