Well, I guess I don't.

If minimal memory consumption is part of the "spec" that you are attempting to deliver (i.e. a non-functional requirement), then I'd call it core and wouldn't worry about an extra scalar and code to implement the larger savings. If it isn't part of your spec, then this may be a case of premature optimization.

On the weak reference idea, it was really just equivalent to your #1 -- but using an extra (weak) reference instead of just copying the reference from the first value in the hash. No better and perhaps even worse.

# simplified example
use Scalar::Util qw/weaken/;

my $weak_class_ref;

sub BUILD {
    my ($self, $ident, $args_ref) = @_;

    $weak_class_ref ||= {};
    $example_hash_of{$ident}  = $weak_class_ref;
    weaken( $weak_class_ref );
}
[download]

The other idea I had was to have a separate singleton class that you assign into every object during initialization. But that still just moves the problem of cleaning up when no "live" objects exist into the singleton class.

You want common data to spring into existence when you have objects and spring back out of existence when there are no objects. I don't see any alternative to keeping a count of objects or copying the reference over during construction (and thus letting Perl keep the reference count for you).