But it leaves me worried that $_ is being used in say() as a global...

You can't "use as a global", you can just "use a global", and you're the one who asked to use $_.

So the question is, is that $_ referenced in the say() sub a global from the implicit $_ in the mainline loop, or not...?

Yes, because $_ is a global. Actually, it's a "super global", meaning "$_" refers to the same variable ($::_) no matter which package you're in.

Note: It is possible to make $_ a lexical, but that feature was also introduced in 5.10, it's experimental, and I don't think it will survive. (A step was already taken by core to avoid using it.)

So what you get is:

BEGIN {
   if ($] < 5.010) {
      *say = sub { local $\ = "\n"; print @_ ? @_ : $_ };
   } else {
      require feature;
      feature->import('say');
   }
}
[download]

There's probably a module that does that.