Vormav, your getting your terms mixed (for a good reason). Let me state it as I understand it?
Traditionaly Mental Ray can produce both diffuse and specular (caustic) GI using photon mapping. The only issue with this is that the photon only solution is usually patchy and a little coarse looking (sooty) and also prone to fluctuating from frame to frame if you had moving objects or lights. So again traditionally in MR GI was photon mapping based yeah? And the photon map was based on world space, not camera space, so you could move your camera and the photons would still be valid which is cool but you could move objects and not re-render the photons and that would leave ghost objects (as seen in your corner sphere test image).
Now… to arrive at a smoother solution a final sampling system was added to sample some final GI values which could be used to reach a much smoother/pleasing result. This was called FINAL GATHER (i.e. final as in a final stage to acheive good quality). So, the Final Gather option was something added to help out with photon mapped GI. And the Final Gather result WAS camera dependant as it took what it already saw in the render and refined it. Move the camera and the Final Gather solution would have to be re-calculated. And this solution was a mixture of Photon based light and Final Gathers re-working of that info.
Ok, so this is where it gets messy. Some people noted that you actually got a pretty good skylight like GI effect if you didn’t bother with the photon mapping and simply just rendered with Final Gather turned on (with a self luminous sphere or an env map to act as a sky colour). I say skylight “like” though as Final Gather also calculated bounced light also, with colour bleeding, and imho because of this I’ve found it un-suitable for the role of a pure skylight (its gives crazy colour bleeding effects in places and also overbright light bouncing in others.
So after a while when people refered to GI in Mental Ray they often called it “Final Gather” where as in fact the Final Gather system was originally designed as a buff for its photon based GI.
So, there we are. If your talking about caching the “Final Gather” solution (i.e. no photons involved) then it IS camera dependant. If your rendering using ONLY the photon solution then they are already cached and that is not camera dependant. Problem is, photon only GI in MR tends to be pretty cack, so lets forget this option. Third scenario is rendering photons with final gather turned on to smooth out the photon sootiness effect. This scenario again IS camera dependant because Final Gather is involved, and it is view dependant…
So, there we go, hope that makes sense. So MR’s caching of the final gather result is actually it saving out the GI only component of the current render to re-use in the next render. That can be Photons + FG, or FG only. If you move the camera that cache file is useless. Believe me, tested it (3 years ago now) and its how it works (the cache file if you can find it on disk can actiually be loaded in photoshop and it looks like a GI only render).