FEAR draws the same geometry twice, the second time using zfunc=equal.
In both cases it sets a huge depth bias of -0.5, presumably to get
better precision for the fragile Z comparison. The GL polygon offset we
set ends up being so large that it pulls the geometry into the negative
Z range. It isn't clipped (or no longer, older NV drivers probably had a
separate bug there), but the Z value gets clamped to 0.0 in the first
draw and doesn't match the incoming Z in the second draw.