🚀 go-pugleaf

It would certainly work.  By extending it just a little
bit farther, you could handle discontinuous elements by
chaining elements together (let's assume we also validate
STARTREF/ENDREF pairing in the parser so we can have
a single GI name:

<discont id=start endid=next1>...
<discont id=next1 startid=start nextid=next2>
...
<discont id=next2 endid=next3 previd=next2>
...
<discont id=next3 startid=next2>

This is a lot of pointing about though.  But like you said,
there's no barrier to implementing this today with
application-defined validation.  Note too that either
the NEXTID= or the PREVID= attribute could be eliminated.
I've included both for completeness.  Allowing forward
references could be expensive for some applications.

I'm afraid I prefer modifying the content models, if
possible, since it removes the requirement to manage
a lot of IDs.  ID management is probably the single
biggest cost associated with SGML authoring, especially
in dynamic, large-document, hypertext environments like
that emerging within IBM.

Eliot Kimber                      Internet:  drmacro@ralvm13.vnet.ibm.com
Dept E14/B500                     IBMMAIL:   USIB2DK9@IBMMAIL
Network Programs Information Development     Phone: 1-919-543-7091
IBM Corporation
Research Triangle Park, NC 27709