If this idea would lead to significant reduction in HTTP overhead then
that's fine. But how do we know this is the case? I know that sometimes
the header (e.g. Mosaic's 1K header) is as big or bigger than the document.
On the other hand, some documents are huge and so the header is insignificant.
It would be good to have real numbers to estimate the expected reduction.
Then we can compare this with the added complexity in the protocol. I
assume this extension would be optional, not mandatory, right? If optional,
how would a client know whether the server supported it?
It also sounds like this is trying to make an end-run around
statelessness, since the idea would really only work if the server
caches the Accept-URI or Header-URI. In otherwords, you've introduced
state into the protocol, but not in a very clean way. Maybe a better
answer is that high performance HTTP (HPHTTP) will have to be stateful.
It might be better to keep HTTP real simple.