The git tree is the source code

FWIW, I don't believe that such a requirement would fly in a court; at least not for quite a few years. Said that, size arguments are BS:
$ du -s .
652664 .
$ du -s .git
118828 .git
and that - on a tree with alternates pointing to straight mirror of kernel.org linux-2.6.git (which obviously doesn't need to be distributed). Note that unpacked kernel source eats about 4.5 times more than everything in .git. IOW, it's really noise. Granted, that's unpacked (i.e. working tree, not package being distributed), but packed (tar.bz2) will give only ~2 times increase compared to that of source without history - more than 1.2, but not very much more.

_Legally_ git doesn't qualify as "preferred source", but for all practical purposes it is strongly preferred as far as kernel work is concerned. gcc is a different story - they prefer suckversion and _that_ is a space hog, indeed. binutils... IIRC, they also use svn these days.

I don't know how to express that in license without running into insane corners, like "you must never rebase / cherry-pick / fold incremental fixes". On the other hand there's patently obnoxious behaviour several groups used to demonstrate - once a year or so they ran diff between the mainline and whatever had been in their CVS tree and post megabytes of non-differentiated garbage to e.g. l-k, usually reverting a bunch of fixes in process. Hell knows... TBH, I doubt that license is the right tool here...