Date: Mon, 26 Oct 2015 20:49:55 -0400
   From: John Cowan <xxxxxx@mercury.ccil.org>

   I have *not* added Riastradh's proposed fence instruction, since it
   is not found in any existing version of ephemerons.

Not true: it is found in MIT Scheme, and used nontrivially there.  It
is critical in order for SET-EPHEMERON-DATUM! to be useful.  At the
very least, you'll want it in order to write any automatic tests for
ephemerons.

I don't understand your objection to SET-EPHEMERON-DATUM!, by the way.
A concurrent GC needs to handle racing mutation by the mutator anyway.
It is not clear to me why SET-EPHEMERON-DATUM! is special.