1: Yes, vtf's take up lots of ram, vmt's don't.
2: No, this will just waste ram, Mipmaps automatically do what you're asking about.
3: Not much of a performance question
But in source, brush models aren't "lit" the way other models are, they're not lit per vertex or per pixel, but instead their diffuse textures are multiplied by a separate "light texture" that's generated per face during map compilation. There's no alternative to this lighting method so you'll have to achieve whatever effect you're trying to create with some clever workaround, what exactly are you trying to do? As far as casting shadows goes, I'm not sure for brush models, but I know that some dynamic entities have a disable shadow input, instead of making your brush a plane brush, see if you can make it into a different brush entity and disable shadows that way.
4:Mipmaps do increase filesize (by 33%) but they do not decrease performance, in fact they actually increase performance. It's difficult to explain why without talking about the pixel pipeline, but long story short: it's faster for the gpu to sample colors from smaller textures, so the further away the object is from the camera, the faster the texture lookup will be on that fragment, because it's sampling from a smaller (therefor faster) mip texture. Most of the calculations that are needed for mip mapping are implemented directly in hardware, so they're really really really fast.