Let me first try to summarise the arguments I believe I've heard in this 
debate.

There's agreement that a few applications (such as editors) need access to 
the whole XML character stream.  It's clear that the Infoset is not 
constraining such applications -- they won't be using it.

There's agreement that many applications don't need (or want) to be told 
about some superficial aspects of the character stream (e.g. most 
whitespace and character entity references).

There's disagreement over exactly which properties should be regarded as 
superficial.  This disagreement comes in two flavours:

(a) The Infoset is a useful concept, but it should be amended to (e.g.) 
retain more DTD information, or to discard comments.

(b) Every application will have distinct opinions about what constitutes a 
significant difference, so trying to standardise on one set is doomed to 
failure.


I personally _do_ believe the Infoset to be useful.  The XML world is not 
one where applications exist in a one-to-one relationship with document 
types.  Once a document is out there, it'll be processed by any number of 
general and special purpose tools.  If those tools are not using compatible 
definitions of what constitutes significant content, there'll be 
considerable confusion.  If FunkyML chooses to make attribute order 
significant, then it's going to be a problem if that ordering information 
is not available to XSLT.

I accept that trying to come up with that common definition will cause 
strain.  There's bound to be conflict between the desire for fidelity and 
the desire for simplicity.  But it's worth doing.
--
Cheers,
John