The lefthand side of the above page is decoded as the default charset instead of gb2312. The righthand side is correctly decoded as gb2312. The lefthand side is loaded into an iframe in the page. The document loaded into the iframe at http://star.vnet.cn/Comments/three.asp?ChannelID=116&ClassID=170 does not specify charset anywhere (html meta or http). On the other hand, its parent document has meta charset declaration. In this scenario, usually inheriting charset from its parent helps.
We do not inherit the charset from parent frame to make Google Images (and Google Cache) work better, see bug 6118. This also fixes sites that use User-Agent sniffing to choose an encoding (e.g. the main frame's server may send x-mac-cyrillic to Safari, while a subframe may rely on the browser default being windows-1251, as that's the Windows Cyrillic encoding). We also match Firefox here. So, although it's a regression from shipping Safari/WebKit for this site, I think it's an evangelism issue.
One way to fix this site without breaking the cases I mentioned above would be to inherit charsets specified in META, but not in an actual HTTP header. But it's not clear whether such a change would fix more sites than it would break.
Matching other browsers is more important than matching older versions of Safari, so I think Alexey's right and we want the behavior we have currently.
Moving to evangelism component.
The URL is dead.
Closing now. Note that we have actually changed the behavior to inherit the charset as long as security considerations allow (other browsers did change in the meanwhile, too).