chore(sync): zsy_developcc 全量同步至 qyqy_develop（W26 口径）

- 分支内容对齐 qyqy_develop b6ec3aa，树完全一致（同步后 git diff 为空）
- 覆盖本轮全部交付：客服 Agent 重构（安全路由 / 五出口 / 记忆与画像 / RAG 全链路）
  + 开发文档 62 份编号体系（D1.1 v1.17 索引）
  + 新增 D2.10-客服Agent端到端答辩文档-2026-09-21.html
- 基线：e239eb7（2026-09-17 品牌口径统一快照），本提交为其直接后继

tests/unit/service/test_model_gateway.py

@@ -147,3 +147,54 @@ async def test_resolve_skips_endpoints_without_declared_capabilities(
         agent_type="customer_service", task_type="chat"
     )
     assert [e.endpoint_code for e in resolved] == ["chat-primary"]
+
+
+@pytest.mark.asyncio
+async def test_resolve_warns_when_multiple_embedding_endpoints_declared(
+    monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture
+) -> None:
+    """声明 `embedding` 能力的 active 端点超过一个时必须告警（配置守卫，只告警不改行为）。
+
+    为什么要守：Milvus 集合的向量维度是固定的，索引侧与查询侧必须用**同一个**模型。
+    两个 embedding 端点共存时，`ModelDispatchService.embed` 只试前 `max(1, max_attempts)`
+    （默认 2）个端点，实际用哪个取决于表的行顺序 —— 一旦索引与查询落到不同模型，
+    `COSINE` 相似度整体失真却**不会报任何错**，表现为“越答越差”的哑故障。
+    """
+    from app.service import model_gateway
+
+    rows = [
+        _endpoint("embedding-primary", ["embedding"]),
+        _endpoint("embedding-shadow", ["embedding"]),
+        _endpoint("chat-primary", ["chat"]),
+    ]
+    monkeypatch.setattr(model_gateway, "SessionFactory", lambda: _FakeScalarSession(rows))
+
+    with caplog.at_level("WARNING"):
+        resolved = await model_gateway.DatabaseModelEndpointResolver().resolve(
+            agent_type="customer_service", task_type="embedding"
+        )
+
+    # 只告警不改行为：仍返回全部声明 embedding 的端点，且顺序与表行顺序一致。
+    assert [e.endpoint_code for e in resolved] == ["embedding-primary", "embedding-shadow"]
+    warnings = [r for r in caplog.records if r.levelname == "WARNING"]
+    assert warnings, "多 embedding 端点时必须留下 WARNING 痕迹"
+    assert "embedding-shadow" in warnings[-1].getMessage()
+
+
+@pytest.mark.asyncio
+async def test_resolve_is_silent_with_single_embedding_endpoint(
+    monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture
+) -> None:
+    """恰有一个 embedding 端点（现网配置）时不得产生告警噪音。"""
+    from app.service import model_gateway
+
+    rows = [_endpoint("embedding-primary", ["embedding"]), _endpoint("chat-primary", ["chat"])]
+    monkeypatch.setattr(model_gateway, "SessionFactory", lambda: _FakeScalarSession(rows))
+
+    with caplog.at_level("WARNING"):
+        resolved = await model_gateway.DatabaseModelEndpointResolver().resolve(
+            agent_type="customer_service", task_type="embedding"
+        )
+
+    assert [e.endpoint_code for e in resolved] == ["embedding-primary"]
+    assert [r for r in caplog.records if r.levelname == "WARNING"] == []