Yeah, it's a length-caching bug

As a test script:

#!/usr/bin/perl
use strict;
use warnings;

use Test::More tests => 8;

{ # Baseline.
   my $s = "\xE8\xAB\x86\x0A";
   utf8::downgrade($s); is(length($s), 4); is($s, "\xE8\xAB\x86\x0A");
   utf8::decode($s);    is(length($s), 2); is($s, "\x{8AC6}\n");
}
{ # Check for length-caching bug.
   my $s = "\xE8\xAB\x86\x0A";
   utf8::upgrade($s); is(length($s), 4); is($s, "\xE8\xAB\x86\x0A");
   utf8::decode($s);  is(length($s), 2); is($s, "\x{8AC6}\n");
}

1;
[download]

1..8
ok 1
ok 2
ok 3
ok 4
ok 5
ok 6
not ok 7
#   Failed test at a.pl line 15.
#          got: '4'
#     expected: '2'
ok 8
# Looks like you failed 1 test of 8.
[download]