What is the idiomatic approach to this class of problems?
Is there a way to simplify the problem?
Many of the following remarks could be added to many programs here on SO.
Imperative names
Every time, you write an imperative name for something that is a relation you will reduce your understanding of relations. Not much, just a little bit. Many common Prolog idioms like append/3 do not set a good example. Think of append(As,As,AsAs). The first argument of find_min(Min, As) is the minimum. So minimum_with_nodes/2 might be a better name.
findall/3
Do not use findall/3 unless the uses are rigorously checked, essentially everything must be ground. In your case it happens to work. But once you generalize foo/2 a bit, you will lose. And that is frequently a problem: You write a tiny program ; and it seems to work.
Once you move to bigger ones, the same approach no longer works. findall/3 is (compared to setof/3) like a bull in a china shop smashing the fine fabric of shared variables and quantification. Another problem is that accidental failure does not lead to failure of findall/3 which often leads to bizarre, hard to imagine corner cases.
Untestable, too specific program
Another problem is somewhat related to findall/3, too. Your program is so specific, that it is quite improbable that you will ever test it. And marginal changes will invalidate your tests. So you will soon give up to perform testing. Let’s see what is specific: Primarily the foo/2 relation. Yes, only an example. Think of how to set up a test configuration where foo/2 may change. After each change (writing a new file) you will have to reload the program. This is so complex, chances are you will never do it. I presume you do not have a test harness for that. Plunit for one, does not cover such testing.
As a rule of thumb: If you cannot test a predicate on the top level you never will. Consider instead
minimum_with(Rel_2, Min, Els)
With such a relation, you can now have a generalized xfoo/3 with an additional parameter, say:
xfoo(o, A,B) :-
foo(A,B).
xfoo(n, A,B) :-
newfoo(A,B).
and you most naturally get two answers for minimum_with(xfoo(X), Min, Els). Would you have used findall/3 instead of setof/3 you already would have serious problems. Or just in general: minmum_with(\A^B^member(A-B, [x-10,y-20]), Min, Els). So you can play around on the top level and produce lots of interesting test cases.
Unchecked border cases
Your version 3 is clearly my preferred approach, however there are still some parts that can be improved. In particular, if there are answers that contain variables as a minimum. These should be checked.
And certainly, also setof/3 has its limits. And ideally you would test them. Answers should not contain constraints, in particular not in the relevant variables. This shows how setof/3 itself has certain limits. After the pioneering phase, SICStus produced many errors for constraints in such cases (mid 1990s), later changed to consequently ignoring constraints in built-ins that cannot handle them. SWI on the other hand does entirely undefined things here. Sometimes things are copied, sometimes not. As an example take:
setof(A, ( A in 1..3 ; A in 3..5 ), _) and setof(t, ( A in 1..3 ; A in 3.. 5 ), _).
By wrapping the goal this can be avoided.
call_unconstrained(Goal_0) :-
call_residue_vars(Goal_0, Vs),
( Vs = [] -> true ; throw(error(representation_error(constraint),_)) ).
Beware, however, that SWI has spurious constraints:
?- call_residue_vars(all_different([]), Xs).
Xs = [_A].
Not clear if this is a feature in the meantime. It has been there since the introduction of call_residue_vars/2 about 5 years ago.