I did try that in my initial experimentation, and I revisited it again after your suggestion, but still no dice... after I redirected STDERR to a scalar I ended up with nothing there. I did this:

    my $test_output;
    my $test_err;
    tie(*STDOUT, 'IO::Scalar', \$test_output);
    tie(*STDERR, 'IO::Scalar', \$test_err);
    eval {
        runtests(@testfiles);
    };
    $test_output .= $@ if ($@);
    untie(*STDOUT);
    untie(*STDERR);
    $test_output .= $test_err;
[download]

$test_output was no different than it was before. No failed test report.