The corner cases (1) of my previous proposal are making the algorithm difficult for you (I ve never had to zoom out from the globe and still don t understand why you still don t just use the hemisphere in that case), so here is another method that can work.
Take your projection frustum as a 3D object and intersect it with the globe (as a 3D object). There are a variety CSG algorithms that can give you union, intersection, and difference. Use intersection. This will result in a 3D mesh of the pieces of the globe that intersect with the frustum. The extents of that mesh are the extents of the bounding box. (Project along the major axis of the frustum.)
This is horribly less efficient than my previous proposal. :-)
(1) Pun completely intended.