Regexes ARE slow (of course everyone here knows). To the extent I can trust my Strawberries and/or my crippled Kaby Lake to benchmark anything, the next is twice as fast (+ I guess rnkeysort should have been used (and it's slower)):

sub keysort1 {
   my @sorted = rnkeysort { substr $_, rindex( $_, '%' ) - 3, 3 } @uns
+orted;
}
[download]

And assuming $s keeps the (un|pre)split input, and if each 2nd '%' is guaranteed to be an anchor, even if records have different length/layout, then the next unclean ugly one is still faster than keysort for me:

sub test1 {
    my $i = my $j = 0;
    my @nums;
    $j ^= 1 or push @nums, substr $s, $i - 3, 3
        while -1 != ( $i = index $s, '%', $i + 1 );
    my @sorted = @unsorted
        [ sort { $nums[$b] <=> $nums[$a] } 0 .. $#nums ]
}

__END__

          Rate keysort   test1
keysort 21.6/s      --    -30%
test1   30.6/s     42%      --
[download]