Compare per-iteration seeds as a set in the parallel-generation test

Samples within an iteration are generated concurrently, so the order in which the
fake generator records their seeds is not deterministic. The test compared the
recorded list directly and failed intermittently (about 1 run in 12 locally).

Compare each iteration's seeds as a sorted group, which still asserts that the same
seeds are reused across iterations.

tests/test_agentic_upsampling.py

@@ -434,7 +434,9 @@ def test_runner_can_disable_early_stop_and_select_best_sample(tmp_path: Path) ->
 
     result = runner.run_item(_item("8", "exactly 12 balloons with exact color counts"))
 
-    assert generator.seeds == [1000, 1001, 1002, 1000, 1001, 1002]
+    # Samples are generated concurrently, so compare each iteration's seeds as a set.
+    assert sorted(generator.seeds[:3]) == [1000, 1001, 1002]
+    assert sorted(generator.seeds[3:]) == [1000, 1001, 1002]
     assert rewriter.previous_scores == [9.0]
     assert result["best_iteration"] == 1
     assert result["best"]["selected_sample_index"] == 1