(1) if the VRAM is available, why not use it to cache things?

(2) VRAM usage monitoring is mostly useless.  The I/O on and off the card is far more important.

(3) GC3 really isn't fully optimized yet, so I would expect a bit of slop in things like texture caching strategies.

(4) We'd have to ask the Dev's, but there's a possibility the game is using the GPU for compute purposes, not just video display.

(5) GC3 uses a fair amount of very large textures, rather than lots of little small ones. The latter (in FPS and RPG games) has a very high "use and evict" ratio, wherein textures are quickly discarded after being displayed a couple of times, in place of new ones being loaded in.  GC3, being a much more static game, almost certainly reuses a large amount of textures repeatedly, so it pays to cache them on-card.

Also, what else did you have running before you started GC3?  I suspect a good amount of that 3GB of textures are leftover from whatever was running before, and GC3 didn't evict it because it didn't need the extra space.

I/O off the card refers to the load time to get a texture from main memory onto the card.  Basically, bandwidth into the card.

You're far better off pre-loading as much textures onto the card as is can hold, rather than doing a "just in time" load, because the time to pull from main memory to VRAM is far, far higher than merely just accessing what's in the VRAM already.  In essence, VRAM is a L3 cache for a video card (that's not a perfect analogy, but close enough).

VRAM bandwidth is in the 5GB/s range.  I/O bandwidth of a PCI 2.0 16x card is less than half that, in real life use, when copying textures from main RAM. It's 0.1% of that if it has to pull the texture from static storage.