transcoding nearly certainly wrong?

Why is transcoding nearly certain to be wrong with XML?

Or, to put it another way, why not limit the use
of text/xml to XML instances for which transcoding
is certain not to be wrong, and for which US-ASCII
is acceptable (because the XML uses numeric character
references or character entities, is only used to
code a limited schema with numeric data, etc.?)

Limiting the scope is less radical than deprecating.