Coercion rule for comparisons

[Rem@IMSSS]

Perhaps the rule for binary (two-arg) comparison should be first coerced
to least-accurate (a la arithmetic) to see if they can be distinguished
(shown non-equal) at that level of precision. If so, we can immediately
answer the pending question at minimal cost. If they look equal at
least precision, then coerce the other way (most precision) to get
a definitive result.

If that were the rule for binary comparisons, would transitivity then
hold for all the standard comparison operations except not-equal?
-------