修复数据链路 P0 瓶颈：缓存炸弹、Context 重渲染、LGBM 循环、TTL 不匹配

P0-1 sessionStorage 限制：
- writeCityDetailCacheBundle 只保留最近 3 个城市的详情
- 避免 3-10MB JSON 序列化阻塞主线程

P0-2 Context 拆分：
- 新增 CityDetailsContext 独立管理 cityDetailsByName 变更
- 新增 useCityDetails hook，只订阅详情的组件不再因 cities/proAccess 变化重渲染
- DashboardStoreContext 保持不变，向后兼容

P0-3 扫描终端 TTL：
- SCAN_TERMINAL_PAYLOAD_TTL_SEC 30s → 120s
- 匹配 ThreadPoolExecutor(4) x 60 城的实际重算耗时

P0-4 LGBM 循环依赖：
- LGBM 预测值不再作为 DEB 输入参与权重计算
- 保留为独立参考字段 lgbm.prediction 输出给前端展示
- 消除 DEB → LGBM 训练 → DEB 的循环

新增 docs/data-architecture-review.md 完整数据链路审查报告

Tested: npx tsc --noEmit, python -m ruff check .

frontend/lib/dashboard-client.ts

@@ -542,15 +542,19 @@ export const dashboardClient = {
     meta: Record<string, CityCacheMeta>,
   ) {
     if (!isClient()) return;
+    // Keep only the 3 most-recently-accessed cities to prevent sessionStorage bloat
+    const MAX_CACHED_CITIES = 3;
+    const allEntries = Object.entries(details).map(([cityName, detail]) => ({
+      cityName,
+      cachedAt: meta[cityName]?.cachedAt || 0,
+      detail,
+      revision: meta[cityName]?.revision || getCityRevision(detail),
+    }));
+    const topEntries = allEntries
+      .sort((a, b) => b.cachedAt - a.cachedAt)
+      .slice(0, MAX_CACHED_CITIES);
     const entries = Object.fromEntries(
-      Object.entries(details).map(([cityName, detail]) => [
-        cityName,
-        {
-          cachedAt: meta[cityName]?.cachedAt || Date.now(),
-          detail,
-          revision: meta[cityName]?.revision || getCityRevision(detail),
-        },
-      ]),
+      topEntries.map((e) => [e.cityName, { cachedAt: e.cachedAt, detail: e.detail, revision: e.revision }]),
     );
     window.sessionStorage.setItem(
       CACHE_KEY,