wren runs its model on one BLAS thread

threadpoolctl scopes the limit to wren.act, so other numpy calls in the
engine keep their threads.

backend/tests/flow/test_wren.py

@@ -143,3 +143,25 @@ def test_nothing_is_written_unless_applied_and_sure(fake_wren, params, confidenc
             }
         ]
     }
+
+
+def test_the_model_runs_on_one_blas_thread(fake_wren, monkeypatch):
+    """About twice as fast for a model this small; numpy is loaded by `mlp`."""
+    from threadpoolctl import threadpool_info
+
+    threads: list[int] = []
+    act = fake_wren.act
+
+    def counted(command: str, house: Any) -> tuple[list[Any], float]:
+        threads.extend(
+            pool["num_threads"]
+            for pool in threadpool_info()
+            if pool["user_api"] == "blas"
+        )
+        return act(command, house)
+
+    monkeypatch.setattr(fake_wren, "act", counted)
+    _node(("kitchen_lamp", DType.BOOL)).execute({"house.command": "lights on"})
+
+    assert threads
+    assert set(threads) == {1}