In the attached test case, the fieldset has a minwidth that is bigger than other browsers. Clear whitespace breaking opportunities exist, and yet they are somehow not being factored into the minwidth. This is a regression caused by http://bugs.webkit.org/show_bug.cgi?id=8438 although I think it just exposed a min width bug.
Created attachment 14457 [details] Reduction Reduction.
<rdar://problem/5146757>
Created attachment 14459 [details] This fixes it. Any text run that gets thoroughly ignored but that contains only stripped out whitespace will have beginWS and endWS set to true. Combine that with autoWrap and you know when a break opportunity exists.
Created attachment 14460 [details] Better patch. After conversation with mitz on IRC, new patch.
Comment on attachment 14460 [details] Better patch. r=me. fix indentation of the first 'if'!
Fixed in r21365.