LSID data does not change meaning it always has the same bit pattern for a given LSID. Since XML allows different bit level expressions for equivalent records there is a mismatch with the LSID mechanism.
There is only a mismatch if you try to return XML as LSID "data". I don't see any reason to do this, unless the XML file *is* the object to which the LSID is applied (as opposed to the object that the XML content attempts to describe, such as a specimen or a taxon name). If, for some reason, someone would want to encapsulate an XML file as the LSID-identified "data", then you would have to do it in a way that "locked in" the bytestream of the XML in a way that is bit-level persistent.
The community can live with this as long as there are additional constraints put on the generation of XML-based records. For the sake of simplicity keep the fixed bitlevel exppression. The existing metadat mechanism handles the semantics of interpretation of the data (sorry to use the word "semantics" but it is nothing really special, just a definition of the "meaning" of the data)
No problem on the use of "sematics", because it's clear which meaning you intended from the context of how you used it (i.e., the semantics of the word semantics was not opaque... :-) )
Aloha, Rich