1.) Do you mean "parse it until there are no more errors"? I thought of that, too. But there can be errors even when the document is complete: when it doesn't follow the DTD, for instance.

2.) Again, I thought of that, too. ;-)
But if the sender forgets this, the server appears to hang and eats more and more memory (as more data comes in).

The best approach seems really to be a content-type including the length at the beginning (before the document itself).

This way it would also be easy to use a different format, e.g. YAML, without having to guess it.