mutate: replace_element_with_copy no longer reproducible under a fixed rand seed (source drawn from unordered pointer set)

[domg123]

Since 81.2.5 (#624), replace_element_with_copy draws its copy source from anywhere in the tree rather than just siblings — a great capability for GP (it effectively gives mutate a native subtree-graft move). However, the source node is selected by indexing into ValuesFromTreeData::allNodes, which is an EvaluableNode::ReferenceSetType — an unordered set keyed on node pointers. Its iteration order depends on the nodes' memory addresses, so the same random seed selects a different source node depending on allocation history. mutate is therefore no longer reproducible under set_rand_seed.

Because replace_element_with_copy is part of the default operation-type split, this affects every mutate/mutate_entity call, not just callers that opt into the op.

Repro

Two identical (set_rand_seed "fixed") + (mutate ...) calls, separated only by allocation churn (unrelated mutates whose results are discarded), return different results:

(seq
	(declare (assoc
			mutate_fixed (lambda (seq
				(set_rand_seed "fixed")
				(unparse (mutate
					(parse "(+ (* x x) (sin (* x x)) (- (/ y 2) 7))")
					0.5
					(assoc "symbol" 0.5 "number" 0.5)
					(assoc "replace_element_with_copy" 1.0)))))
		)
		(declare (assoc a (call mutate_fixed))
			;; allocation churn: unrelated mutates whose results are discarded
			(map
				(lambda (mutate
					(parse "(+ (* x x) (sin (* x x)) (- (/ y 2) 7))")
					0.5
					(assoc "symbol" 0.5 "number" 0.5)
					.null))
				(range 0 200))
			(declare (assoc b (call mutate_fixed))
				(print "a:         " a "\n")
				(print "b:         " b "\n")
				(print "identical: " (= a b) "\n")))))

On the 81.3.0 darwin-arm64 release binary (amalgam-st):

a:         (+ (* (- (/ y 2) 7) x x) (- (/ (* x x) 2) 17) (sin (+ (* x x) (- (/ y 2) 7) (sin (* x x)))))
b:         (+ (* (sin (* x x)) x x) (- (/ 2 2) 17) (sin x))
identical: .false

Same script on 81.1.1: identical: .true.

Downstream this shows up as flaky same-seed test failures: which tests fail changes from run to run, since the divergence depends on allocation layout at the moment the op fires.

Suggested fix

In GetStringsFromTree, also collect nodes into a std::vector<EvaluableNode *> in traversal order (keeping the existing set for the already-visited check on cyclic structures), and have ENBISI_replace_element_with_copy index into that vector. Traversal order is structural, so selection becomes deterministic again — and indexing the vector is O(1) instead of the current O(n) walk over the set.

[domg123]

Still reproduces on the 81.4.1 release binary (darwin-arm64): the repro above prints identical: .false on every run. Current main still iterates allNodes in ENBISI_replace_element_with_copy (EvaluableNodeTreeManipulation.cpp ~L1453-1457).