This does in deed work and I've found out why. The problem is that Test::Builder duplicates STDOUT and STDERR (comments say so you can change them without affecting Test::More) but it does NOT duplicate the IO layers. The funny thing is has code to do it but it is commented out:

sub _open_testhandles {
    my $self = shift;   
    return if $Opened_Testhandles;
    # We dup STDOUT and STDERR so people can change them in their
    # test suites while still getting normal test output.
    open( $Testout, ">&STDOUT") or die "Can't dup STDOUT:  $!";
    open( $Testerr, ">&STDERR") or die "Can't dup STDERR:  $!";

#    $self->_copy_io_layers( \*STDOUT, $Testout );
#    $self->_copy_io_layers( \*STDERR, $Testerr );
    
    $Opened_Testhandles = 1;
}

sub _copy_io_layers {
    my($self, $src, $dst) = @_;
    
    $self->_try(sub {
        require PerlIO;
        my @src_layers = PerlIO::get_layers($src);

        binmode $dst, join " ", map ":$_", @src_layers if @src_layers;
    });
}
[download]

I've no idea why this is commented out. However, setting utf8 io layer on STDOUT and then calling failure_output to reset Test::More's idea of failure output works - just seems a bit of a hack.