It's hard to form a full picture of someone else's product from a few words, or to imagine all the possible consequences of doing it one way or another. Off the top of my head I would say if the various parts are inter-compatible (eg same JPs) then they should go in one cr2. If there is to be a helmet this should probably be a separate item (so it can be carried in the hand, or placed on a table). Probably the same should go for gauntlets.

To me putting all you can into one cr2 sounds like less work for the end user as they only need to load one file, then all the options are available through geom switching. To leave an item out an empty obj can be used as one of the geometries. I have never tried using ERC on a 'geomChan', but can't think of any reason why it should not work, so perhaps all the geom swapping could be done from the BODY, or some dedicated ghost actor, via ERC. This could cut down on work by allowing things like switching right and left thighs and shins via one dial, instead of needing to set four dials.

Mazak, makes a good point about geometry switching and people who don’t want render in Poser. Perhaps an optional add on pose pack could be made to inject the various geometries, rather than dialling them in, to accommodate those rendering in other apps. I don't know how other apps handle geometry, so am not sure if this would work. Making such a pose pack should be relatively quick and easy as you will already have most of the obj files to hand for the geom switching.

I know you are trying to keep the file size down, but IMO you should allow for future expansion my putting 3 empty targetGeom channels in every actor, include a few spare materials, and perhaps even some spare actors.