There's really no reason to do this. Carp is a lightweight module.

Although the proponent(s) of this practice have yet to post data to back it up where I recall having seen it, I don't recall you having posted data to support your claim either. Note that this practice predates the following in Carp.pm:

    require Carp::Heavy unless $INC{"Carp/Heavy.pm"};
[download]

Note that I get quite a kick out of the last half of that statement. I'm sure the very first thing that require does is exactly that anyway, so I'd never add such trivial code in an obvious attempt at micro-optimization. But perhaps this whole hubbabaloo is just micro-optimization. (Update: Which brings up my long-standing catch phrase, "Nothing is obvious unless you are overlooking something" -- see ikegami's reply below.)

It is my belief that more than just unfounded attempts are micro-optimization were required to motivate so much reworking of these modules. I could certainly be wrong in that belief, but your say-so isn't enough to convince me otherwise.

Also note that I delay loading Carp.pm for my own reasons that I find few people share. I don't load modules that I likely don't need because I don't like my modules failing in those rare environments when some module that most people presume that everyone has is actually missing. I've certainly run into such environments many times. In fact, checking my soon-to-be-released module, the code I wrote is actually:

BEGIN {
    my $croak;
    sub _croak
    {
        if(  ! $croak  ) {
            if(  eval { require Carp; 1 }  ) {
                $croak= \&Carp::croak;
            } else {
                $croak= sub {
                    die @_;
                };
            }
        }
        return $croak;
    }
}
[download]

Which means that my module works fine even if Carp.pm is nowhere to be found.

Update: I also get a kick out of the fact that most of the code has been moved out of Carp.pm and yet "require Carp" has to read over a ton of POD in order to get to that small bit of code. I'd micro-optimize that as well and move the POD after __END__. (: