fix(intern-decision-serve): code-review fixes

- a failure while building the response (a non-finite number included) is a 500 inside the
  envelope, never a 422 or a render crash outside it
- an engine ValueError keeps its message but is released and raised unchained, like an OOM
- the prompt is built (and the model's own validation run) before the forward
- the row cap is counted before any ordering is built
- /health reads a device name cached at load, so it makes no driver call off the inference thread

services/intern-decision-serve/tests/test_app.py

@@ -464,3 +464,28 @@ def test_the_app_uses_the_executor_it_is_given():
     assert client.post("/decide", json=ROW, headers=AUTH).status_code == 200
     assert engine.threads == {home}
     executor.shutdown()
+
+
+class InfiniteEngine(FakeEngine):
+    """A server-side defect: a probability of +inf, which breaks the averaging arithmetic."""
+    def predict(self, request):
+        response, sha = super().predict(request)
+        for answer in response["answers"].values():
+            first = next(iter(answer["probabilities"]))
+            answer["probabilities"][first] = float("inf")
+        return response, sha
+
+
+def test_a_failure_while_building_the_response_is_500_never_a_422():
+    response = make_client(InfiniteEngine()).post("/decide", json={**ROW, "orderings": "rotations"}, headers=AUTH)
+    assert response.status_code == 500
+    assert response.json()["error"]["code"] == "scoring_failed"
+
+
+def test_a_huge_orderings_request_is_refused_with_its_true_row_count():
+    engine = FakeEngine()
+    body = {"state": "s", "decisions": [{**d, "orderings": "rotations"} for d in decisions(3000, opts(16))]}
+    response = make_client(engine, max_body_bytes=16 * 2**20).post("/decide/shared", json=body, headers=AUTH)
+    assert response.status_code == 422
+    assert "48000 rows" in response.json()["error"]["message"]
+    assert engine.calls == []