I think (1) and (3) don't necessarily have a single, correct answer; but, I can offer a few thoughts.
(1) We've generally had good success in initializing retrospective MPAS-A simulations using CFSR, CFSv2, NCEP FNL, ERA-Interim, and ERA5. There are tradeoffs to using each of these datasets, of course, so it's difficult to make a specific recommendation for the general case.
(2) We've performed real-time simulations aimed at capturing TCs using the 15-3 km variable-resolution meshes (circular and elliptical refinement), as well as the quasi-uniform 15-km mesh. Whether these meshes would work for you would in part depend on what computational resources you have available to you.
Regarding nudging (2), there aren't any options available in the current release (v7.x) of MPAS-Atmosphere for doing this. A group at the US EPA has implemented analysis nudging in an older version of MPAS (v4.0); here's the paper describing that work:
Adding four-dimensional data assimilation by analysis nudging to the Model for Prediction Across Scales – Atmosphere (version 4.0) . We are aiming to incorporate that work into the main MPAS-Atmosphere development branch, but unfortunately we don't yet have a timeline for when that might be released.
Generally, it's probably worth taking any of the analyses/reanalyses listed above (e.g., CFSv2), and, say, the quasi-uniform 15-km mesh, and running a simulation or two for a short but representative time period. This will help you to gauge whether the computational costs are manageable or not, and whether the model results look useful. This would also be an opportunity to make sure that you'll have the post-processing tools in place before committing to the full three months of simulations. But, the lack of a nudging option in the MPAS v7.x release may ultimately prove to be a significant hindrance.
