Bad practice: Overriding HTTP content-type with a URI reference

In the last call working draft of SGRS,
http://www.w3.org/TR/speech-grammar/#S2.2.2
a reference to another document can specify the content type of the 
destination object in a way that overrides any content-type provided by 
the HTTP server . The same was true  of (svg? smil?).  I understand that 
the intent was specifically to be able to reference a lot of existing 
data in legacy (probably unregistered) mime types.

This is obviously counter to the architecture, but meets a real need
in practice for servers which just can't be configured by the people
who publish on them.  This is fact a social/tools mess - if the
config files took local directory input then all would be well.

Maybe a compromise is to only allow the link to specify the
content-type when the server is FTP (or something else
with no content-type control) or the HTTP server returns
text/plain or octet-steam, which seem to be used for
"don't know" types.

It is of course consistent to indicate "the contents of this document 
expressed in whatever language but then reinterpreted as 
application/whatever" but in practice such references are kludges and 
would for example make the introduction of new versions of 
application/whatever more difficult.

Tim