The more I look at this, the more I'm coming around to your point of view. When I look at

ext/File-Glob/t/global.t

they appear to be doing extensive testing to make sure a glob overload behaves in the same way as the core glob, including loops using <> and explicit invocations of glob. If I modify global.t as follows

4,5c4
<     chdir 't' if -d 't';
<     @INC = '../lib';
---
>     chdir '/home/jpl/src/perl-5.14.1/t';
94a94,111
> 
> @s = ();
> while (glob("/tmp/dir with spaces/*")) {
>     push @s, $_;
>     last if (@s > 1000);
> }
> print(scalar(@s), "\n");
> print(join(" ", @s), "\n");
> 
> package Bar;
> use File::Glob ':glob';
> @s = ();
> while (glob("/tmp/dir with spaces/*")) {
>     push @s, $_;
>     last if (@s > 1000);
> }
> print(scalar(@s), "\n");
> # print(join(" ", @s), "\n");
[download]

I get

2
/tmp/dir with
1001
[download]

at the end of the output (the directory has 3 entries). It's above my pay grade to figure out why this happens, or how to make the replacement using :glob behave as core glob does, but I'm running out of arguments why the current behavior is correct. Your suggestion of a new tag to get this behavior seems reasonable.