Keep provider reads behind deterministic evidence guards

DeepSeek can return a polished city forecast that conflicts with already-computed stale-observation, observed-break, or peak-window evidence. The completion path now carries the deterministic fallback guard state and overwrites only the critical fields when provider text or numbers contradict those local facts.

Constraint: Airport-read latency optimization keeps provider output narrow, so backend completion remains the authority for final highs and evidence conflicts.

Rejected: Trust provider final wording when present | it can reintroduce stale METAR anchors or miss observed high breaks.

Confidence: high

Scope-risk: narrow

Tested: pytest tests/test_web_observability.py -q

Tested: npm run build

CHANGELOG.md

@@ -8,6 +8,7 @@
 - 城市决策卡兜底判断新增实测突破识别：当最新 METAR/观测已高于 DEB 中枢或模型上沿时，改为提示最高温中枢需要上修
 - 城市决策卡兜底判断补充实测偏低和峰值窗口已过分支：峰后未追上模型时提示下修压力，峰前偏低时只提示等待确认
 - 城市决策卡新增过旧 METAR/观测识别：过旧报文只作为背景参考，不再触发强实况锚点、上修或下修判断；AI 缓存键同步纳入观测时间与 stale 状态
+- 城市决策卡新增 AI 结果后处理护栏：完整 DeepSeek 返回若与过旧观测、实测突破、峰后下修等确定性证据冲突，会以后端规则覆盖关键数值和结论文案
 - 城市决策卡市场层改用完整 `all_buckets` 并严格识别 exact / range / or higher / or lower 温度桶方向，避免最高温中枢错配到不合理尾部桶
 - 温度桶标签统一规范化 `C/F/°C/°F`，修复 `31°°C` 这类重复单位展示
 - 决策卡展示文案将“概率差”收口为“模型-市场差”，明确口径为 `模型概率 - 市场隐含概率`