The question is whether the C-code should be truly generic, or if
(reasonable) assumptions about the hardware can be made.

If you're going to write code that goes in the machine-independant portion
of the tree, it should be truly generic.  If it's to go in the machine-
dependant portions of the tree, it can make unreasonable assumptions about
hardware.  Trying to explore the middle ground is just asking to make
someone's life difficult in the future when "reasonable" changes.  (Remember
when 8-bit processors were "enough"?)

(htonXX and friends are, I think, perfectly good candidates for being
machine-dependant.)