the perldoc entry for length (which I checked beforehand to make sure it wouldn't count bytes -- hence my confusion)

It "normally deals in logical characters", but its logic doesn't cover all the intricacies of unicode.

Do you have any specific languages or complex data in mind with which it might fail?

Yes, Thai language is the main one I'm involved with. The modified script below shows that length counts diacriticals in Thai, which may or may not be what is wanted, and is inconsistent with the results for Latin diacriticals in your dataset, which length isn't counting separately. I'm using pre tags so that the Thai will display correctly and shortened lines to facilitate copy/paste.

#!/usr/bin/env perl
use warnings;
use v5.14;
use Unicode::Normalize qw/NFD/;
binmode STDOUT, 'utf8';
binmode DATA, 'encoding(utf-8)';
 
while (<DATA>) {
    chomp;
    print $_, ': ';
    s/[A-Za-z]//g;
    my $alphacount = () = /\p{Alpha}/g;
    say "non-(A-Za-z) symbols <$_>", 
        " contain $alphacount", 
        " alphabetic characters and ",
        getdia($_), " diacritical chars.";
    say "length() thinks there are ", 
        length, " characters\n";
}

sub getdia {
    my $normalized = NFD($_[0]);
    my $diacount = () = 
        $normalized =~ /\p{Dia}/g;
    return $diacount;
}

__DATA__
เป็น
ผู้หญิง
เมื่อวันก่อน
æðaber
æðahnútur
æðakölkun