In all honnesty this a question I'd very much like to avoid.
The simple truth is: we haven't decided yet. There are many technical unknown factors in this and deciding on a definitive number is impossible for now.
So the rule of thumb is to keep it 'reasonnable'.
(if that is any help)
The one thing I like to emphasize, is that the limitation is more on the number of polygons that can be seen from the street. Because when we'll have lots of shells those are the ones that will slow down someone's computer when walking in the street looking at all those shells. Polygons that are hidden from the view shouldn't matter be cause they shouldn't be taken into account.
Polygons that can be seen from the streets are: those that are behind the entrance of your shell and those that can be seen above the walls.
NowI say 'should/shouldn't" because at this point this is mostly technical speculation on my part. There are many things the blender plugin can do, but those kinds of 'rendering optimizations' are not part of those.
So these polygon limitations are more something for the future, when the plugin will have evolved enough to include that, or when we'll get tools from Cyan which use that. So for now let's just make things 'reasonnable' and we should all be fine.