In case of a tie, I believe the convention is to round to the
nearest even integer. So round(x) = floor(x+1/2) isn't an
identity. I think that a good reference for this is
"Concrete Mathematics," by Graham et al. (I don't have
a copy to look at right now.) Another reference
is the NMC (http://www.nmconsortium.org/), but
http://www.nmconsortium.com/docs/NMC%20Technical%20Specification%20v1.0%20_DRAFT.pdf
doesn't seem to mention the round to even rule.
One more thing: the NMC document says that 'fix' should round towards
zero. Our fix is a floor, I believe.
Of course, with the related, but different functions round, floor, ceiling,
fix, entier, and truncate, we can generate expressions that vanish
that Maxima won't be able to crunch to zero. And that's bad.
Barton