Yea, thanks. I'm hoping the info in this thread will pile up, because I don't, by any means, live under the impression that what I did here is a comprehensive analysis on what goes into Mantra's dealing with geometry generation at render time, either from an optimization pov or a looks one. In order to get there, this thread will have to wait for me to become a Mantra expert or to have other users chip in with their knowledge and practical results.
I know the user base in general is not focused on this type of work, but I'm sure there a few out there and I'm hoping they'll take notice and respond.
Imho, generally you'll want to use render time subdivision, and better dicing quality if first option is not enough (both are under parameters of your geometry). That's, well, generally. Mantra seems to be using a sort of Renderman style, render time tessellation, so if mesh is not tessellated enough, it can't place the texture. That's how for example 3delight is working, but, while with 3delight plugin for Softimage, two or three parameters are enough for pleasant life (together with built-in automatic texture conversion), and error is just a blurred image - in Mantra, there are much more parameters able to do something wrong, object local and global render options. For example I'm not sure is there a way to get MIP mapping from external tifs, or there's always necessary to manually convert to RAT.
Not an answer, anyway somehow I got a habit to avoid any bitmap input in H, except environment texture. Good luck.