The rendering is incorrect
and the test is marked as flaky on Linux or on GPU rendering,
but this should help catch further regressions.

The test is modified to add the following CSS:

    .intro { display: none }
    html #top { margin-top: 0 }

To disable the "scrolling" part of the test.

The reference is modified to match the current rendering of the test:

* Rows 4 and 5 (eyes) have a red background.
* The nose is not quite where it should be.
* Row 12’s height is 3px more than it should be.

(See http://www.webstandards.org/action/acid2/guide/ for row numbers.)

The nose issue seems be related to rounding vs truncating
when pixel-snapping borders, but this is only a guess.