I propose to decouple those two and make __NetBSD_Version__ only an
  indicator of the kernel ABI, completely independent of the version
  reported by "uname -r".

I think this could be confusing.  Your problem is caused, in my view,
not by the __NetBSD_Version__ value being wrong -- your system did
indeed change from 3.0 to 3.1, but by the lkm machinery interpreting
system version number as a kernel ABI version.

So, perhaps we should introduce __NetBSD_KABI_Version__ and leave it
at 300000000 on netbsd-3, and keep changing it as we do for current.

I realize my comments can be viewed as nitpicking over name choices.
Your proposal of keeping the variable as-is for current and not
changing it along the stable branch (where KABI changes are
prohibited) seems very reasonable.