That reminds me -- I've been conceptualizing, on and off, a way to take a closed mesh and break it into convex shapes so it could be used for collision detection. I'm not talking about the crappy way current tools do it where they make a brush for every face, but an actual "smart" algorithm that would break things up in close to the most efficient way possible. Ideally I'd like to have it so you could plop a mesh into your map and have it get compiled in like solid geometry. Even more ideally, it would have level of detail, so the whole world could be dynamic and reduce polygons for stuff that's further away or on systems that are slower. Mixing collision and LOD will be difficult, though. It won't be too bad for players, but it would look weird for things like splat decals if a paintball hits an invisible detail because the LOD mesh doesn't match the collision mesh.
Another thing I'd like to do is have a more advanced radiosity system -- something kind of like Brazil, maybe. You've probably seen the really pretty no-texture renders floating around. Things like this:
http://home.xtra.co.nz/hosts/polycount3d/wip004.jpg . It would also be nice to have the level of quality controllable by time. Like, the longer you run it, the better it looks, but you can just run a quick preview, and maybe it would be a bit splotchy like this:
http://www.christopher-thomas.net/pages/brazil/images/set1/lightmesh_v01.jpg . Also, it would be cool to have the resolution of lightmaps automatically scale to the detail present, so if you had something complex like a grate casting a shadow on something, it would have a higher resolution lightmap than some flat surface with little to no shadowing.
Maybe someday...