>So, if a program is ONLY_FOR_ARCHES mipseb (because say we're running
>an SGI binary, hey, it could happen), then my solution works and yours
>doesn't.

And more to the point, if the target machine is a mipseb64, then
you're saying we can't install it.  Or should we never list mipseb in
${MACHINE_ARCH} unless we also add mipseb64?

I keep saying that fragmenting the ${MACHINE_ARCH}es all the way down
leads to database normalization problems and it's not the right
engineering decision.

That's a completely separate distinction from what magic numbers are
in a file and what the kernel reports for uname -a.  I'm even agnostic
about what uname -a should be, as I've said to Todd.