Good question. In theory you could use same_term/2 to locate the goal inside the term as provided by prolog_load_context(term, Term). That is all rather hacky though. The fact that the compiler toplevel is a failure driven loop could also be used to maintain a global variable to keep track of the terms processed. I doubt it is a good idea to forever commit to the failure driven approach though.
So, what else? I have some really early thoughts about wrapping at a low level. Not sure whether and how that will see daylight. I might need something like that to enhance the tabling implementation.
We could also rewrite expand_goal/2 to block expanding the same goal. So, if goal_expansion/2 succeeds for X, we go down into the expanded term but will not pass X to goal expansion again should we find it.
I had a look at the SICStus docs. That also suggests using a naive fixed point approach. Any systems that have a solution?