Let 'RA' be some registration service.
There's a list of experimental features. RA registers the features.
A feature registration gives a 'feature name' and the HTML of the
document describing the feature. (Not just the URL, the entire HTML
documentation, so that stuff doesn't disappear.)
There's a list of browsers/browser strings. RA registers the browsers,
and, for each browser string, the list of features supported.
http://RA/features?<browser>
returns a text/plain list of the features supported for <browser>.
Web servers might have to contact RA if they come across a browser
they've never seen before to find out what features it supports, but
the server could cache that information.
This database is pretty small. Maybe popular sites wouldn't mind
replicating it.
Hey, I just had another idea: Maybe the browser string could have the
URL of the browser's feature list?