fix: enforce registered auto reasoning capabilities

tests/unit/test_providers.py

@@ -57,14 +57,10 @@ class TestDefaultProfiles:
             assert w.off == {"reasoning_effort": "none"}, name
             assert w.effort_key == "reasoning_effort", name
 
-    def test_minimax_on_tier_carries_a_tier_value(self):
-        """issue #21 的权宜之计: minimax 的"开"必须真写一个档位值,不能是空片段。
-
-        断言反复过一次: T2 按"这些模型默认就推理"的推定把它改成 `{}`,T10 真实
-        网关实测推翻推定(M3 不发推理参数时 5/5 轮不推理),故逐字恢复旧版的 medium。
-        """
+    def test_minimax_on_does_not_select_a_tier(self):
+        """形态不代替模型能力，也不替调用者选择付费档位。"""
         w = get_provider("minimax").thinking
-        assert w.on_base == {"reasoning_effort": "medium"}
+        assert w.on_base == {}
         assert w.off == {"reasoning_effort": "none"}
         assert w.effort_key == "reasoning_effort"
 

tests/unit/test_thinking.py

@@ -285,16 +285,13 @@ class TestResolveThinking:
                 get_provider("zhipu"), cap, Effort.NONE, model="glm-5.3", fallback="nearest"
             )
 
-    def test_phase4_only_blocks_the_off_direction(self):
-        """关不掉 ≠ 开不了: M2.x 默认就在推理,开的方向不该被拦。
-
-        期望片段 2026-09-05 由 `{}` 改成 minimax 的 `on_base` 实际值: issue #21 把
-        该段的"开"改回带 medium(T2 的"开档不注入"是推定,T10 实测推翻)。本用例守的
-        是 Phase 4 只拦关闭方向,注入什么由 wire 决定,故随 wire 走。
-        """
-        cap = get_capability("MiniMax-M2.7")
-        got = resolve_thinking(get_provider("minimax"), cap, Effort.AUTO, model="MiniMax-M2.7")
-        assert got.payload == {"reasoning_effort": "medium"}
+    @pytest.mark.parametrize("model", ["MiniMax-M2.5", "MiniMax-M2.7"])
+    def test_phase4_only_blocks_the_off_direction(self, model):
+        """已登记 AUTO 只发开启片段，不由库代选 medium。"""
+        got = resolve_thinking(
+            get_provider("minimax"), get_capability(model), Effort.AUTO, model=model
+        )
+        assert got.payload == {}
         assert got.applied_effort is Effort.AUTO
 
     def test_phase4_passes_when_none_is_registered(self):
@@ -351,17 +348,42 @@ class TestResolveThinking:
         assert got.payload == {"thinking": {"type": "enabled"}, "reasoning_effort": "max"}
         assert got.applied_effort is Effort.MAX
 
-    def test_auto_never_trips_phase5(self):
-        """`auto` = 不指定档位,可满足性只取决于 wire 有没有 on_base。
+    @pytest.mark.parametrize(
+        "provider,model", [("deepseek", "deepseek-v4-pro"), ("minimax", "MiniMax-M3")]
+    )
+    @pytest.mark.parametrize("fallback", ["error", "nearest"])
+    def test_unregistered_auto_choice_is_rejected(self, provider, model, fallback):
+        """有开启形态也不代表已登记模型支持 AUTO，nearest 不可代选。"""
+        with pytest.raises(ThinkingUnsupportedError) as exc:
+            resolve_thinking(
+                get_provider(provider),
+                get_capability(model),
+                Effort.AUTO,
+                model=model,
+                fallback=fallback,
+            )
+        assert model in str(exc.value)
+        assert "auto" in str(exc.value)
+        assert "EFFORT_FALLBACK=nearest" not in str(exc.value)
 
-        它不是写进 `effort_key` 的取值,故不受档位清单约束。反过来判会让存量的
-        `ENABLE_THINKING=true`(T5 起等价于 auto)在 deepseek/glm-5.3 这类清单里
-        没有 auto 的模型上当场报错——设计 §12 明确承诺存量配置继续可跑。
-        """
-        cap = get_capability("deepseek-v4-pro")  # (none, high, max),清单里没有 auto
-        got = resolve_thinking(get_provider("deepseek"), cap, Effort.AUTO, model="deepseek-v4-pro")
-        assert got.payload == {"thinking": {"type": "enabled"}}
+    def test_minimax_explicit_medium_restores_old_wire(self):
+        got = resolve_thinking(
+            get_provider("minimax"), get_capability("MiniMax-M3"), Effort.MEDIUM, model="MiniMax-M3"
+        )
+        assert got.payload == {"reasoning_effort": "medium"}
+        assert got.applied_effort is Effort.MEDIUM
+
+    @pytest.mark.parametrize("provider", ["openai", "qwen"])
+    def test_unknown_auto_warns_without_promising_effect(self, provider):
+        messages, sink = _warnings()
+        try:
+            got = resolve_thinking(
+                get_provider(provider), None, Effort.AUTO, model="unregistered-model"
+            )
+        finally:
+            logger.remove(sink)
         assert got.applied_effort is Effort.AUTO
+        assert any("不保证" in str(message) for message in messages)
 
     # —— nearest 映射(fallback 的逃生口)——