If that is indeed the case (and you may very well be right), than how can copyright prevent distribution of binary modules built against those headers? If the actual header code itself describing a particular interface of the kernel isn't copyrightable (as the same expression or as derivative of the expression), then how can anything be derived from it?
If you say, "well, it's deriving from the interface, not the header code" then how is the header code likewise not also equally implicating this abstract interface? Certainly the header code often describes this interface in more exacting detail than the binary module. Some of the material you cited is predicated on the notion that the interface being described isn't copyrightable for very specific reasons relevant to the particular interfaces at issue--an extremely fact-centric analysis not likely to apply to Google's largely indiscriminate copying, __KERNEL__ guards notwithstanding.
Further more, your substantial similarity argument applies equally well--both quantitatively and qualitatively--to the binary modules' proposed infringement. Clearly, how we slice and dice the scope of the infringement is critical.
For something to be derivative, it must create a sufficient impression in one's mind of identifiable, copy-protected elements of the derived work. If a header file cannot do this, how can a binary module do this? It hardly matters that the binary module doesn't work without the Linux kernel. I can write a story in such a manner that every third page seques into a substantial Phillip K. Dick paragraph. And I can distribute that story, sans those paragraphs, if without those paragraphs you couldn't identify specific copy-protectable elements of Dick's stories. That when you put them together it becomes a derived work hardly matters by itself.
I'm not challenging the soundness of the cases you cite. I'm merely suggesting that those aren't necessarily applicable to Google's wholesale copying of huge portions of Linux kernel header code. And if Google is legally in the clear (which IMHO is doubtful with such a huge collection of code, no matter how thin the copyright), then there are tremendous, potentially undesirable consequences for copyleft enforcement. (Though a BSD-style license proponent might not care, and even might welcome it, as strict license enforcement is of lesser consequence and often creates more confusion than clarity.)
It would be interesting to read your reply, but this shall be my last. I don't mind if you get the last word; in fact I would prefer it ;)