Sorry for the late reply.

I added "should" here because it is what Gambit and Chez do for e.g.
(atanh 1), and I think it is the correct behavior: the function is
undefined at exactly that point, and exact operations should have exact
results (even if that result is the raising of an exception).

The wording "should" is in the IETF sense[1], where it is still allowed
to *not* raise an exception. Hence all of the behaviors you gave in your
email are conforming, and reasons such as backward compatibility,
consistency, and numeric tower extensions are (in my opinion) good
reasons to not raise an exception.

However, if one were to write a Scheme implementation from scratch and
adopt this proposal, I recommend raising an exception.

-- Peter McGoron

[1]: https://www.rfc-editor.org/info/rfc2119/