No, it doesn't. The function gets created by sub before \& is executed.

print "a:", exists(&frobnicate)?1:0, "\n";             # 1

my %OPS = (
     frob => \&frobnicate,
     knob => \&knobnicate,
);

BEGIN { print "b:", exists(&frobnicate)?1:0, "\n"; }   # 0

sub frobnicate { '...' }
sub knobnicate { '...' }
[download]

Hm, so it does. The only non-contrived example I can think of is:

BEGIN {
    print exists(&frobnicate)?1:0, "\n";

    my %OPS = (
        frob => \&frobnicate,     
        knob => \&knobnicate,
    );

    sub runops { print $_->(), "\n" for values %OPS }
}

sub frobnicate { 'frobnicate' }
sub knobnicate { 'knobnicate' }
[download]

I have done it in the past in order to keep the dispatch table visible only to the function using it, but I find the extra indentation rather ugly so I tend to avoid it. I don't know if I'd call it strange and rare, but it certainly isn't unexpected. The need for it is also removed in 5.10, but I mention this only so that someone else won't.

Edit: I realize now this has gotten somewhat far afield of the original question. You asked me to substantiate that if you remove auto-vivification by reference then all subroutines would have to be defined before they can be referred to. It's true, they would need only be declared, not defined, before taking a reference.