The web is based on de facto specs, requiring a lot of investigation to find out what exactly the platform does. One question about the de facto behavior is being reraised by code-splitting efforts: just how bad is synchronous XHR, if used in a place that the application may as well pause anyway? This question comes up because you can't use GWT's code-splitting approach without good static analysis. You can use dynamic analysis, but whenever the dynamic analyzer guesses wrong, the system must fall back on synchronous XHR.
Mark S. Miller sent me a link to Mark Pruett, who did some actual experiments to see what happens in practice. Pruett concludes that all browsers but Firefox 2 are fine.
Kelly Norton is less sanguine. While he likes Opera's behavior reasonably well, he's unsure about IE 6, and he thinks the Safari discussion is inaccurate. It's not clear to what extent browsers are going to mimic Opera.
Overall, I come away thinking that synchronous XHR is reasonable for code splitting so long as the dynamic analyzer is only very infrequently wrong. The app will freeze when it happens, which is bad, but it should be infrequent. Further, the development effort will need to put time into setting up a suite of interaction cases to feed to the dynamic analyzer, and that will consume time. I guess extra time is expected, though, if you want better results.