Well, I tried a longer test of my wbinvd() kernel.  It corrupted the disks.
(Though there is a small chance that this was due to a previous bit of
corruption causing a crash at an inopportune moment.)  I suspect, then, that
this may not be a case of hardware not snooping but rather of some more
fundamental bug in the caching logic, though what I can't imagine.  (Perhaps
the writeback cache ignores the writeback signal.  Yuck.)

Or maybe I put it in the wrong place.  I'll have to look over what I did
more carefully.