test: pin the tier fallback and the assembly guard to real behaviour

Both were wired but unwitnessed: hardcoding the transport's fallback, or
blanking the source tier the assembly guard reads, left the whole unit
suite green. Cover them where the value is visible -- the bytes on the
wire for nearest-vs-error, and the assembly-time refusal that must name
low as the executable alternative.

tests/unit/test_config.py

@@ -758,6 +758,24 @@ class TestCrossFieldInvariants:
         with pytest.raises(ValueError, match="register_provider"):
             GatewayClient.from_settings(settings, registry=register_provider(mystery))
 
+    def test_source_tier_the_model_cannot_satisfy_fails_at_assembly(self):
+        """守卫必须读**源级档位**,而不只是 `enable_thinking`(设计 §4.2 三层优先级)。
+
+        变异实测: 把 `_guard_thinking` 的 `source_effort=source.reasoning_effort`
+        改成 `None`,全套单测依然全绿(2026-09-05 独立验证查出)。文案里必须出现
+        可执行替代 `low`——glm-5.3 官方关不掉推理,只报"不行"会把人推回
+        `extra_body` 那条绕过治理的老路(issue #20 的成因)。
+        """
+        base = self._base()
+        src = dataclasses.replace(
+            base.sources[0], provider="zhipu", model="glm-5.3", reasoning_effort=Effort.NONE
+        )
+        with pytest.raises(ValueError) as exc:
+            GatewayClient.from_settings(dataclasses.replace(base, sources=(src,)))
+        message = str(exc.value)
+        assert "glm-5.3" in message
+        assert "low" in message
+
     def test_openai_segment_now_assembles_with_the_standard_field(self):
         # 行为变更(2026-09-04): reasoning_effort 是 OpenAI 官方字段而非厂商方言,
         # 经网关的兼容端点收得下,故兜底段不再把"关闭"判为形态未知