fix: reject conflicting raw reasoning overrides before sending

tests/unit/test_openai_compat.py

@@ -623,8 +623,8 @@ class TestThinkingReconciliation:
         try:
             await _complete(transport, self._minimax(False))
             await _complete(transport, self._minimax(False))
-            await _complete(transport, self._minimax(True))
-            await _complete(transport, self._minimax(True))
+            await _complete(transport, self._minimax(True), reasoning_effort=Effort.MEDIUM)
+            await _complete(transport, self._minimax(True), reasoning_effort=Effort.MEDIUM)
         finally:
             logger.remove(sink_id)
         hits = [m for m in messages if "MiniMax-M3" in m]
@@ -705,75 +705,44 @@ class TestNonStreamFastPath:
 
 
 class TestRequestShaping:
-    @pytest.mark.parametrize(
-        ("enable_thinking", "expected"),
-        [(True, {"enable_thinking": True}), (False, {"enable_thinking": False}), (None, {})],
-    )
-    async def test_thinking_tri_state_injection(self, enable_thinking, expected):
+    @pytest.mark.parametrize("tier", [Effort.MEDIUM, Effort.NONE])
+    async def test_minimax_explicit_tier_is_sent(self, tier):
         seen = {}
 
         def handler(request):
             seen.update(json.loads(request.content))
             return _sse_stream(_chunk(content="x"), _chunk(usage=_USAGE))
 
-        await _complete(_transport_for(handler), _source(enable_thinking=enable_thinking))
-        assert {k: seen[k] for k in expected} == expected
-        if enable_thinking is None:
+        transport = _transport_for(handler)
+        try:
+            result = await _complete(
+                transport, _source(provider="minimax", model="MiniMax-M3"), reasoning_effort=tier
+            )
+            assert seen["reasoning_effort"] == tier.value
+            assert result.applied_effort is tier
             assert "enable_thinking" not in seen
-        assert seen["stream_options"] == {"include_usage": True}
+        finally:
+            await transport.aclose()
 
-    @pytest.mark.parametrize(
-        ("enable_thinking", "expected"),
-        # 本条断言反复过一次,记下原委以免第三次改回去:
-        # T2(2026-09-04)按"MiniMax 开启档本就无需参数"的**推定**把 medium 改成不注入;
-        # T10(2026-09-05)真实网关实测推翻该推定——M3 不发任何推理参数时 5/5 轮不推理,
-        # 故 medium 回归(issue #21 的权宜之计,正解是让 auto 受能力表约束)
-        [(True, "medium"), (False, "none")],
-    )
-    async def test_minimax_injects_reasoning_effort(self, enable_thinking, expected):
-        """issue #5: MiniMax 认的是 reasoning_effort,不是 enable_thinking。"""
+    async def test_raw_only_keeps_source_then_request_priority(self):
         seen = {}
 
         def handler(request):
             seen.update(json.loads(request.content))
             return _sse_stream(_chunk(content="x"), _chunk(usage=_USAGE))
 
-        source = _source(
-            name="mm", provider="minimax", model="MiniMax-M3", enable_thinking=enable_thinking
-        )
-        await _complete(_transport_for(handler), source)
-        if expected is None:
-            assert "reasoning_effort" not in seen
-        else:
-            assert seen["reasoning_effort"] == expected
-        assert "enable_thinking" not in seen  # 旧形态实测被静默丢弃,不再下发
-
-    async def test_extra_body_overrides_the_profile_slot(self):
-        """注入顺序即优先级: profile → extra_body → overlay,两行不可调换。
-
-        固定用 **zhipu + glm-5.3 + 源级 low** 这组: 判据必须落在一个 profile
-        **真的写了值**的键上,两边写同一个键才谈得上谁覆盖谁。不挑 minimax 是因为
-        它的 `on_base` 只写 `reasoning_effort` 一个键(issue #21 的权宜之计),
-        覆盖发生后看不见"profile 独有的那半边仍在",判据少一半。
-        """
-        seen = {}
-
-        def handler(request):
-            seen.update(json.loads(request.content))
-            return _sse_stream(_chunk(content="x"), _chunk(usage=_USAGE))
-
-        source = _source(
-            name="zp",
-            provider="zhipu",
-            model="glm-5.3",
-            reasoning_effort="low",
-            extra_body={"reasoning_effort": "high"},
-        )
-        await _complete(_transport_for(handler), source)
-        # profile 注入的是 low,extra_body 后写故发出去的是 high;顺序一调换就变 low,
-        # 即下游写在 extra_body 里的覆盖被库悄悄顶掉(issue #20 的成因形态)
-        assert seen["reasoning_effort"] == "high"
-        assert seen["thinking"] == {"type": "enabled"}  # profile 独有的那半边仍在
+        transport = _transport_for(handler)
+        try:
+            result = await _complete(
+                transport,
+                _source(extra_body={"reasoning_effort": "low", "temperature": 0}),
+                overlay={"reasoning_effort": "high", "temperature": 1},
+            )
+            assert seen["reasoning_effort"] == "high"
+            assert seen["temperature"] == 1
+            assert result.applied_effort is None
+        finally:
+            await transport.aclose()
 
     async def test_model_that_cannot_disable_is_rejected_not_silently_ignored(self):
         """M2.x 关不掉推理: 必须是四分类之一的 RequestRejected,不是裸 ValueError。
@@ -1056,3 +1025,86 @@ class TestLifecycle:
         await _complete(transport, _source())
         await transport.aclose()
         await transport.aclose()
+
+
+class TestManagedReasoningOwnership:
+    """通过真实 transport 验证拒绝发生在 HTTP 之前。"""
+
+    @pytest.mark.parametrize(
+        "raw",
+        [
+            {"reasoning_effort": "high"},
+            {"enable_thinking": True},
+            {"thinking": {}},
+            {"thinking_budget": 100},
+            {"reasoning": {}},
+            {"thinkingConfig": {}},
+            {"output_config": {"effort": "low"}},
+        ],
+    )
+    @pytest.mark.parametrize("layer", ["source", "request", "shadowed"])
+    async def test_raw_control_is_rejected_before_http(self, raw, layer):
+        sent = []
+
+        def handler(request):
+            sent.append(request)
+            return _sse_stream(_chunk(content="ok"), _chunk(usage=_USAGE))
+
+        source = _source(
+            provider="openai",
+            reasoning_effort=Effort.HIGH,
+            extra_body=raw if layer != "request" else {},
+        )
+        overlay = raw if layer != "source" else {}
+        transport = _transport_for(handler)
+        try:
+            with pytest.raises(RequestRejectedError):
+                await _complete(transport, source, overlay=overlay)
+            assert sent == []
+        finally:
+            await transport.aclose()
+
+
+class TestDefaultClientFactory:
+    """直接检查生产 factory，不用取证测试的另一套 factory 代替。"""
+
+    @pytest.mark.parametrize("key", ["fake-a", "fake-b"])
+    async def test_authorization_uses_source_api_key(self, key):
+        from polygateway.transports.openai_compat import _default_client_factory
+
+        client = _default_client_factory(_source(api_key=key))
+        try:
+            assert client.headers["Authorization"] == f"Bearer {key}"
+            assert (
+                client.build_request("POST", "https://gw.example/v1/chat/completions").headers[
+                    "Authorization"
+                ]
+                == f"Bearer {key}"
+            )
+        finally:
+            await client.aclose()
+
+    @pytest.mark.parametrize("timeout", [17.0, 53.0])
+    async def test_timeout_uses_source_timeout_for_all_phases(self, timeout):
+        from polygateway.transports.openai_compat import _default_client_factory
+
+        client = _default_client_factory(_source(timeout_s=timeout))
+        try:
+            assert [
+                client.timeout.connect,
+                client.timeout.read,
+                client.timeout.write,
+                client.timeout.pool,
+            ] == [timeout] * 4
+        finally:
+            await client.aclose()
+
+    @pytest.mark.parametrize("trust_env", [True, False])
+    async def test_trust_env_uses_source_setting(self, trust_env):
+        from polygateway.transports.openai_compat import _default_client_factory
+
+        client = _default_client_factory(_source(trust_env=trust_env))
+        try:
+            assert client.trust_env is trust_env
+        finally:
+            await client.aclose()