I am very hostile towards the idea of requiring UAs to implement any XML parsing features that are in the realm of the XML 1.0 spec but that the XML 1.0 spec does not require. This means processing the DTD beyond checking the internal subset for well-formedness.
That hostility may be justified as far as browser-type UAs go, but I would rather you didn't apply it to server-side and authoring tools.
Those who want to use entities for input, should parse and reserialize as UTF-8 in their own lair and not expose their entity references (or parochial legacy encodings) to the public network.
For those of us writing HTML by hand, this is not a practical solution, particularly when invisible characters are involved. Invisible characters aside, I don't want to go digging through a Unicode character map every time I want → or τ.
~fantasai
