Very nice solution. I was also thinking of a succession of ternary operators to handle the undef or "null" special cases, but did not want in my earlier post to decide on those cases without the OP specifying the required behavior for those cases (which is why I suggested an example sorting on another field). Having said that, if the idea is just to sort the (final) sub-arrays (as I understood the OP's requirement), then the two calls to the map function seem unnecessary. We could have just this (using your test data):

use strict;
use warnings;

my %data = (
    A => [ 0, 1, 2, 3, 99,      5, 6 ],
    B => [ 0, 1, 2, 3, 9.9999,  5, 6 ],
    C => [ 0, 1, 2, 3, "",      5, 6 ],
    D => [ 0, 1, 2, 3, undef,   5, 6 ],
    E => [ 0, 1, 2, 3, 11,      5, 6 ],
    F => [ 0, 1, 2, 3, 123.0,   5, 6 ],
    G => [ 0, 1, 2, 3, '',      5, 6 ],
    H => [ 0, 1, 2, 3, undef,   5, 6 ],
    I => [ 0, 1, 2, 3, -11,     5, 6 ],
    J => [ 0, 1, 2, 3, -1.1,    5, 6 ],
);

my @rows = sort {
    (! defined $a->[4]) ? -1 :
    (! defined $b->[4]) ?  1 :
    (! length $a->[4])  ? -1 :
    (! length $b->[4])  ?  1 :
    ($a->[4] <=> $b->[4])
} values %data;
[download]

But here again, we don't really know what the OP exactly wants to get.