> In my understanding, LKMs need to be compiled against at
> least module.h, and in practical terms against more headers.
> If any of those headers could just be un-GPL'd the way Google
> claims it can un-GPL the raw headers for Bionic, there might
> still be other arguments remaining for which one can claim
> that LKMs are derived works, but one argument would be gone.
Indeed, that is the heart of the issue. A compiler similarly "strips" comments and normalizes code; and in no sense is a piece of code dependent on a comment, or on the amount of white space in a declaration or definition; those things are semantically meaningless in a source file. Google posits, therefore, that all those things with semantic meaning in those header files is not copyrightable.
If Google isn't violating the GPL, then neither would any company distributing closed-sourced binary modules built against kernel headers. At best they would be liable under some sort of contributory infringement theory, except a non-distributing end-user would have a license to use the GPLd work, so there would be no infringement to contribute to. Under Google's theory, the derivative-ness doesn't happen until being loaded into the kernel. (Because if the headers are not copyrightable, it implies that neither are the interfaces to which they're intended to be run against; in other words, that there are other potential non-infringing Linux clones that might load that module. So until it's actually loaded, it's not derivative.)
People's defense of Google here is baffling. It's one thing to critique lawyers' poor technical comprehension. It's another to defend a copyright theory which affects such a clear and momentous change in Free Software licensing.