I've got a statistic for everything, and while I am now unable to measure what happens if I remove my Win32 demo (as my affiliates are spreading that about too) I can tell you that of the 35% of users who were asked to download OpenGL drivers to make the game work, only 1% actually bother.
Seems thats a more scary problem, that people can't be bother to download GL drivers.
So: faced with the cold hard stats, could anyone possibly say with a straight face that Webstart is currently a good deployment mechanism? And given that you'd lose a lot of credibility by trying to argue otherwise, what would the collective's easy quick-fix interim solution consist of?
I still don't quite go with your stats that webstart is bad because more people downloaded your win32 distribution, but that aside your logic is sound, clicking on a JNLP link and being faced with nothing but a XML page is utterly useless.
As above, the solution would seem to be to have a native plugin which orchestrates the download itself (via JNLP or what ever other method you'd like to reinvent). Better still would to have the plugin available on every platform before you get there (e.g. ship it with the machine) but that doesn't seem possible since most machines are already shipped

. I suppose it would be possible to write a deployment app in Flash if you really wanted to?
At the end of the day tho, the only thing that seems to make _your_ end customers happy is to have a tiny download. If thats the case then a binary distribution of just the bits you need seems to be the only thing you can do.
Cas, out of interest, why did you consider webstart knowing that it would require you users to get a JRE? You already has a JET compiled small binary didn't you? EDIT: Oh, I only every played the webstart version, was there an installer with the Win32 version?
Kev