Your setup seems correct. It even does not break for me when delta is smaller than 0.1 (HFSS 17).
The simulation produces extremely low meshing, so I would stick with my initial guess that the simulation is not adequately converged. The problem in this case is that you are converging in the passband, where S11 is extremely low, and S21 is extremely high, which won't change much even if the mesh is not adequate, due to the bandwidth of the passband.
Set the maximum number of passes to at least 20, delta S is OK at 0.002, but minimum number of converged passes should be something like at least 10. You could also try converging at the top of the passband instead of the middle.
Probably a better solution is to specify a seed mesh, with minimum length equal to at least half the radius of the corrugations.