Sorry to answer my own quesion, but it seems like it was videocard/firefox rendering related (I have one of comparatively old Matrox videocards)
I narrowed it to a simple fragment (referencing only a single background image). It happens that for three properties at once
- body background
- floating element
- hovering with border-bottom
firefox repaints the whole float element no matter how big it is (for stackexchange it's the full questions list) when the mouse hovers over the links. Supposedly the mouse in my case was frozen because of the matrox videocard bitblitting temporary disabled the mouse. It seems that other videocards not showing the problem just did the blitting fast. The issue that ff repaints all the area can be easily reproducable if you fill the screen with random pixels and try to move mouse afterward seeing which parts are beeing freed from these pixels.
Below is the fragment that reproduces full repaint behavior (in my case when I try to intersect the link with the mouse, it is as though it holds the pointer a litte ("rubber" effect):
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
background:#fff url(bg-site.png) repeat top left;
border-bottom:1px dotted #ab2020;
<a href="/questions/19416/">Is argument or parameter considered more appropriate for a function/method argument/parameter?</a>
update: I'm not sure whether it's a big issue, but actually it means that on all windows machines large parts of the page (at least the whole mini-list with questions) is redrawn when mouse doesn't activate anything, but just moves. When I modified the program to fill the screen randomly on a constant basis, I noticed that ff repaints the whole page on every key press, even here, on meta.stackoverflow.com. Probably time to send a bug report to mozilla's bugzilla.
update interesting that although the users of modern fast videocards usually don't see anything wrong, they still can suffer a little when the site is accessed on a notebook working from battery. The reason is that full repainting consumes more resources than partial. You can see this effect just by opening the Task manager and looking at ff cpu% when moving in and out of the link to a quesion on stackexhange sites. With ATI-bundled notebook I see almost full core consumption for stackexchange and about 20% for stackoverflow