[P1][Runtime Identity] Do not claim an inactive inference engine is currently serving

[ShuhaoZhangTony]

现象

在 sage-mate-vllm-engine.service 明确为 inactive、公开 /health 的 llm_status 为 not_checked 时，运行时身份问答仍回答“当前后端实际提供的是 DeepSeek-V4”，并把 deployment lock 当作当前在线服务证据。

影响

状态问答会把期望/锁定配置误报为实时可用状态，破坏运维可信度，也使部署后对话验收产生假阳性。

验收标准

• 运行时身份将 configured/locked identity 与 live serving availability 分开。
• live engine 不可达时不得使用“当前实际提供/正在运行”等措辞；应明确模型是锁定目标且引擎当前不可用或未验证。
• used_model、Support、health 与回答使用同一 availability snapshot。
• 覆盖 live healthy、live unavailable + lock present、lock absent、probe timeout/unknown。
• 合并 main、部署后，以真实 inactive engine 对话验证回答不再误导且仍返回引用。

发现来源

Ascend CI runner #8 合并后的强制对话验收。

[ShuhaoZhangTony]

Post-merge acceptance completed on main commit 70c21be (v4.6.32).

Evidence:

• PR fix: distinguish configured and live runtime identity #28 merged after all 5 CI checks passed.
• Public health: runtime_identity_status=receipt, runtime_serving_available=false; engine service remains inactive.
• Real public dialogue returned HTTP 200 in 2.4 s with 1 runtime evidence hit and 1 Support basis item labeled 部署配置与可用性.
• The answer explicitly says 部署锁定目标 and 不代表引擎当前正在提供推理.
• used_model=runtime-identity-provider; no offline model was falsely attributed.
• Full operational wording matrix: 24/24 passed, reference coverage 100%, max contradiction score 0.