> I also think this will definitely break backwards compatibility for
> non-base prefixed numbers.  That's actually worse than the complexity
> (which is managable).

Hmm. If we assume the most restricted syntax proposal so far for
non-prefixed numbers, that would mean that somebody wrote tokens like
1_234_567 in a source file and expected them to parse as symbols. Or
generated source files containing tokens like that -- but most languages
outside the Lisp family parse that as a number anyway, since it starts
with a digit, so it's not clear for what purpose those things would have
been generated.

Internally it would still be possible to represent symbols like
1_000_000, i.e. (symbol? (string->symbol "1_000_000")) would still work
and return #t. Those symbols could just not be read in by the reader.
Unless you use pipes '|1_000_000| in which case it would be parsed as a
symbol as it now is.