release: v1.32.6 — 两周改动深度审查全面修复（回测口径三件套/LLM 安全加固/涨停价舍入/时区统一/缓存与竞态等 58 处）

对 v1.21→v1.32.5 的 249 文件 4.2 万行改动做六路专项审查，本轮落地全部发现：

回测正确性：组合收益 fillna(0) 虚增、轮动停牌日过期价成交、单标的 WF 逐窗指标
被预热区稀释（三件套均带先红后绿回归）；worst_drawdown 方向、grading 容错、
组合体检品种费率、寻优端点费率透传。

安全：LLM api_url 仅 http/https 且禁 userinfo（封死 file:// 读取与 Key 外送链）、
错误响应不回显原始 body、响应体 2MB 上限、配置原子写、坏配置字段级防御。

数据：涨跌停价整数分币舍入（67/318/90 个价位错 1 分漏判清零）、交易时段/采样/
provisional 统一沪时区、warehouse 增量缺口自动全量重拉、provisional 定点转正、
baostock 真故障抛错 + W/M 去 tradestatus（实测服务端报错，周月兜底此前从未工作）
+ 指数 vol 股→手（实测锚定）、ccpm 结构变更抛错。

Web API：缓存键补 count/vipdoc、NaN 清洗先于缓存、count>800 分页取全量、
submit 透传真实状态、pending 不再被淘汰成幽灵、watchlist/server 入参约束。

公式：FILTER 去副作用、0-1 值域误判收严、递归深度上限、REF 负移位显式禁止。

前端：4 处请求竞态序号守卫、Sparkline viewBox、北交所 market=2 映射、
空数据缓存死角、AI 弹窗卸载中止轮询、量能/资金日历口径修正。

CLI/CI：warehouse sync 失败 exit 1、参数校验干净报错、release 真实发布 SHA256、
CI 超时与缓存、spec 补 baostock 前提。

约 60 条回归测试先红后绿；pytest 1820 全过，ruff/mypy/vue-tsc/node --test 全绿。

.github/workflows/ci.yml

@@ -9,6 +9,7 @@ on:
 jobs:
   test:
     runs-on: ${{ matrix.os }}
+    timeout-minutes: 60 # Windows runner 偶发卡顿，6 格矩阵整体留足余量
     strategy:
       fail-fast: false
       matrix:
@@ -22,6 +23,10 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+          cache: "pip"
+          cache-dependency-path: |
+            pyproject.toml
+            requirements-dev.txt
       # 前端 dist 被 pyproject.toml 的 force-include 引用，pip install -e .
       # 要求它存在。先构建前端（type-check + vite build）。
       - uses: actions/setup-node@v4
@@ -43,11 +48,14 @@ jobs:
 
   mypy:
     runs-on: ubuntu-latest
+    timeout-minutes: 30
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
           python-version: "3.13"
+          cache: "pip"
+          cache-dependency-path: pyproject.toml
       # 同 test job：force-include 要求 web-ui/dist 存在。
       - uses: actions/setup-node@v4
         with:
@@ -67,6 +75,7 @@ jobs:
   # 与 Python 版本无关，故独立 job，避免在 test matrix 里跑 6 遍。
   frontend:
     runs-on: ubuntu-latest
+    timeout-minutes: 60 # Playwright 安装 + E2E 首跑较慢
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-node@v4
@@ -84,7 +93,15 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.13"
+          cache: "pip"
+          cache-dependency-path: pyproject.toml
       - run: pip install -e ".[web]"
+      # 浏览器二进制按 lockfile 哈希缓存（版本变 → 新 key → 重下）
+      - name: Cache Playwright browsers
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/ms-playwright
+          key: playwright-${{ runner.os }}-${{ hashFiles('web-ui/package-lock.json') }}
       - run: npx playwright install --with-deps chromium
         working-directory: web-ui
       - run: npm run test:e2e

.github/workflows/release.yml

@@ -57,10 +57,22 @@ jobs:
           mv dist/easy-tdx.exe "dist/easy-tdx-${VERSION}-windows.exe"
           ls -lh dist/
 
+      # 生成 SHA256 清单并随产物上传——Release 正文引导用户核对该哈希
+      # （防杀软误报时自行验证完整性），必须真实存在而非只写在文案里。
+      - name: Compute SHA256
+        shell: bash
+        run: |
+          cd dist
+          FILE=$(ls easy-tdx-*-windows.exe)
+          sha256sum "$FILE" > SHA256SUMS.txt
+          cat SHA256SUMS.txt
+
       - uses: actions/upload-artifact@v4
         with:
           name: easy-tdx-windows-exe
-          path: dist/easy-tdx-*-windows.exe
+          path: |
+            dist/easy-tdx-*-windows.exe
+            dist/SHA256SUMS.txt
           if-no-files-found: error
 
   release:
@@ -73,37 +85,64 @@ jobs:
           name: easy-tdx-windows-exe
           path: dist
 
+      # 重算哈希（与构建 job 的 SHA256SUMS.txt 同源同值），嵌入 Release 正文，
+      # 使「核对 Release 说明中公布的 SHA256」的指引真实可执行。
+      - name: Compute SHA256 for release notes
+        run: |
+          EXE=$(ls dist/easy-tdx-*-windows.exe)
+          SHA=$(sha256sum "$EXE" | cut -d' ' -f1)
+          echo "EXE_NAME=$(basename "$EXE")" >> "$GITHUB_ENV"
+          echo "EXE_SHA256=$SHA" >> "$GITHUB_ENV"
+
+      # 引号 heredoc 防止正文里的反引号被 bash 当命令替换，先写占位符再 sed 替换
+      - name: Generate release body
+        run: |
+          cat > release_body.md <<'BODY'
+          ## 下载使用
+
+          1. 下载下方 `__EXE_NAME__`（约 40-50MB）
+          2. 双击运行
+          3. 浏览器会自动打开回测界面（地址 `http://localhost:8000`）
+
+          ## ✅ 完整性核对（SHA256）
+
+          ```
+          __EXE_SHA256__  __EXE_NAME__
+          ```
+
+          PowerShell 核对：`Get-FileHash __EXE_NAME__`，与上方一致即为官方构建。
+          同份哈希也在附件 `SHA256SUMS.txt` 中。
+
+          ## ⚠️ SmartScreen 提示
+
+          本版本未做代码签名，首次运行 Windows 会弹出"已保护你的电脑"：
+
+          1. 点击 **更多信息**
+          2. 点击 **仍要运行**
+
+          Phase 2 将引入代码签名消除此提示。
+
+          ## 🛡️ 杀软误报说明（Windows Defender 报毒？请先读这里）
+
+          部分 Windows Defender 用户会看到 `Trojan:Win32/Wacatac.C!ml` 报警。
+          `.exe` 由 GitHub Actions 在公开的 tag 提交上从源码构建（本文件即构建脚本），
+          为**未签名的 PyInstaller 单文件**打包——"自解压 + 无签名"特征是杀软
+          机器学习引擎（`!ml` 后缀即 ML 判定）的经典误报源头，v1.32.1 起各版本
+          构建方式相同。自行核验：
+
+          1. 核对上方公布的 SHA256（PowerShell：`Get-FileHash __EXE_NAME__`）
+          2. 可上传 [VirusTotal](https://www.virustotal.com/) 交叉验证：典型误报特征是
+             少数 ML 启发式引擎报警、主流特征码引擎（Kaspersky/ESET/BitDefender 等）不报
+          3. 哈希一致仍想加速白名单，可向微软提交误报申诉：
+             [microsoft.com/en-us/wdsi/filesubmission](https://www.microsoft.com/en-us/wdsi/filesubmission)（选"软件开发者"）
+          BODY
+          sed -i "s/__EXE_NAME__/${EXE_NAME}/g; s/__EXE_SHA256__/${EXE_SHA256}/g" release_body.md
+
       - name: Create release
         uses: softprops/action-gh-release@v2
         with:
-          files: dist/easy-tdx-*-windows.exe
+          files: |
+            dist/easy-tdx-*-windows.exe
+            dist/SHA256SUMS.txt
           generate_release_notes: true
-          body: |
-            ## 下载使用
-
-            1. 下载下方 `easy-tdx-*-windows.exe`（约 40-50MB）
-            2. 双击运行
-            3. 浏览器会自动打开回测界面（地址 `http://localhost:8000`）
-
-            ## ⚠️ SmartScreen 提示
-
-            本版本未做代码签名，首次运行 Windows 会弹出"已保护你的电脑"：
-
-            1. 点击 **更多信息**
-            2. 点击 **仍要运行**
-
-            Phase 2 将引入代码签名消除此提示。
-
-            ## 🛡️ 杀软误报说明（Windows Defender 报毒？请先读这里）
-
-            部分 Windows Defender 用户会看到 `Trojan:Win32/Wacatac.C!ml` 报警。
-            `.exe` 由 GitHub Actions 在公开的 tag 提交上从源码构建（本文件即构建脚本），
-            为**未签名的 PyInstaller 单文件**打包——"自解压 + 无签名"特征是杀软
-            机器学习引擎（`!ml` 后缀即 ML 判定）的经典误报源头，v1.32.1 起各版本
-            构建方式相同。自行核验：
-
-            1. 核对 Release 说明中公布的 SHA256（PowerShell：`Get-FileHash *.exe`）
-            2. 可上传 [VirusTotal](https://www.virustotal.com/) 交叉验证：典型误报特征是
-               少数 ML 启发式引擎报警、主流特征码引擎（Kaspersky/ESET/BitDefender 等）不报
-            3. 哈希一致仍想加速白名单，可向微软提交误报申诉：
-               [microsoft.com/en-us/wdsi/filesubmission](https://www.microsoft.com/en-us/wdsi/filesubmission)（选"软件开发者"）
+          body_path: release_body.md

CHANGELOG.md

@@ -2,6 +2,48 @@
 
 本文件记录 easy-tdx 的版本变更。格式遵循 [Keep a Changelog](https://keepachangelog.com/zh-CN/)。
 
+## [1.32.6] — 2026-09-06
+
+**两周改动深度审查后的全面修复**——对 v1.21→v1.32.5 的 249 个文件、4.2 万行改动做六路专项审查，本轮落地全部修复：回测数字可信性（组合收益虚增 / 轮动停牌成交 / WF 指标稀释三件套）、LLM 安全加固（封死 file:// 读取与 API Key 外送链）、涨跌停价舍入漏判、数据源与缓存正确性、前端请求竞态等，共 58 处行为修复、约 60 条"先红后绿"回归测试。
+
+### 回测正确性（影响数字可信性，建议重点升级）
+
+- **组合收益不再虚增**（[portfolio_engine.py](src/easy_tdx/backtest/portfolio_engine.py)、[multi_strategy_engine.py](src/easy_tdx/backtest/multi_strategy_engine.py)）：组合内标的起始日期不齐时（次新股/取数截断），合并净值曲线前导缺口旧代码填 0，导致曲线首值小于总投入、`total_return` 被系统性虚增（实测两标的各投 10 万可显示 +126%，真实约 +13%）；改为前导缺口按初始资金回填，与组合 Walk-Forward 口径一致，`total_return` 恒等于资金加权收益率。
+- **轮动回测停牌日不再按过期价成交**（[rotation.py](src/easy_tdx/backtest/rotation.py)）：停牌日挂单此前会以停牌前旧开盘价成交（实测卖出价虚增 43%）；改为当日真实有交易才可成交，挂单顺延至复牌开盘，符合真实挂单语义。另修：首个调仓日照常产生信号、历史不足 5 根的次新股不再以 0 分混入买入候选。
+- **单标的 Walk-Forward 逐窗指标不再被预热区稀释**（[walkforward.py](src/easy_tdx/backtest/walkforward.py)）：上下文预热 bar 此前计入窗口绩效，默认参数下逐窗 sharpe/年化被零收益段大幅稀释（实测 8.84 vs 正确 14.96）；改为只用窗内净值与成交计算，与组合级同口径。另修：`worst_drawdown` 方向取反、单窗不足 20 根强制跳过、窗口失败记 warning 不再静默、int 日期列窗口标签显示 1970-01-01。
+- **评分容错**（[grading.py](src/easy_tdx/backtest/grading.py)）：组合回撤序列缺行/NaN 不再截断或放大水下期计数；组合级"适配性体检"对 ETF/可转债组合按品种费率解析（此前错收股票印花税）。
+
+### 安全
+
+- **LLM 接口加固**（[ai/llm.py](src/easy_tdx/ai/llm.py)）：`api_url` 仅允许 http/https 且禁止携带 URL 凭据（封死 `file://` 读本地配置文件与内网 SSRF 链）；HTTP 错误不再回显原始响应体（改为只提取 provider 错误 message，杜绝错误通道变任意内容回读）；响应体 2MB 上限；配置文件原子写（崩溃不再半写损坏）；手工编辑的 llm.json 坏字段自动回退默认并告警，不再打挂全部 AI 端点；anthropic 路径思考型模型空正文给出可操作报错而非静默空回复。
+
+### 数据与统计
+
+- **涨跌停价舍入修复**（[limitup.py](src/easy_tdx/screen/limitup.py)）：旧浮点实现 `floor(x*100+0.5)` 在半分边界受浮点误差影响，±10% 档 67/318 个、±5% 档 90 个价位会算低 1 分（如 33.05×1.1 误算 36.35，交易所 36.36），真实涨跌停被静默漏判；改为纯整数分币运算，全价位对账与交易所零差异。涨停家数/连板高度/炸板率/离线回补一并修正，ST 5% 与 3 元低价门槛改为逐 bar 判定。
+- **时区统一沪市时间**（[session.py](src/easy_tdx/realtime/session.py) 等）：`is_trading_time`、情绪采样、warehouse provisional 标记/转正、热点"今日"列此前用主机本地时区，非中国时区主机上情绪页永久无数据、当日 K 线被长期标 provisional；统一为固定 UTC+8（中国无夏令时），与主机时区无关。
+- **warehouse 增量同步**（[sync.py](src/easy_tdx/warehouse/sync.py)、[store.py](src/easy_tdx/warehouse/store.py)）：停用超过约三周后再同步，尾部 15 根覆盖不到的缺口此前永久丢失且 summary 照常 ok；现在检测到缺口自动全量重拉并告警。provisional 盘中临时值改为"拉取成功后仅转正到本次拉到的最新 bar"，数据源失败不再把未定值洗成 completed。
+- **baostock 兜底**（[sources/baostock.py](src/easy_tdx/sources/baostock.py)）：真故障（登录失败/查询报错）改为记日志并抛错，仅真无数据返回空，`--source baostock` 不再静默"无数据"；周/月线去掉服务端明确报错的 `tradestatus` 字段（实测 error_code=10004012——周/月兜底此前从未工作过）；指数兜底 vol 按实测（sh.000001）从股换算为手对齐 /bars/index 契约；`/bars` 数字周期串（如 `category=4`）不再绕过兜底。
+- **ccpm**：中金所页面结构变更时抛 `CcpmError` 并附结构线索，不再静默返回空表被当成"无数据"。
+- **Web API 正确性**（[board_mac.py](src/easy_tdx/web/routers/board_mac.py)、[market.py](src/easy_tdx/web/routers/market.py)、[schemas.py](src/easy_tdx/web/schemas.py)、[bars.py](src/easy_tdx/web/routers/bars.py) 等）：板块总览/涨停缓存键补齐 `count`/`vipdoc`（15s TTL 内不同参数不再互相串台）；响应 NaN 递归清洗且先清洗后入缓存（一行 NaN 导致稳定 500 的问题消除）；回测/寻优/公式端点 `count>800` 改分页取全量（TDX 单次协议上限，此前静默截短回测窗口）；任务提交响应透传真实状态（极快完成的任务不再谎报 running）；任务淘汰不再把未起跑的 pending 淘汰成"永久 pending 幽灵"。
+
+### 公式解析
+
+- [formula.py](src/easy_tdx/formula.py)、[MyTT.py](src/easy_tdx/MyTT.py)：FILTER 不再原地改写输入序列（`FILTER(C,2); MA(C,2)` 的 MA 此前被污染）；值域兜底收严（RSI/100、价格比率等 0~1 数值列不再被误判为买卖信号列）；解析深度上限 100 层（5000 层嵌套从 RecursionError 裸崩改为 FormulaError）；REF 负移位显式禁止（未来函数入口不再依赖类型巧合拦截）。
+
+### 前端
+
+- **请求竞态守卫**（[HotspotView.vue](web-ui/src/views/HotspotView.vue)、HotspotCorrelation、BoardOverviewView、BoardDialog 四处同款）：快速切换类型/参数时旧响应后到不再覆盖新数据、不再杀死构建轮询、不再产生幽灵翻红翻绿事件，休市/后台标签页不再停留错误状态。
+- 其余：Dashboard 指数分时不再被裁掉约 39%（Sparkline viewBox 随尺寸）；北交所个股从榜单点开能正确拉到行情（market=2 映射与 920xxx 前缀修正）；大盘日历空数据不再被永久缓存（重试可用）；AI 解读弹窗关闭即中止轮询（此前最长空转 20 分钟）；量能图与资金日历口径/正负号展示修正。
+
+### CLI / CI / 打包
+
+- `warehouse sync` 全部失败退出码改 1（对齐 ccpm）；`--period` 参数枚举校验；`--source baostock` + 非日线周期前置拦截；公式/仓库命令缺冒号参数改为干净报错（此前裸 traceback）。
+- Release 工作流现在真实计算并发布 SHA256SUMS.txt（正文"核对哈希"指引从此成立）；CI 各 job 加超时与 pip/Playwright 缓存；PyInstaller spec 构建前提补 baostock extra 并在缺包时明确报错。
+
+### 测试
+
+- 约 60 条回归测试全部"先在旧码跑出失败、再修复转绿"；两项存疑以实测定案：baostock 周/月线字段与指数 vol 单位（真实登录对账）、深市 ETF vol 单位（本机 vipdoc 与服务器双源一致，现有 ÷100 换算正确）。全量 1820 通过，ruff / ruff format / mypy 严格模式 / `vue-tsc` / 前端 node --test 23 条全绿。
+
 ## [1.32.5] — 2026-09-06
 
 **回撤持续统计修复：不再恒为 1**——单标的回测「绩效指标 → 风险 → 回撤持续」此前无论什么股票都显示 1，本版修复计算错误，并把三处实现的口径统一为指标文档承诺的「最长水下期」：从净值峰值跌落到重新创新高的最长天数（末日仍未修复则计到最后一天），即"最长一次套牢了多久"。

easy_tdx.spec

@@ -3,7 +3,9 @@
 
 构建前提（CI 会自动完成，本地手动构建需自行执行）::
 
-    1. pip install -e ".[web]" pyinstaller
+    1. pip install -e ".[web,packaging,baostock]" pyinstaller
+       # baostock 为 EXE 内置兜底数据源（见下方 hiddenimports），缺它会
+       # 被 spec 顶部的显式检查拦下——这是有意设计，防止静默产出无兜底的 EXE
     2. cd web-ui && npm ci && npm run build   # 产出 web-ui/dist/
     3. pyinstaller easy_tdx.spec              # 产出 dist/easy-tdx.exe
 
@@ -24,8 +26,19 @@ from PyInstaller.utils.hooks import collect_data_files, collect_submodules
 hiddenimports: list[str] = []
 hiddenimports += collect_submodules("uvicorn")
 hiddenimports += collect_submodules("easy_tdx")
-# baostock 兜底数据源（v1.33：/bars 与 warehouse 的最后一级回退）在
+# baostock 兜底数据源（/bars 与 warehouse 的最后一级回退）在
 # sources/baostock.py 里经 importlib 懒加载，静态分析扫不到，需显式声明。
+# 显式检查而非依赖 collect_submodules 的失败形态（不同 PyInstaller 版本下
+# 可能返回空列表静默跳过）——缺 baostock 时立即报错，防止静默产出
+# 「兜底源缺失」的 EXE（运行期 /bars 自动回退静默失效，极难排查）。
+import importlib.util
+
+if importlib.util.find_spec("baostock") is None:
+    raise SystemExit(
+        "打包错误: 未找到 baostock——EXE 内置兜底数据源需要它。"
+        '请先执行 pip install baostock（或 pip install -e ".[web,packaging,baostock]"）'
+        "再运行 pyinstaller easy_tdx.spec"
+    )
 hiddenimports += collect_submodules("baostock")
 # pandas / numpy / scipy 由 PyInstaller 自带 hook 处理（见
 # PyInstaller/hooks/hook-pandas.* 等），无需手动 collect_submodules——

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "easy-tdx"
-version = "1.32.5"
+version = "1.32.6"
 description = "通达信 TCP 协议行情数据客户端，支持在线行情、离线数据读取与写入同步"
 readme = "README.md"
 requires-python = ">=3.10"

src/easy_tdx/MyTT.py

@@ -183,9 +183,13 @@ def EXIST(S, N):  # EXIST(CLOSE>3010, N=5)  n日内是否存在一天大于3000
 
 
 def FILTER(S, N):  # FILTER函数，S满足条件后，将其后N周期内的数据置为0, FILTER(C==H,5)
-    for i in range(len(S)):
-        S[i + 1 : i + 1 + N] = 0 if S[i] else S[i + 1 : i + 1 + N]
-    return S  # 例：FILTER(C==H,5) 涨停后，后5天不再发出信号
+    # 无副作用实现：在副本上置零。曾直接改写输入序列——公式通道里
+    # FILTER(C, N) 会把同一公式后续语句引用的 C 一并污染（或对只读
+    # 数组直接报错）。
+    out = np.array(S, copy=True)
+    for i in range(len(out)):
+        out[i + 1 : i + 1 + N] = 0 if out[i] else out[i + 1 : i + 1 + N]
+    return out  # 例：FILTER(C==H,5) 涨停后，后5天不再发出信号
 
 
 def BARSLAST(S):  # 上一次条件成立到当前的周期, BARSLAST(C/REF(C,1)>=1.1) 上一次涨停到今天的天数

src/easy_tdx/ai/llm.py

@@ -29,9 +29,12 @@ from __future__ import annotations
 import asyncio
 import json
 import logging
+import math
 import os
+import tempfile
 import time
 import urllib.error
+import urllib.parse
 import urllib.request
 from dataclasses import asdict, dataclass, field, replace
 from pathlib import Path
@@ -45,6 +48,7 @@ __all__ = [
     "mask_key",
     "resolve_config",
     "save_config",
+    "validate_api_url",
 ]
 
 logger = logging.getLogger(__name__)
@@ -157,29 +161,108 @@ def _read_config_file() -> dict[str, Any]:
 
 
 def load_config() -> LlmConfig:
-    """加载配置：llm.json 显式字段 > 环境变量兜底（未填字段仍为空，调用时再取预设）。"""
+    """加载配置：llm.json 显式字段 > 环境变量兜底（未填字段仍为空，调用时再取预设）。
+
+    字段级防御：llm.json 常被手工编辑，单个字段类型不对（``"temperature":
+    null`` / ``"abc"`` 等）只记 warning 并回退默认值，绝不让 load_config
+    抛异常打挂全部 /llm/* 端点。
+    """
     data = _read_config_file()
     env_url = os.environ.get("LLM_BASE_URL", "")
-    cfg = LlmConfig(
-        provider=str(data.get("provider") or os.environ.get("LLM_PROVIDER", "") or "deepseek"),
-        api_url=str(data.get("api_url") or env_url or ""),
-        api_key=str(data.get("api_key") or os.environ.get("LLM_API_KEY", "") or ""),
-        model=str(data.get("model") or os.environ.get("LLM_MODEL", "") or ""),
-        temperature=float(data.get("temperature", 0.3)),
-        max_tokens=int(data.get("max_tokens", 16000)),
-        timeout=float(data.get("timeout", 180.0)),
-        system_prompt=str(data.get("system_prompt", "") or LlmConfig.system_prompt),
+    return LlmConfig(
+        provider=_clean_str(data.get("provider") or os.environ.get("LLM_PROVIDER", ""), "provider")
+        or "deepseek",
+        api_url=_clean_str(data.get("api_url") or env_url, "api_url"),
+        api_key=_clean_str(data.get("api_key") or os.environ.get("LLM_API_KEY", ""), "api_key"),
+        model=_clean_str(data.get("model") or os.environ.get("LLM_MODEL", ""), "model"),
+        temperature=_clean_float(data.get("temperature"), "temperature", 0.3, minimum=0.0),
+        max_tokens=_clean_int(data.get("max_tokens"), "max_tokens", 16000, minimum=1),
+        timeout=_clean_float(data.get("timeout"), "timeout", 180.0, minimum=0.1),
+        system_prompt=_clean_str(
+            data.get("system_prompt") or LlmConfig.system_prompt, "system_prompt"
+        )
+        or LlmConfig.system_prompt,
     )
-    return cfg
+
+
+def _clean_str(value: Any, field_name: str) -> str:
+    """字符串字段清洗：非 None 的非字符串（数字/列表等）记 warning 后回退空串。"""
+    if value is None:
+        return ""
+    if not isinstance(value, str):
+        logger.warning(
+            "LLM 配置字段 %s 应为字符串，收到 %s(%r)，已忽略并回退默认值",
+            field_name,
+            type(value).__name__,
+            value,
+        )
+        return ""
+    return value
+
+
+def _clean_float(
+    value: Any, field_name: str, default: float, *, minimum: float | None = None
+) -> float:
+    """数值字段清洗：非法/非有限/低于下限均 warning 后回退默认。"""
+    if value is None:
+        return default
+    try:
+        f = float(value)
+    except (TypeError, ValueError):
+        logger.warning(
+            "LLM 配置字段 %s 应为数字，收到 %r，回退默认值 %s", field_name, value, default
+        )
+        return default
+    if not math.isfinite(f) or (minimum is not None and f < minimum):
+        logger.warning(
+            "LLM 配置字段 %s 超出合理范围（%r），回退默认值 %s", field_name, value, default
+        )
+        return default
+    return f
+
+
+def _clean_int(value: Any, field_name: str, default: int, *, minimum: int | None = None) -> int:
+    """整数字段清洗：接受 "8192.9" 这类字符串的宽松矫正（截断到 int）。"""
+    if value is None:
+        return default
+    try:
+        i = int(value)
+    except (TypeError, ValueError):
+        try:
+            i = int(float(value))
+        except (TypeError, ValueError):
+            logger.warning(
+                "LLM 配置字段 %s 应为整数，收到 %r，回退默认值 %s", field_name, value, default
+            )
+            return default
+    if not math.isfinite(i) or (minimum is not None and i < minimum):
+        logger.warning(
+            "LLM 配置字段 %s 超出合理范围（%r），回退默认值 %s", field_name, value, default
+        )
+        return default
+    return i
 
 
 def save_config(cfg: LlmConfig) -> Path:
-    """写入 llm.json（WebUI 保存入口；目录惰性创建）。"""
+    """写入 llm.json（WebUI 保存入口；目录惰性创建；原子写）。
+
+    先写同目录临时文件再 ``os.replace``——写一半崩溃/断电不会留下损坏的
+    llm.json（损坏的后果是下次 load 静默回空配置，用户要重填 key）。
+    """
     p = config_path()
     p.parent.mkdir(parents=True, exist_ok=True)
-    p.write_text(
-        json.dumps(cfg.to_dict(), ensure_ascii=False, indent=2), encoding="utf-8", newline="\n"
-    )
+    text = json.dumps(cfg.to_dict(), ensure_ascii=False, indent=2)
+    fd, tmp_name = tempfile.mkstemp(dir=str(p.parent), prefix=".llm-", suffix=".tmp")
+    try:
+        with os.fdopen(fd, "w", encoding="utf-8", newline="\n") as f:
+            f.write(text)
+        os.replace(tmp_name, p)
+    except BaseException:
+        try:
+            os.unlink(tmp_name)
+        except OSError:
+            pass
+        raise
     return p
 
 
@@ -188,10 +271,12 @@ def resolve_config(cfg: LlmConfig | None = None) -> LlmConfig:
 
     - ``api_url`` 空 → 预设 ``base_url``；
     - ``model`` 空 → 预设 ``default_model``；
-    - provider 无预设（拼错）→ 按 custom 处理，url/model 必须已填。
+    - provider 无预设（拼错）→ 按 custom 处理，url/model 必须已填；
+    - ``api_url`` 强制 http/https 且禁 userinfo（SSRF/本地文件读取防线）。
 
     Raises:
-        ValueError: 补齐后仍缺 api_url 或 model（custom 未填全）。
+        ValueError: 补齐后仍缺 api_url 或 model（custom 未填全），
+            或 api_url 非法（非 http/https、携带 user:pass@）。
     """
     c = replace(cfg or load_config())
     preset = PROVIDER_PRESETS.get(c.provider, PROVIDER_PRESETS["custom"])
@@ -203,6 +288,7 @@ def resolve_config(cfg: LlmConfig | None = None) -> LlmConfig:
         raise ValueError(
             f"LLM 配置不完整：provider={c.provider} 缺少 api_url 或 model，请在 AI 设置中补全"
         )
+    validate_api_url(c.api_url)
     return c
 
 
@@ -226,6 +312,80 @@ class LlmError(RuntimeError):
         self.status = status
 
 
+#: 响应体大小上限：正常 chat 响应远小于此（max_tokens 128k 的纯文本约几百 KB），
+#: 超限说明对端异常（如把 api_url 配成了下载地址），及时中止防内存被撑爆。
+_MAX_RESPONSE_BYTES = 2 * 1024 * 1024
+_READ_CHUNK = 64 * 1024
+
+#: 允许的 URL scheme（SSRF / 本地文件读取防线：file:// 会读本地文件、
+#: ftp:// 与内网 http 可被当跳板——resolve_config 里强制校验）。
+_ALLOWED_URL_SCHEMES = ("http", "https")
+
+
+def validate_api_url(url: str) -> None:
+    """api_url 安全校验：仅 http/https、禁止携带 userinfo（user:pass@）。
+
+    Raises:
+        ValueError: scheme 非法/缺失，或 URL 携带用户凭据（web 层已有
+            ValueError→错误响应通道，CLI 场景同样可直接展示）。
+    """
+    if not url:
+        return
+    parts = urllib.parse.urlsplit(url)
+    scheme = parts.scheme.lower()
+    if scheme not in _ALLOWED_URL_SCHEMES:
+        raise ValueError(
+            f"LLM api_url 非法：仅允许 http/https 地址（收到 scheme={scheme!r}）——"
+            "file/ftp 等协议已禁用；缺前缀时请补 http:// 或 https://"
+        )
+    if parts.username or parts.password:
+        raise ValueError(
+            "LLM api_url 非法：不允许携带用户凭据（user:pass@host 形式）——"
+            "请把鉴权放到 API Key 字段（请求头），而不是 URL 里"
+        )
+
+
+def _read_capped(fp: Any, cap: int = _MAX_RESPONSE_BYTES) -> bytes:
+    """分块读响应体，超过 cap 字节即中止（防异常网关撑爆内存）。"""
+    buf = bytearray()
+    while True:
+        chunk = fp.read(_READ_CHUNK)
+        if not chunk:
+            break
+        buf.extend(chunk)
+        if len(buf) > cap:
+            raise LlmError(
+                f"LLM API 响应超过 {cap // (1024 * 1024)}MB 上限——对端不是正常的 chat 接口"
+                "（请检查 api_url 是否填错），已中止"
+            )
+    return bytes(buf)
+
+
+def _extract_error_message(body: str) -> str | None:
+    """从 provider 错误响应中提取可读 message（不回显原始 body）。
+
+    OpenAI/Anthropic 兼容网关的惯例是 ``{"error": {"message": ...}}``；
+    部分网关 error 直接是字符串，或把 message 放顶层。提取结果截到 300 字符。
+    """
+    try:
+        obj = json.loads(body)
+    except (json.JSONDecodeError, ValueError):
+        return None
+    if not isinstance(obj, dict):
+        return None
+    err = obj.get("error")
+    if isinstance(err, dict):
+        msg = err.get("message") or err.get("msg") or err.get("code")
+        if msg:
+            return str(msg)[:300]
+    elif isinstance(err, str) and err:
+        return err[:300]
+    msg = obj.get("message")
+    if msg:
+        return str(msg)[:300]
+    return None
+
+
 def _post_json(
     url: str, headers: dict[str, str], payload: dict[str, Any], timeout: float
 ) -> dict[str, Any]:
@@ -235,6 +395,7 @@ def _post_json(
     超时单独成类报错：非流式 chat 接口要等模型**整段回复生成完**才回包，
     大 Prompt（如整份回测报告解读）生成 1-3 分钟很正常，读超时≠网络故障，
     报错必须把「调大超时」这个动作说清楚（v1.29.1 实测踩坑）。
+    HTTP 错误只回显 provider 的 error.message（≤300 字符），不透传原始 body。
     """
     req = urllib.request.Request(
         url,
@@ -244,11 +405,16 @@ def _post_json(
     )
     try:
         with urllib.request.urlopen(req, timeout=timeout) as resp:
-            data: dict[str, Any] = json.loads(resp.read().decode("utf-8"))
+            raw = _read_capped(resp)
+            data: dict[str, Any] = json.loads(raw.decode("utf-8", errors="replace"))
             return data
     except urllib.error.HTTPError as exc:
-        body = exc.read().decode("utf-8", errors="replace")[:500]
-        raise LlmError(f"LLM API HTTP {exc.code}: {body}", status=exc.code) from exc
+        try:
+            body = exc.read(_READ_CHUNK).decode("utf-8", errors="replace")
+        except OSError:
+            body = ""
+        detail = _extract_error_message(body) or "接口返回错误响应"
+        raise LlmError(f"LLM API HTTP {exc.code}: {detail}", status=exc.code) from exc
     except urllib.error.URLError as exc:
         if isinstance(exc.reason, TimeoutError):
             raise LlmError(_timeout_message(timeout)) from exc
@@ -256,7 +422,9 @@ def _post_json(
     except TimeoutError as exc:
         raise LlmError(_timeout_message(timeout)) from exc
     except json.JSONDecodeError as exc:
-        raise LlmError(f"LLM API 响应不是合法 JSON: {exc}") from exc
+        raise LlmError(
+            "LLM API 响应不是合法 JSON——api_url 可能不是 chat 接口端点，请检查 AI 设置"
+        ) from exc
 
 
 def _timeout_message(timeout: float) -> str:
@@ -325,8 +493,14 @@ class LlmClient:
         except LlmError:
             raise
         except (KeyError, IndexError, TypeError) as exc:
-            raw = json.dumps(data, ensure_ascii=False)[:300]
-            raise LlmError(f"LLM 响应格式异常: {raw}") from exc
+            # 不回显原始响应体：内容可能包含网关内部信息，且 historical 上
+            # 曾被当作任意 URL 响应的回读通道。只描述缺什么 + 顶层键名。
+            hint = sorted(data.keys()) if isinstance(data, dict) else type(data).__name__
+            raise LlmError(
+                "LLM 响应格式异常：未找到 choices[0].message 字段"
+                f"（响应顶层字段: {hint}）——请检查 api_url 是否为正确的"
+                " chat/completions 端点、模型名是否正确"
+            ) from exc
 
     def _extract_reply_openai(self, message: dict[str, Any], finish: str) -> str:
         """从 OpenAI 兼容响应的 message 里提取正文，处理思考型模型的空白正文。
@@ -352,8 +526,11 @@ class LlmClient:
             raise LlmError(
                 "模型输出被 max_tokens 截断且无正文，请在「AI 设置」调大 Max Tokens 后重试"
             )
-        raw = json.dumps(message, ensure_ascii=False)[:300]
-        raise LlmError(f"LLM 响应 message.content 为空: {raw}")
+        keys = sorted(message.keys()) if isinstance(message, dict) else type(message).__name__
+        raise LlmError(
+            f"LLM 响应 message.content 为空（message 字段: {keys}，"
+            f"finish_reason={finish or 'unknown'}）——请检查模型名与 api_url 是否匹配"
+        )
 
     def _chat_anthropic(self, prompt: str, system: str) -> str:
         cfg = self._cfg
@@ -371,11 +548,64 @@ class LlmClient:
         url = f"{cfg.api_url.rstrip('/')}/messages"
         data = _post_json(url, headers, payload, cfg.timeout)
         try:
-            blocks = data["content"]
-            return "".join(str(b.get("text", "")) for b in blocks if b.get("type") == "text")
-        except (KeyError, TypeError) as exc:
-            raw = json.dumps(data, ensure_ascii=False)[:300]
-            raise LlmError(f"LLM 响应格式异常: {raw}") from exc
+            return self._extract_reply_anthropic(data)
+        except LlmError:
+            raise
+        except (KeyError, TypeError, AttributeError) as exc:
+            # 与 openai 路径同口径：不回显原始响应体，只描述问题
+            hint = sorted(data.keys()) if isinstance(data, dict) else type(data).__name__
+            raise LlmError(
+                f"LLM 响应格式异常：content 字段不可解析（响应顶层字段: {hint}）——"
+                "请确认 api_url 指向 Anthropic /messages 端点、模型名正确"
+            ) from exc
+
+    def _extract_reply_anthropic(self, data: dict[str, Any]) -> str:
+        """从 Anthropic 响应提取正文，与 openai 路径同口径：绝不返回空串。
+
+        - content 是块列表：拼接 text 块，统计 thinking 块字数（思考耗尽
+          max_tokens 时报可操作错误，而非静默成功空串）；
+        - content 是字符串（部分网关）：直接作为正文；
+        - 空白正文：按 stop_reason 给「调大 Max Tokens」指引。
+        """
+        content = data["content"]
+        if isinstance(content, str):
+            text = content
+            thinking_len = 0
+            block_types: list[str] = []
+        elif isinstance(content, list):
+            parts: list[str] = []
+            thinking_len = 0
+            block_types = []
+            for b in content:
+                if not isinstance(b, dict):
+                    continue
+                b_type = str(b.get("type") or "")
+                block_types.append(b_type)
+                if b_type == "text":
+                    parts.append(str(b.get("text") or ""))
+                elif b_type in ("thinking", "redacted_thinking"):
+                    thinking_len += len(str(b.get("thinking") or b.get("data") or ""))
+            text = "".join(parts)
+        else:
+            raise TypeError(f"content 应为块列表或字符串，收到 {type(content).__name__}")
+        if text.strip():
+            return text
+        finish = str(data.get("stop_reason") or "")
+        if thinking_len:
+            raise LlmError(
+                f"模型只返回了思考链（thinking {thinking_len} 字），未生成正文——"
+                f"max_tokens={self._cfg.max_tokens} 大概率被思考耗尽"
+                f"（stop_reason={finish or 'unknown'}）。"
+                "请在「AI 设置」把 Max Tokens 调大（思考型模型建议 ≥16000）后重试"
+            )
+        if finish == "max_tokens":
+            raise LlmError(
+                "模型输出被 max_tokens 截断且无正文，请在「AI 设置」调大 Max Tokens 后重试"
+            )
+        raise LlmError(
+            f"LLM 响应 content 为空（block 类型: {block_types or '无'}，"
+            f"stop_reason={finish or 'unknown'}）——请检查模型名与 api_url 是否匹配"
+        )
 
     async def test(self) -> dict[str, Any]:
         """连通性测试：发一句极短 ping，返回 ok/延迟/样例回复。"""

src/easy_tdx/backtest/benchmark.py

@@ -357,13 +357,20 @@ def evaluate_portfolio(
         **engine_kwargs,
     ).run()
 
-    # 3. 适配性体检：逐标的跑三段体检，跨标的多数口径聚合
+    # 3. 适配性体检：逐标的跑三段体检，跨标的多数口径聚合。
+    # 每标的传入各自 symbol，使 auto_fees 按品种解析费率——与组合回测主路径
+    # （PortfolioBacktestEngine 逐标的 resolve_fee_model）同口径。此前漏传
+    # symbol：ETF/可转债组合的三段体检被按股票口径错收印花税。
     fitness_kwargs: dict[str, Any] = {
         k: v for k, v in engine_kwargs.items() if k not in ("total_cash", "chanlun_level")
     }
     per_stock_fitness = [
         FitnessEngine(
-            strategy=strategy, split=split, context_bars=context_bars, **fitness_kwargs
+            strategy=strategy,
+            split=split,
+            context_bars=context_bars,
+            symbol=f"{stock.market}{stock.code}",
+            **fitness_kwargs,
         ).evaluate(stock.df)
         for stock in stocks
     ]

src/easy_tdx/backtest/cli.py

@@ -115,6 +115,16 @@ def backtest(
 
     # 1. 加载策略（单策略 or 多因子组合）
     is_combo = combo_strategies is not None
+    # 组合模式暂不支持的分析旗标：显式告警而非静默吞掉（审查修复）
+    if is_combo and walk_forward:
+        click.echo(
+            "警告: --wf（Walk-Forward 样本外验证）暂不支持 --combo-strategies 组合模式，已忽略",
+            err=True,
+        )
+    if is_combo and full_evaluate:
+        click.echo(
+            "警告: --evaluate（一条龙评估）暂不支持 --combo-strategies 组合模式，已忽略", err=True
+        )
 
     if is_combo:
         assert combo_strategies is not None  # narrowed by is_combo
@@ -728,6 +738,13 @@ def optimize(
         raise SystemExit(1)
 
     custom_grid = _parse_param_grid(param_pairs) if param_pairs else None
+    if optimize_all and custom_grid is not None:
+        # --all 逐策略使用各自预设网格，--param 无处安放：显式告警而非静默忽略（审查修复）
+        click.echo(
+            "警告: --param 在 --all 模式下被忽略（一键寻优逐策略使用各自预设网格；"
+            "如需自定义网格请指定 --strategy 单策略寻优）",
+            err=True,
+        )
     if not optimize_all:
         assert strategy_name is not None
         try:

src/easy_tdx/backtest/grading.py

@@ -131,6 +131,11 @@ THRESHOLDS: dict[str, tuple[str, tuple[Anchor, ...]]] = {
             Anchor(0.6, 0),
         ),
     ),
+    # 回撤持续：该维度输入的单位是「bar 数」（performance.max_dd_duration 输出
+    # 水下期的 bar 数），下方天数锚点按日线（1 bar ≈ 1 交易日）校准。分钟级
+    # 周期下 bar 数远大于天数，得分会系统性偏低（偏保守）。注意：评分维度
+    # （grade_performance / grade_portfolio_equity）当前均未使用该表项，
+    # 锚点仅供展示与前端对照。
     "max_dd_duration": (
         "回撤持续",
         (
@@ -400,7 +405,13 @@ def _downweight_unreliable(dimensions: list[DimensionScore]) -> bool:
 
 @dataclass
 class CombinedMetrics:
-    """从净值序列重算的组合级指标（净值可推导的字段子集）。"""
+    """从净值序列重算的组合级指标（净值可推导的字段子集）。
+
+    ``max_dd_duration`` 的单位是 **bar 数**（一根 K 线计 1），与
+    ``performance.max_dd_duration`` 同口径；THRESHOLDS 里对应锚点的天数
+    （30/90/365…）按日线（1 bar ≈ 1 交易日）校准，分钟级周期下该值按 bar
+    直读会显著大于天数（评级维度未使用，仅展示）。
+    """
 
     total_return: float = 0.0
     annual_return: float = 0.0
@@ -463,12 +474,24 @@ def compute_combined_metrics(equity: list[dict[str, Any]]) -> CombinedMetrics:
     # 最大回撤：优先用 drawdown_pct（与前端一致），缺则从 totals 反推。
     # 持续 = 最长水下期（峰值 → 重新创新高；末日未修复则计到最后一点），
     # 与 performance.py / combinedMetrics.ts / max_dd_duration 锚点量纲同口径。
+    # 缺行 / None / NaN 的 drawdown_pct 按「状态延续」处理：沿用上一根的
+    # 水下/峰值状态——此前缺行被当 0（创新高）截断水下期、NaN 永久脱离
+    # 峰值判定，两者都会扭曲 max_dd_duration。
     max_dd = 0.0
     max_dd_dur = 0
     if equity[0].get("drawdown_pct") is not None:
         last_peak = 0
+        prev_dd = 0.0  # 上一根的有效回撤（首根之前视作峰值状态）
         for i, e in enumerate(equity):
-            dd = float(e.get("drawdown_pct") or 0.0)
+            raw = e.get("drawdown_pct")
+            dd: float | None
+            try:
+                dd = None if raw is None else float(raw)
+            except (TypeError, ValueError):
+                dd = None
+            if dd is None or not math.isfinite(dd):
+                dd = prev_dd
+            prev_dd = dd
             if dd > max_dd:
                 max_dd = dd
             if dd == 0:

src/easy_tdx/backtest/multi_strategy_engine.py

@@ -61,7 +61,8 @@ class MultiStrategyResult:
         total_performance: 组合整体绩效（资金加权收益率 + 策略数 + 总资金）。
         individual_results: 每个策略槽位的独立回测结果，key 形如 "{label}@{symbol}"。
         equity_allocation: 每个槽位的资金分配比例（均分时各 1/N）。
-        combined_equity: 组合整体净值曲线（各槽位按日期并集 ffill 对齐后求和），
+        combined_equity: 组合整体净值曲线（各槽位按日期并集 ffill 对齐后求和，
+            晚起步槽位的前导缺口按首个净值=初始资金回填），
             列: datetime / total / drawdown / drawdown_pct。
     """
 
@@ -200,7 +201,8 @@ class MultiStrategyEngine:
         算法与 ``PortfolioBacktestEngine._build_combined_equity`` 一致：
         各策略回测日期范围可能不同（取数差异、停牌），取 datetime 并集，
         每个策略的 total 列 forward-fill 对齐到并集后求和得组合总净值，
-        再算回撤。
+        再算回撤。前导缺口（晚起步槽位）按其首个净值（=初始资金）回填
+        （bfill），保证合并曲线首值等于总投入资金。
         """
         del allocations  # 资金分配不参与曲线形状（各策略独立 full cash 回测，
         # 合并的是 normalized 的净值贡献；保持签名与 Portfolio 版一致便于对照）
@@ -224,8 +226,15 @@ class MultiStrategyEngine:
         if not series_list:
             return empty
 
+        # 外连接对齐（并集日期）：各槽位在缺失日期 forward-fill（持有不动）；
+        # 前导缺口（晚起步槽位）用每列首个有效值回填（bfill）——资金在组合
+        # 起点即已分配，建仓前按初始资金趴账，与 PortfolioBacktestEngine 及
+        # 组合 Walk-Forward 的口径一致。此前前导缺口填 0：合并曲线首值会小于
+        # 总投入资金，total_return 被系统性虚增。
+        # 退化兜底：整列全 NaN（理论不可达——无数据的槽位不会进入 series_list）
+        # 显式落 0，避免 sum 传播 NaN。
         aligned = pd.concat(series_list, axis=1).sort_index()
-        aligned = aligned.ffill().fillna(0)
+        aligned = aligned.ffill().bfill().fillna(0.0)
         total = aligned.sum(axis=1)
 
         # 回撤：drawdown 为绝对回撤额（峰值-当前，正值），drawdown_pct 为相对当时

src/easy_tdx/backtest/portfolio_engine.py

@@ -44,7 +44,9 @@ class PortfolioResult:
         equity_allocation: 每只标的的资金分配比例
         combined_equity: 组合整体净值曲线（按日期对齐各标的求和），
             列: datetime/total/drawdown/drawdown_pct。各标的独立回测日期范围
-            可能不同，此处按日期并集 forward-fill 对齐后求和。
+            可能不同，此处按日期并集 forward-fill 对齐后求和；晚上市标的的
+            前导缺口按其首个净值（=初始资金）回填，保证合并曲线首值等于
+            总投入资金。
         trades: 组合层汇总成交（各标的 concat + ``symbol`` 列标注来源标的），
             供组合级绩效统计（逐标的 FIFO 配对持仓天数）与前端明细表使用。
     """
@@ -265,7 +267,9 @@ class PortfolioBacktestEngine:
         """把各标的独立净值曲线按日期对齐求和，生成组合整体净值曲线。
 
         各标的独立回测的日期范围可能不同（取数差异、停牌等），这里取所有标的
-        datetime 的并集，每个标的的 total 列 forward-fill 对齐到并集后求和。
+        datetime 的并集，每个标的的 total 列 forward-fill 对齐到并集后求和；
+        前导缺口（晚上市标的）按其首个净值（=初始资金）回填（bfill），保证
+        合并曲线首值等于总投入资金。
 
         Returns:
             DataFrame: datetime / total / drawdown / drawdown_pct。
@@ -293,10 +297,15 @@ class PortfolioBacktestEngine:
         if not series_list:
             return empty
 
-        # 外连接对齐（并集日期），forward-fill 各标的在缺失日期的净值（持有不动），
-        # 再求和得组合总净值。缺失值填 0 是为应对某标的完全无该日期数据的情况。
+        # 外连接对齐（并集日期）：各标的在缺失日期 forward-fill（持有不动）；
+        # 前导缺口（晚上市 / 取数晚于组合起点）用每列首个有效值回填（bfill）——
+        # 资金在组合起点即已分配，建仓前按初始资金趴账，与组合 Walk-Forward 的
+        # ffill().bfill() 口径一致。此前前导缺口填 0：晚上市标的上市前贡献 0，
+        # 合并曲线首值 < 总投入资金，total_return 被系统性虚增。
+        # 退化兜底：整列全 NaN（理论不可达——无数据的标的不会进入 series_list）
+        # 显式落 0，避免 sum 传播 NaN。
         aligned = pd.concat(series_list, axis=1).sort_index()
-        aligned = aligned.ffill().fillna(0)
+        aligned = aligned.ffill().bfill().fillna(0.0)
         total = aligned.sum(axis=1)
 
         # 回撤：drawdown 为绝对回撤额（峰值-当前，正值），drawdown_pct 为相对

src/easy_tdx/backtest/rotation.py

@@ -225,16 +225,24 @@ class RotationEngine:
 
             # 1. 推进各标的指针到 ≤ d 的最新一根
             bar_today: dict[str, pd.Series] = {}
+            traded_today: set[str] = set()
             for sym, df in self._dfs.items():
                 dts = self._dt_index(sym)
                 while pointers[sym] + 1 < len(dts) and dts[pointers[sym] + 1] <= d:
                     pointers[sym] += 1
                 if pointers[sym] >= 0:
                     bar_today[sym] = df.iloc[pointers[sym]]
+                    # 当日真实有 bar 才可成交；停牌标的只有旧 bar（估值用）
+                    if dts[pointers[sym]] == d:
+                        traded_today.add(sym)
 
-            # 2. 次开执行昨日信号（用当日开盘价）
+            # 2. 次开执行挂单（用当日开盘价）。停牌标的当日不可成交，挂单顺延、
+            #    复牌开盘成交（真实挂单语义）；第 4 步排队按 symbol+方向去重，
+            #    不会与后续新信号重复排队。
+            still_pending: list[tuple[str, str, str]] = []
             for sym, direction, reason in pending:
-                if sym not in bar_today:
+                if sym not in traded_today:
+                    still_pending.append((sym, direction, reason))
                     continue
                 price = float(bar_today[sym]["open"])
                 if not math.isfinite(price) or price <= 0:
@@ -268,7 +276,7 @@ class RotationEngine:
                             day_i, d_str, sym, "BUY", shares, price, fee, 0.0, reason=reason
                         )
                     )
-            pending = []
+            pending = still_pending
 
             # 3. 止盈止损检查（收盘口径，次日执行）
             for sym in list(positions):
@@ -281,7 +289,8 @@ class RotationEngine:
                 elif self._take_profit is not None and close >= cost * (1 + self._take_profit):
                     pending.append((sym, "SELL", "take_profit"))
 
-            # 4. 调仓判定
+            # 4. 调仓判定（day0 即可产生初始调仓信号，次日开盘执行；排名只用
+            #    截至 day0 收盘的数据，无未来泄漏）
             key = (
                 (d.isocalendar()[0], d.isocalendar()[1])
                 if self._refresh == "weekly"
@@ -289,7 +298,7 @@ class RotationEngine:
             )
             is_rebalance = key != prev_key
             prev_key = key
-            if is_rebalance and day_i >= 1:
+            if is_rebalance:
                 rebalances.append(d_str)
                 ranked = self._rank_all(pointers, d)
                 top_keep = [s for s, _ in ranked[: self._keep_rank]]
@@ -355,12 +364,16 @@ class RotationEngine:
         return pd.DatetimeIndex(self._dfs[sym]["_ts"])
 
     def _rank_all(self, pointers: dict[str, int], d: Any) -> list[tuple[str, float]]:
-        """对全部标的按截至 d 的前缀数据打分并降序排名。"""
+        """对全部标的按截至 d 的前缀数据打分并降序排名。
+
+        历史不足（< 5 根，如次新股）打不出有效分，直接从排名（买入候选）中
+        剔除而非按 0 分参与排序——0 分会排在负动量标的之前导致误买。持仓
+        标的买入时即已满足 ≥5 根且指针只进不退，不受影响。
+        """
         scored: list[tuple[str, float]] = []
         for sym, df in self._dfs.items():
             idx = pointers[sym]
             if idx < 5:
-                scored.append((sym, 0.0))
                 continue
             start = max(0, idx - self._max_history)
             prefix = df.iloc[start : idx + 1].drop(columns=["_ts"], errors="ignore")

src/easy_tdx/backtest/walkforward.py

@@ -30,6 +30,7 @@
 
 from __future__ import annotations
 
+import logging
 from dataclasses import dataclass, field
 from typing import Any
 
@@ -37,9 +38,15 @@ import numpy as np
 import pandas as pd
 
 from easy_tdx.backtest.engine import BacktestEngine
+from easy_tdx.backtest.performance import PerformanceAnalyzer
 from easy_tdx.backtest.strategy import Strategy
 from easy_tdx.backtest.types import to_json_native
 
+logger = logging.getLogger(__name__)
+
+# 单个测试窗最少 bar 数（低于此的窗口跳过，不参与评估）
+MIN_WINDOW_BARS = 20
+
 __all__ = [
     "WalkForwardWindow",
     "WalkForwardResult",
@@ -198,7 +205,7 @@ class WalkForwardEngine:
         for i in range(self._n_windows):
             s = eval_start + i * window_len
             e = s + window_len if i < self._n_windows - 1 else n  # 末窗吃到尾部
-            if e - s < 5:
+            if e - s < MIN_WINDOW_BARS:
                 continue
             win = self._run_window(df, s, e, i)
             if win is not None:
@@ -211,7 +218,9 @@ class WalkForwardEngine:
         """独立回测单个窗口 [s, e)。
 
         带前置上下文（指标预热），用 warmup_bars 压制上下文区间的信号；
-        窗口起点空仓（每窗独立开仓语义）。
+        窗口起点空仓（每窗独立开仓语义）。绩效只用窗内净值与成交计算，
+        上下文预热区不稀释 sharpe/年化/波动等时间口径指标（与组合级
+        ``_ComboWalkForwardBase._run_window`` 的 ``ec.iloc[lead:]`` 同口径）。
         """
         ctx_s = max(0, s - self._context_bars)
         lead = s - ctx_s  # 上下文 bar 数 = 需压制的信号数
@@ -226,9 +235,10 @@ class WalkForwardEngine:
         )
         try:
             bt = engine.run(sub)
-        except Exception:  # noqa: BLE001 — 单窗失败不拖垮整组，跳过该窗
+        except Exception as exc:  # noqa: BLE001 — 单窗失败不拖垮整组，跳过该窗
+            logger.warning("WF 第 %s 窗回测失败，跳过该窗：%s: %s", index, type(exc).__name__, exc)
             return None
-        perf = bt.performance
+        perf = self._window_performance(bt, lead)
 
         dt = self._dates(sub, lead)
         return WalkForwardWindow(
@@ -241,21 +251,48 @@ class WalkForwardEngine:
             max_drawdown=float(perf.get("max_drawdown", 0.0)),
             total_trades=int(perf.get("total_trades", 0)),
             win_rate=float(perf.get("win_rate", 0.0)),
-            performance={k: v for k, v in perf.items()},
+            performance=perf,
         )
 
+    @staticmethod
+    def _window_performance(bt: Any, lead: int) -> dict[str, Any]:
+        """只用窗内净值 + 成交重算绩效（上下文预热区不参与窗指标）。
+
+        上下文区净值恒为初始现金（warmup 压制信号），窗口内回撤/收益不变，
+        但 sharpe/年化/波动等按全序列（含上下文）计算会被零收益段稀释。
+        """
+        equity = bt.equity_curve
+        if len(equity) <= lead:
+            return dict(bt.performance)
+        window_equity = equity.iloc[lead:].reset_index(drop=True)
+        return dict(PerformanceAnalyzer(equity_curve=window_equity, trades=bt.trades).compute())
+
     @staticmethod
     def _dates(sub: pd.DataFrame, lead: int) -> tuple[str, str]:
-        """取窗口起止日期（跳过 lead 根上下文）。"""
+        """取窗口起止日期（跳过 lead 根上下文）。
+
+        int/np 整数（YYYYMMDD，TDX 日线原样）先 str 再按 %Y%m%d 解析——
+        直接 ``pd.Timestamp(int)`` 会被当纳秒换算成 1970 年。
+        """
         col = "datetime" if "datetime" in sub.columns else "date"
         vals = sub[col].iloc[lead:]
         if len(vals) == 0:
             return "", ""
         return (
-            pd.Timestamp(vals.iloc[0]).strftime("%Y-%m-%d"),
-            pd.Timestamp(vals.iloc[-1]).strftime("%Y-%m-%d"),
+            WalkForwardEngine._fmt_date(vals.iloc[0]),
+            WalkForwardEngine._fmt_date(vals.iloc[-1]),
         )
 
+    @staticmethod
+    def _fmt_date(v: Any) -> str:
+        """单个日期值 → YYYY-MM-DD（int/float YYYYMMDD 与 Timestamp/datetime64 兼容）。"""
+        ts: str
+        if isinstance(v, int | float | np.integer | np.floating) and not isinstance(v, bool):
+            ts = str(pd.to_datetime(str(int(v)), format="%Y%m%d").strftime("%Y-%m-%d"))
+        else:
+            ts = str(pd.Timestamp(v).strftime("%Y-%m-%d"))
+        return ts
+
     @staticmethod
     def _aggregate(result: WalkForwardResult) -> None:
         """聚合各窗指标（空列表安全）。"""
@@ -270,7 +307,8 @@ class WalkForwardEngine:
         result.worst_window = float(np.min(rets))
         result.best_window = float(np.max(rets))
         result.mean_sharpe = float(np.mean([w.sharpe for w in ws]))
-        result.worst_drawdown = float(min(w.max_drawdown for w in ws))
+        # max_drawdown 为正数幅度（(peak-total)/peak），“最差窗回撤”应取最大值
+        result.worst_drawdown = float(max(w.max_drawdown for w in ws))
         result.total_trades = int(sum(w.total_trades for w in ws))
 
 
@@ -352,7 +390,7 @@ class _ComboWalkForwardBase:
         for i in range(self._n_windows):
             s = eval_start + i * window_len
             e = s + window_len if i < self._n_windows - 1 else n  # 末窗吃到尾部
-            if e - s < 5:
+            if e - s < MIN_WINDOW_BARS:
                 continue
             win = self._run_window(timeline, s, e, i)
             if win is not None:
@@ -413,7 +451,14 @@ class _ComboWalkForwardBase:
             )
             try:
                 bt = engine.run(sub)
-            except Exception:  # noqa: BLE001 — 单槽位失败不拖垮整窗
+            except Exception as exc:  # noqa: BLE001 — 单槽位失败不拖垮整窗
+                logger.warning(
+                    "WF 组合第 %s 窗槽位 %s 回测失败，跳过：%s: %s",
+                    index,
+                    slot.key,
+                    type(exc).__name__,
+                    exc,
+                )
                 continue
 
             # 只取窗内净值点（上下文区恒为现金，不参与窗指标，避免稀释波动率）
@@ -453,7 +498,6 @@ class _ComboWalkForwardBase:
             if trade_frames
             else pd.DataFrame(columns=["symbol", "direction", "pnl", "rejected"])
         )
-        from easy_tdx.backtest.performance import PerformanceAnalyzer
 
         perf = PerformanceAnalyzer(equity_curve=window_equity, trades=all_trades).compute()
 

src/easy_tdx/ccpm/client.py

@@ -174,6 +174,20 @@ def parse_xml(text: str) -> list[dict[str, Any]]:
             "product": _g(node, "productid"),
         }
 
+    if not cells:
+        if len(root) > 0:
+            # XML 合法但 0 个 <data> 数据节点，且根元素下存在其他子结构——
+            # 官网模板/字段改版的典型形态（正常发布日必有 <data>；真实
+            # 「无数据日」走 302 → CcpmNoDataError）。静默返回空表会让
+            # 改版长期伪装成"无数据"，这里显式报错。
+            tags = ",".join(sorted({child.tag for child in root}))[:200]
+            raise CcpmError(
+                f"XML 解析出 0 条 <data> 数据节点（根元素 <{root.tag}>，"
+                f"子节点: {tags}）——中金所页面结构可能已变更"
+            )
+        # 根元素无任何子节点（如空 <positionRank/>）：按当日无数据处理
+        return []
+
     rows: list[dict[str, Any]] = []
     ranks = sorted({r for (_, _, r) in cells})
     for instrument in sorted(meta):

src/easy_tdx/cli/cmd_formula.py

@@ -23,6 +23,17 @@ def formula() -> None:
     """通达信公式：计算 / 选股 / 回测（命名布尔输出即信号）。"""
 
 
+def _parse_symbol(sym: str) -> tuple[str, str]:
+    """解析 ``市场:代码``；格式不对抛 click.BadParameter（而非裸 ValueError）。"""
+    market, sep, code = sym.strip().partition(":")
+    if not sep or not market.strip() or not code.strip() or ":" in code:
+        raise click.BadParameter(
+            f"标的格式应为 市场:代码（如 SH:600519），收到: {sym!r}",
+            param_hint="--symbols",
+        )
+    return market.strip().upper(), code.strip()
+
+
 def _load_formula(text: str | None, file: str | None) -> str:
     from easy_tdx.formula import compile_formula
 
@@ -130,12 +141,13 @@ def formula_screen(
         ]
     else:
         symbol_list = [s.strip() for s in symbols.split(",") if s.strip()]
+    # 前置校验全部标的格式（缺冒号等在发请求前就报错，而非循环中途裸崩）
+    symbol_pairs = [_parse_symbol(s) for s in symbol_list]
 
     compiled = compile_formula(source)
     hits: list[dict[str, Any]] = []
     errors: list[dict[str, str]] = []
-    for sym in symbol_list:
-        market, code = sym.split(":", 1)
+    for sym, (market, code) in zip(symbol_list, symbol_pairs, strict=True):
         try:
             df = _fetch(market, code, count, adjust)
             if df is None or len(df) == 0:

src/easy_tdx/cli/cmd_offline.py

@@ -561,6 +561,10 @@ def _sync_one_daily(client: TdxClient, filepath: Path) -> tuple[int, str]:
     # 协议返回的成交量单位是股；encode_daily_bar 在 vol_coeff=0.01 时按 ×100
     # 写入（.day 原始字段为股，读取端 ×0.01 还原为手），故写入前须换算为手。
     # 否则日成交 > 4295 万股（如招商银行等大盘股）的 bar 会溢出 uint32。
+    # 深市基金/ETF 同口径已实测锚定（2026-09-06）：sz159915.day 2026-06-12
+    # 原始 vol=1,249,608,655 与 get_security_bars 返回值完全一致，且
+    # amount/vol≈close——.day 原始字段与协议 vol 同为「股」，SZ_FUND
+    # （vol_coeff=0.01）走本分支换算正确，无需例外。
     if vol_coeff == 0.01:
         for b in bars:
             b.vol /= 100

src/easy_tdx/cli/cmd_warehouse.py

@@ -20,6 +20,32 @@ from typing import Any
 
 import click
 
+#: sync 支持的周期（与 Period 枚举名对齐，仓库按同名键存储）。
+_PERIOD_CHOICES = ["DAILY", "WEEKLY", "MONTHLY", "MIN_1", "MIN_5", "MIN_15", "MIN_30", "MIN_60"]
+#: baostock 是 EOD 源，只覆盖日线及以上（与 sources/baostock.py 的能力边界一致）。
+_BAOSTOCK_PERIODS = ("DAILY", "WEEKLY", "MONTHLY")
+
+
+def _parse_symbol(sym: str) -> tuple[str, str]:
+    """解析 ``市场:代码``；格式不对抛 click.BadParameter（而非裸 ValueError）。"""
+    market, sep, code = sym.strip().partition(":")
+    if not sep or not market.strip() or not code.strip() or ":" in code:
+        raise click.BadParameter(
+            f"标的格式应为 市场:代码（如 SH:600519），收到: {sym!r}",
+            param_hint="--symbols",
+        )
+    return market.strip().upper(), code.strip()
+
+
+def _validate_source_period(source: str, period: str) -> None:
+    """--source baostock 不支持的周期在参数层直接报错（而非静默空转 exit 0）。"""
+    if source == "baostock" and period.upper() not in _BAOSTOCK_PERIODS:
+        raise click.BadParameter(
+            f"--source baostock 仅支持日线及以上周期 {'/'.join(_BAOSTOCK_PERIODS)}，"
+            f"收到: {period}（baostock 为 EOD 源，无分钟线）",
+            param_hint="--period",
+        )
+
 
 def _require_warehouse(db_path: str | None) -> Any:
     """惰性导入 warehouse（duckdb 可选依赖），失败给友好错误。"""
@@ -40,7 +66,13 @@ def warehouse() -> None:
 @click.option(
     "--symbols", required=True, help="标的列表：逗号分隔（SH:600519,SZ:000001）或 @文件（每行一个）"
 )
-@click.option("--period", default="DAILY", help="K 线周期（默认 DAILY）")
+@click.option(
+    "--period",
+    "period",
+    default="DAILY",
+    type=click.Choice(_PERIOD_CHOICES, case_sensitive=False),
+    help="K 线周期（默认 DAILY；baostock 源仅支持日线及以上）",
+)
 @click.option("--max-bars", default=8000, type=int, help="首同步最大拉取根数（默认 8000）")
 @click.option("--tail-bars", default=15, type=int, help="增量同步尾部根数（默认 15）")
 @click.option(
@@ -72,6 +104,7 @@ def warehouse_sync(
     db_path: str | None,
 ) -> None:
     """增量同步行情进仓库（首同步全量、此后只补尾部）。"""
+    _validate_source_period(source, period)
     if symbols.startswith("@"):
         from pathlib import Path
 
@@ -86,6 +119,9 @@ def warehouse_sync(
         ]
     else:
         symbol_list = [s.strip() for s in symbols.split(",") if s.strip()]
+    # 前置校验全部标的格式（缺冒号等在发请求前就报错，而非循环中途裸崩）
+    for sym in symbol_list:
+        _parse_symbol(sym)
 
     with _require_warehouse(db_path) as wh:
         from easy_tdx.warehouse import WarehouseSyncer
@@ -100,7 +136,7 @@ def warehouse_sync(
             syncer = WarehouseSyncer(
                 kline_client, wh, max_bars=max_bars, tail_bars=tail_bars, adjust=adjust
             )
-            summary = syncer.sync(symbol_list, period=period, progress=_progress)
+            summary = syncer.sync(symbol_list, period=period.upper(), progress=_progress)
         else:
             from ..cli.conn import get_mac_client
 
@@ -115,13 +151,18 @@ def warehouse_sync(
                 syncer = WarehouseSyncer(
                     kline_client, wh, max_bars=max_bars, tail_bars=tail_bars, adjust=adjust
                 )
-                summary = syncer.sync(symbol_list, period=period, progress=_progress)
-        click.echo(
-            json.dumps(
-                {k: v for k, v in summary.items() if k != "details"},
-                ensure_ascii=False,
-            )
-        )
+                summary = syncer.sync(symbol_list, period=period.upper(), progress=_progress)
+        # source 口径标注进 summary（与 /bars 响应的 source 字段呼应）
+        payload = {k: v for k, v in summary.items() if k != "details"}
+        payload["source"] = source
+        # 失败明细打到 stderr（JSON 主体保持机器可读）
+        for d in summary.get("details", []):
+            if isinstance(d, dict) and d.get("error"):
+                click.echo(f"  {d['symbol']}: {d['error']}", err=True)
+        click.echo(json.dumps(payload, ensure_ascii=False))
+        if summary.get("failed"):
+            click.echo(f"错误: {summary['failed']}/{summary['total']} 个标的同步失败", err=True)
+            raise SystemExit(1)
 
 
 @warehouse.command("query")
@@ -189,11 +230,13 @@ def warehouse_check(symbols: str | None, db_path: str | None) -> None:
     with _require_warehouse(db_path) as wh:
         market = code = None
         if symbols:
-            first = [s.strip() for s in symbols.split(",") if s.strip()][0]
-            market, code = first.split(":", 1)
-            if ":" not in symbols and len(symbols.split(",")) > 1:
+            # 先校验「只支持一个标的」，再解析（旧码顺序颠倒：含冒号的多标的
+            # 绕过校验后在 split 处裸崩；无冒号单标的直接 ValueError）
+            sym_list = [s.strip() for s in symbols.split(",") if s.strip()]
+            if len(sym_list) > 1:
                 click.echo("错误: --symbols 自检模式一次只支持一个标的", err=True)
                 raise SystemExit(1)
+            market, code = _parse_symbol(sym_list[0])
         report = wh.health_check(market=market, code=code)
     click.echo(json.dumps(report, ensure_ascii=False, default=str))
     if report["issues"]:

src/easy_tdx/formula.py

@@ -43,6 +43,10 @@ import pandas as pd
 
 __all__ = ["FormulaError", "FormulaResult", "CompiledFormula", "compile_formula"]
 
+# 表达式嵌套深度上限（递归下降防爆栈：每层约 8 个 Python 栈帧，100 层
+# 远低于 CPython 默认递归上限，超出按 FormulaError 报错而非 RecursionError）
+_MAX_EXPRESSION_DEPTH = 100
+
 # ── Token ─────────────────────────────────────────────────────────────────────
 
 _TOKEN_RE = re.compile(
@@ -108,6 +112,7 @@ class _Parser:
     def __init__(self, tokens: list[_Token]) -> None:
         self._tokens = tokens
         self._i = 0
+        self._depth = 0
 
     def _peek(self) -> _Token:
         return self._tokens[self._i]
@@ -172,7 +177,13 @@ class _Parser:
 
     # 表达式优先级：OR < AND < 比较 < 加减 < 乘除 < 一元 < 原子
     def parse_expression(self) -> _Node:
-        return self._parse_or()
+        self._depth += 1
+        if self._depth > _MAX_EXPRESSION_DEPTH:
+            raise FormulaError(f"公式嵌套过深（超过 {_MAX_EXPRESSION_DEPTH} 层）")
+        try:
+            return self._parse_or()
+        finally:
+            self._depth -= 1
 
     def _parse_or(self) -> _Node:
         left = self._parse_and()
@@ -213,15 +224,22 @@ class _Parser:
         return left
 
     def _parse_unary(self) -> _Node:
-        if tok := self._match_op("-", "+"):
-            child = self._parse_unary()
-            if tok.value == "-":
-                return _Node(kind="un", value="neg", children=[child])
-            return child
-        if tok := self._match_op("!", "NOT"):
+        tok = self._match_op("-", "+", "!", "NOT")
+        if tok is None:
+            return self._parse_primary()
+        # 一元运算符链也计入深度（防 "!!!!…" 型超长链爆栈）
+        self._depth += 1
+        if self._depth > _MAX_EXPRESSION_DEPTH:
+            raise FormulaError(f"公式嵌套过深（超过 {_MAX_EXPRESSION_DEPTH} 层）")
+        try:
             child = self._parse_unary()
+        finally:
+            self._depth -= 1
+        if tok.value == "-":
+            return _Node(kind="un", value="neg", children=[child])
+        if tok.value in ("!", "NOT"):
             return _Node(kind="un", value="not", children=[child])
-        return self._parse_primary()
+        return child  # 一元正号
 
     def _parse_primary(self) -> _Node:
         tok = self._peek()
@@ -319,6 +337,18 @@ def _build_functions() -> dict[str, Callable[..., Any]]:
     ):
         if hasattr(mytt, name):
             fns[name] = getattr(mytt, name)
+
+    # REF 负移位 = 引用未来数据（未来函数），显式禁止。此前仅靠负数字面量
+    # 经一元负号转成 float 在 pandas 层报错这一巧合拦截。MyTT 库内直调
+    # （如 ICHIMOKU 迟行带画图 REF(C, -SHIFT)）不走公式白名单，不受影响。
+    def _ref_no_lookahead(S: Any, N: Any = 1) -> Any:
+        n_arr = np.asarray(N)
+        if n_arr.size and float(np.min(n_arr)) < 0:
+            raise FormulaError(f"REF 不允许负移位（引用未来数据）：N={N}")
+        return mytt.REF(S, N)
+
+    fns["REF"] = _ref_no_lookahead
+
     # numpy 补齐（TDX 语义）
     fns["POW"] = np.power
     fns["SQRT"] = np.sqrt
@@ -373,7 +403,10 @@ class _Evaluator:
 
     @staticmethod
     def _is_boolean(expr: _Node, val: Any) -> bool:
-        """输出归类：比较/逻辑/CROSS 节点或 0/1 值域 → 信号列。"""
+        """输出归类：比较/逻辑/CROSS 节点 → 信号列；否则仅当全部有限值
+        ∈ {0.0, 1.0} 才兜底判为信号（真 0/1 布尔指标）——含其他小数的
+        连续值（价格比率、归一化振荡器等）一律归数值列。
+        """
         if expr.kind in ("cmp", "logic"):
             return True
         if expr.kind == "call" and expr.value in _BOOL_FUNCS:
@@ -382,7 +415,7 @@ class _Evaluator:
         finite = arr[np.isfinite(arr)]
         if finite.size == 0:
             return False
-        return bool(finite.min() >= 0.0 and finite.max() <= 1.0)
+        return bool(np.all((finite == 0.0) | (finite == 1.0)))
 
     def eval(self, node: _Node) -> Any:
         if node.kind == "num":

src/easy_tdx/mac/commands/unusual.py

@@ -34,8 +34,12 @@ UNUSUAL_TYPE_NAMES: dict[int, str] = {
 }
 
 
-def _describe_unusual(unusual_type: int, data: bytes, hour: int = 9) -> tuple[str, str]:
-    """根据异动类型解析描述和数值。hour 用于区分竞价/尾盘双时刻信号（0x15）。"""
+def _describe_unusual(unusual_type: int, data: bytes, hour: int) -> tuple[str, str]:
+    """根据异动类型解析描述和数值。
+
+    hour 必传：来自报文时间槽（offset 29），用于区分竞价/尾盘双时刻信号
+    （0x15）——缺省会按 9 点把 15:00 的收盘信号误标成「竞价」。
+    """
     if len(data) < 13:
         return "", ""
     v1, v2, v3, v4 = struct.unpack_from("<B2fI", data)

src/easy_tdx/mac/qfq_check.py

@@ -145,7 +145,9 @@ def detect_ex_dividend_gaps(
     with np.errstate(divide="ignore", invalid="ignore"):
         ratio = open_arr[1:] / prev_close[:-1] - 1.0
     out: list[str] = []
-    for i in np.where(~np.isfinite(ratio) | (ratio < threshold))[0]:
+    # 仅「有限且低于阈值」计跳空：前收缺失（NaN/0）产生的非有限比率是数据
+    # 缺口而非除权（除权跳空必然是有限值），计入只会制造假 unexplained_gap。
+    for i in np.where(np.isfinite(ratio) & (ratio < threshold))[0]:
         out.append(_fmt(df["_dt"].iloc[i + 1]))
     return out
 

src/easy_tdx/realtime/feed.py

@@ -212,7 +212,11 @@ class RealtimeDataFeed:
                 :meth:`stop`。
         """
         if self._stop_requested:
-            return  # 启动前已请求停止（见 __init__ 的竞态说明）
+            # 启动前已请求停止（见 __init__ 的竞态说明）：消费该一次性标志并
+            # 立即退出；复位后同一实例可再次 start（修复 start→stop→start
+            # 二次启动被永久静默吞掉的问题）。
+            self._stop_requested = False
+            return
         self._running = True
         try:
             count = 0
@@ -262,7 +266,9 @@ class RealtimeDataFeed:
     async def _run_sync_loop(self, client: Any, max_iterations: int | None) -> None:
         """同步客户端的轮询循环：阻塞调用丢到 executor。"""
         if self._stop_requested:
-            return  # 启动前已请求停止（见 __init__ 的竞态说明）
+            # 同 run_async：消费一次性停止标志并复位（实例可复用）
+            self._stop_requested = False
+            return
         self._running = True
         try:
             count = 0
@@ -279,7 +285,8 @@ class RealtimeDataFeed:
         """请求停止轮询（下一轮 sleep 结束后生效）。
 
         在 ``run_async`` / ``run_sync`` 首次获得调度之前调用同样有效
-        （启动即退出），见 ``__init__`` 的竞态说明。
+        （启动即退出）；该停止请求是一次性的——被某次 run_* 消费后，
+        同一实例可以再次 start。
         """
         self._stop_requested = True
         self._running = False

src/easy_tdx/realtime/session.py

@@ -17,12 +17,22 @@
 
 from __future__ import annotations
 
-from datetime import datetime, time, tzinfo
+from datetime import datetime, time, timedelta, timezone, tzinfo
 from typing import Any
 
-__all__ = ["SESSION_WINDOWS", "SESSION_DESC", "is_trading_time", "session_info"]
+__all__ = [
+    "SHANGHAI_TZ",
+    "SESSION_WINDOWS",
+    "SESSION_DESC",
+    "is_trading_time",
+    "session_info",
+]
 
-#: 有效行情时段（本地时间）。窗口 = (start, end)，含两端。
+#: A 股行情统一按沪市时区判断。中国无夏令时，固定 UTC+8 即可精确表达，
+#: 不依赖系统时区/zoneinfo 数据库（Windows 无 IANA tzdata）。
+SHANGHAI_TZ = timezone(timedelta(hours=8), "Asia/Shanghai")
+
+#: 有效行情时段（沪市时间）。窗口 = (start, end)，含两端。
 #: - 早盘 09:15:00-11:30:30：09:15 起集合竞价可看，11:30:30 容纳尾单撮合散点；
 #: - 午盘 13:00:00-15:05:00：15:00-15:03 为收盘集合竞价，留 2 分钟余量。
 SESSION_WINDOWS: tuple[tuple[time, time], ...] = (
@@ -42,13 +52,14 @@ def is_trading_time(now: datetime | None = None, *, tz: tzinfo | None = None) ->
     而节假日行情本就不动，手动刷新始终可用）。
 
     Args:
-        now: 待判断时间，None = 取本地当前时间。
-        tz: 未传 ``now`` 时使用的时区，None = 系统本地时区。
+        now: 待判断时间，None = 取当前时间。
+        tz: 未传 ``now`` 时使用的时区，None = :data:`SHANGHAI_TZ`（与主机
+            时区无关；非中国时区的服务器/海外机器不会错位）。
 
     Returns:
         True = 盘中（含集合竞价缓冲窗）。
     """
-    t = now or datetime.now(tz=tz)
+    t = now or datetime.now(tz=tz or SHANGHAI_TZ)
     if t.weekday() >= 5:  # 周六/周日
         return False
     for start, end in SESSION_WINDOWS:
@@ -62,7 +73,7 @@ def session_info(now: datetime | None = None, *, tz: tzinfo | None = None) -> di
 
     前端以本地判断为主（每 15s 重估），本接口用于校准服务器侧视角。
     """
-    t = now or datetime.now(tz=tz)
+    t = now or datetime.now(tz=tz or SHANGHAI_TZ)
     return {
         "is_trading_time": is_trading_time(t),
         "sessions": [

src/easy_tdx/screen/limitup.py

@@ -17,7 +17,6 @@
 
 from __future__ import annotations
 
-import math
 from dataclasses import dataclass, field
 from pathlib import Path
 
@@ -34,13 +33,25 @@ __all__ = [
 ]
 
 
-def _round_price(x: float) -> float:
-    """四舍五入到分（Python round 是银行家舍入，交易所是四舍五入，不能混用）。"""
-    return math.floor(x * 100 + 0.5) / 100
+def _to_cents(price: float) -> int:
+    """元 → 分。.day 价格本身按 ×100 存 uint，round 消除读回的浮点表示误差。"""
+    return int(round(price * 100))
 
 
-def _eq_price(a: float, b: float) -> bool:
-    return abs(a - b) < 1e-4
+def _limit_price_cents(prev_cents: int, pct: int) -> int:
+    """交易所涨跌停价（分）：前收 × (1 ± pct%)，四舍五入到分（半进位）。
+
+    纯整数运算 ``(prev_cents * (100 + pct) + 50) // 100``，与交易所逐价位
+    对账零差异。不能用 float 乘完再 ``floor(x*100+0.5)``：乘法在半分边界
+    受浮点表示误差影响，±10% 档 67/318 个、±5% 档 90/884 个价位会算低
+    1 分（如 33.05×1.1 → 误算 36.35，交易所 36.36），导致真实涨跌停被漏判。
+    """
+    return (prev_cents * (100 + pct) + 50) // 100
+
+
+def _limit_price(prev: float, pct: int) -> float:
+    """交易所涨跌停价（元）：pct 为整数百分数（正=涨停档，负=跌停档）。"""
+    return _limit_price_cents(_to_cents(prev), pct) / 100.0
 
 
 def _limit_ratio(code: str) -> float:
@@ -113,27 +124,29 @@ def _entry_from_closes(
     pct = (closes[-1] / prev - 1.0) * 100.0
     entry = LimitUpEntry(code=code, market=market, pct=round(pct, 2))
 
-    up_ratio = _limit_ratio(code)
-    limit_up_price = _round_price(prev * (1 + up_ratio))
+    up_pct = 20 if _limit_ratio(code) == 0.20 else 10
+    # 全程分币整数比较，杜绝浮点舍入在半分边界错 1 分（漏判涨跌停）
+    cents = [_to_cents(c) for c in closes]
+    high_cents = _to_cents(last_high)
+    prev_c = cents[-2]
+    limit_up_c = _limit_price_cents(prev_c, up_pct)
+    limit_down_c = _limit_price_cents(prev_c, -up_pct)
     # 主板 5%：疑似 ST 涨停。低价股（< 3 元）最小报价单位 0.01 占比过大，
     # +5% 整的巧合概率骤增，跳过 ST 判定（宁可漏报不误报）。
-    st_applicable = up_ratio == 0.10 and prev >= 3.0
-    st_price = _round_price(prev * 1.05) if st_applicable else None
-    limit_down_price = _round_price(prev * (1 - up_ratio))
-    st_down_price = _round_price(prev * 0.95) if st_applicable else None
-
-    def _eq(a: float, b: float) -> bool:
-        return abs(a - b) < 1e-4
+    st_applicable = up_pct == 10 and prev >= 3.0
+    st_price_c = _limit_price_cents(prev_c, 5) if st_applicable else None
+    st_down_price_c = _limit_price_cents(prev_c, -5) if st_applicable else None
 
     def _is_up(i: int) -> bool:
-        """第 i 根是否涨停（用第 i-1 根收盘作前收）。"""
+        """第 i 根是否涨停（用第 i-1 根收盘作前收；ST/3 元门槛逐 bar 判定，
+        避免「按最新前收定性整段历史」在价格穿越 3 元时漏计/多计）。"""
         if i < 1:
             return False
-        p = closes[i - 1]
-        c = closes[i]
-        if _eq(c, _round_price(p * (1 + up_ratio))):
+        p_c = cents[i - 1]
+        c_c = cents[i]
+        if c_c == _limit_price_cents(p_c, up_pct):
             return True
-        return st_applicable and _eq(c, _round_price(p * 1.05))
+        return up_pct == 10 and p_c >= 300 and c_c == _limit_price_cents(p_c, 5)
 
     # 连板高度（截至最后一根）
     streak = 0
@@ -142,28 +155,28 @@ def _entry_from_closes(
         streak += 1
         i -= 1
     entry.streak = streak
-    entry.st = bool(streak > 0 and st_price is not None and _eq(closes[-1], st_price))
+    entry.st = bool(streak > 0 and st_price_c is not None and cents[-1] == st_price_c)
 
     if streak > 0:
         entry.blown = False
         return entry
 
-    # 未封住的场合：炸板（high 触及涨停价）或跌停
-    if _eq(last_high, limit_up_price):
+    # 未封住的场合：炸板（high 触及涨停价）或跌停。
+    # 口径说明：炸板仅按 10%/20% 档判定——.day 文件无法识别 ST，若对主板
+    # 额外按 5% 判炸板，非 ST 股恰好摸到 +5.00% 的会误报，故维持漏报方向。
+    if high_cents == limit_up_c:
         entry.blown = True
         return entry
 
-    if _eq(closes[-1], limit_down_price) or (
-        st_down_price is not None and _eq(closes[-1], st_down_price)
-    ):
+    if cents[-1] == limit_down_c or (st_down_price_c is not None and cents[-1] == st_down_price_c):
         down_streak = 0
         j = len(closes) - 1
         while j >= 1:
-            p = closes[j - 1]
-            c = closes[j]
-            hit = _eq(c, _round_price(p * (1 - up_ratio)))
-            if not hit and st_applicable:
-                hit = _eq(c, _round_price(p * 0.95))
+            p_c = cents[j - 1]
+            c_c = cents[j]
+            hit = c_c == _limit_price_cents(p_c, -up_pct) or (
+                up_pct == 10 and p_c >= 300 and c_c == _limit_price_cents(p_c, -5)
+            )
             if not hit:
                 break
             down_streak += 1
@@ -300,22 +313,21 @@ def compute_limitup_history(
             n_files += 1
             if n_files >= max_files:
                 break
-            up_ratio = _limit_ratio(code)
-            closes = [b.close for b in tail]
+            up_pct = 20 if _limit_ratio(code) == 0.20 else 10
+            cents = [_to_cents(b.close) for b in tail]
             date_ints = [b.year * 10000 + b.month * 100 + b.day for b in tail]
             for i in range(1, len(tail)):
-                p, c = closes[i - 1], closes[i]
-                if p <= 0:
+                p_c, c_c = cents[i - 1], cents[i]
+                if p_c <= 0:
                     continue
-                st_applicable = up_ratio == 0.10 and p >= 3.0
                 d = date_ints[i]
                 bucket = counts.setdefault(d, {"limit_up": 0, "limit_down": 0})
-                if _eq_price(c, _round_price(p * (1 + up_ratio))) or (
-                    st_applicable and _eq_price(c, _round_price(p * 1.05))
+                if c_c == _limit_price_cents(p_c, up_pct) or (
+                    up_pct == 10 and p_c >= 300 and c_c == _limit_price_cents(p_c, 5)
                 ):
                     bucket["limit_up"] += 1
-                elif _eq_price(c, _round_price(p * (1 - up_ratio))) or (
-                    st_applicable and _eq_price(c, _round_price(p * 0.95))
+                elif c_c == _limit_price_cents(p_c, -up_pct) or (
+                    up_pct == 10 and p_c >= 300 and c_c == _limit_price_cents(p_c, -5)
                 ):
                     bucket["limit_down"] += 1
         if n_files >= max_files:

src/easy_tdx/sources/baostock.py

@@ -8,20 +8,28 @@
   即启用；未安装时本模块整体静默关闭，核心功能零影响。
 - baostock 客户端是单条全局连接且非线程安全，本模块内部全程持锁串行，
   供 async 调用方经 ``asyncio.to_thread`` 使用。
-- 数据口径：volume 为股（与 /bars 输出契约一致，无需换算）；停牌日
-  （tradestatus=0 或 volume=0）剔除，与通达信 K 线不含停牌日的口径对齐；
-  复权经 adjustflag 原生支持（QFQ/HFQ/NONE），North Exchange（BJ）不覆盖。
+- 数据口径：个股 volume 为股（与 /bars 输出契约一致，无需换算）；指数
+  经 ``is_index=True`` 显式声明后 vol ÷100 转为手（baostock 指数 volume
+  单位为股，而 /bars/index 契约为手）；停牌日（tradestatus=0 或
+  volume=0）剔除，与通达信 K 线不含停牌日的口径对齐；复权经 adjustflag
+  原生支持（QFQ/HFQ/NONE），North Exchange（BJ）不覆盖。
+- 拉取失败（登录失败 / 查询 error_code≠0）记 warning 日志并抛
+  ``RuntimeError``——auto 兜底路径以 except 包裹调用不受影响，
+  ``--source baostock`` 显式使用时不会被伪装成"无数据"。
 """
 
 from __future__ import annotations
 
 import importlib
+import logging
 import os
 import threading
 from datetime import datetime, timedelta
 
 import pandas as pd
 
+logger = logging.getLogger(__name__)
+
 BAOSTOCK_DISABLE_ENV = "EASY_TDX_BAOSTOCK"
 
 # baostock 的全局连接锁（该库单连接、非线程安全）
@@ -33,6 +41,10 @@ _LOCK_TIMEOUT_SECONDS = 30.0
 
 # 支持兜底的周期（baostock frequency）：日线及以上；分钟线/季年线不兜
 _FREQ_BY_CATEGORY: dict[str, str] = {"DAY": "d", "WEEK": "w", "MONTH": "m"}
+# 请求字段：baostock 周线/月线不支持 tradestatus（实测 error_code=10004012
+# 「周线指标参数传入错误:tradestatus」，2026-09-06），仅日线可传。
+_FIELDS_DAILY = "date,open,high,low,close,volume,amount,tradestatus"
+_FIELDS_WEEKLY = "date,open,high,low,close,volume,amount"
 # 复权映射：baostock adjustflag — 1=后复权 2=前复权 3=不复权
 _ADJUST_FLAG = {"NONE": "3", "QFQ": "2", "HFQ": "1"}
 _MARKET_PREFIX = {"SZ": "sz", "SH": "sh"}  # BJ baostock 不覆盖
@@ -85,6 +97,7 @@ def fetch_bars(
     start: int,
     count: int,
     adjust: str,
+    is_index: bool = False,
 ) -> pd.DataFrame | None:
     """拉取日线及以上 K 线，输出对齐 /bars 契约的 DataFrame。
 
@@ -95,11 +108,19 @@ def fetch_bars(
         start: 跳过最新 start 根（与 TDX offset 语义一致）。
         count: 最多返回 count 根。
         adjust: "NONE" / "QFQ" / "HFQ"。
+        is_index: 标的是指数（如 sh.000001）。baostock 指数 volume 单位为
+            股，而 /bars/index 输出契约为手（通达信指数日线原样、周/月
+            ×100 还原后均为手）——True 时 vol ÷100 转手。实测
+            sh.000001 2026-09-04：volume=53,728,616,100 股
+            ÷100 = 537,286,161 手。
 
     Returns:
         按 [date, open, close, high, low, vol, amount] 列序、时间升序的
-        DataFrame；兜底不可用 / 不适用 / 无数据时返回 None（调用方继续
-        维持原错误，不吞异常）。
+        DataFrame；兜底不可用 / 不适用 / 无数据时返回 None。
+
+    Raises:
+        RuntimeError: baostock 登录或查询失败（已记 warning 日志）。auto
+            兜底调用方以 except 包裹即可维持原错误路径。
     """
     global _logged_in
     if not is_enabled():
@@ -116,46 +137,47 @@ def fetch_bars(
     coef, buffer_days = _WINDOW_DAYS[frequency]
     end_date = datetime.now()
     start_date = end_date - timedelta(days=total * coef + buffer_days)
+    fields = _FIELDS_DAILY if frequency == "d" else _FIELDS_WEEKLY
 
     # baostock 全局单连接：持锁串行；等待超时则放弃本次兜底
     if not _bs_lock.acquire(timeout=_LOCK_TIMEOUT_SECONDS):
         return None
     try:
         bs = importlib.import_module("baostock")
-        try:
-            _login_if_needed(bs)
-            rows = _query_rows(
-                bs,
-                code=f"{prefix}.{code}",
-                fields="date,open,high,low,close,volume,amount,tradestatus",
-                start_date=start_date.strftime("%Y-%m-%d"),
-                end_date=end_date.strftime("%Y-%m-%d"),
-                frequency=frequency,
-                adjustflag=adjustflag,
-            )
-        except Exception:
-            # 连接可能中途断开：重置登录态，下次兜底重新登录
-            _logged_in = False
-            raise
-    except Exception:
-        # 兜底源自身的任何失败都不向上抛：调用方按"无兜底数据"处理
-        return None
+        _login_if_needed(bs)
+        rows = _query_rows(
+            bs,
+            code=f"{prefix}.{code}",
+            fields=fields,
+            start_date=start_date.strftime("%Y-%m-%d"),
+            end_date=end_date.strftime("%Y-%m-%d"),
+            frequency=frequency,
+            adjustflag=adjustflag,
+        )
+    except Exception as exc:
+        # 连接可能中途断开：重置登录态，下次兜底重新登录。
+        # 真故障记日志并上抛——auto 兜底调用方（/bars 的 except 分支）接住
+        # 后维持原错误；显式 --source baostock 不会被伪装成"无数据"。
+        _logged_in = False
+        logger.warning("baostock 拉取失败（%s.%s %s）：%s", prefix, code, frequency, exc)
+        raise RuntimeError(f"baostock 拉取失败: {exc}") from exc
     finally:
         _bs_lock.release()
 
     if not rows:
         return None
-    df = pd.DataFrame(
-        rows, columns=["date", "open", "high", "low", "close", "vol", "amount", "tradestatus"]
-    )
+    df = pd.DataFrame(rows, columns=fields.split(",")).rename(columns={"volume": "vol"})
     for col in ("open", "high", "low", "close", "vol", "amount"):
         df[col] = pd.to_numeric(df[col], errors="coerce")
-    # 停牌日剔除（tradestatus=0 或无成交），对齐通达信 K 线不含停牌日的口径
+    # 停牌日剔除（tradestatus=0 或无成交），对齐通达信 K 线不含停牌日的口径。
+    # W/M 无 tradestatus 列（baostock 不支持），停牌周/月靠 vol>0 兜底剔除。
     if "tradestatus" in df.columns:
         df = df[df["tradestatus"] != "0"]
     df = df.dropna(subset=["close"])
     df = df[df["close"] > 0]
     df = df[df["vol"] > 0]
+    if is_index:
+        df["vol"] = df["vol"] / 100.0  # 股 → 手，见 docstring is_index 说明
     if df.empty:
         return None
     df["date"] = pd.to_datetime(df["date"]).dt.normalize()

src/easy_tdx/warehouse/store.py

@@ -29,6 +29,8 @@ from typing import Any
 
 import pandas as pd
 
+from easy_tdx.realtime.session import SHANGHAI_TZ
+
 logger = logging.getLogger(__name__)
 
 __all__ = ["KlineWarehouse", "default_warehouse_path", "MARKET_TO_TDX"]
@@ -36,6 +38,17 @@ __all__ = ["KlineWarehouse", "default_warehouse_path", "MARKET_TO_TDX"]
 # 未收盘 cutoff：15:05（A股 15:00 收盘 + 5 分钟数据落定余量）
 _MARKET_CLOSE_CUTOFF = dt_time(15, 5)
 
+
+def _shanghai_now() -> datetime:
+    """沪市墙钟（naive）：provisional 判定与主机时区无关。
+
+    非 UTC+8 主机（海外服务器）的系统本地时间会把「当日 / 15:05 前」判错
+    （UTC 主机上沪市 18:00 收盘后本地才 10:00，当日 bar 被误标 provisional
+    而被默认查询隐藏），故统一按 A 股时区取墙钟。中国无夏令时，固定 UTC+8。
+    """
+    return datetime.now(SHANGHAI_TZ).replace(tzinfo=None)
+
+
 MARKET_TO_TDX: dict[str, int] = {"SZ": 0, "SH": 1, "BJ": 2}
 _TDX_TO_MARKET: dict[int, str] = {v: k for k, v in MARKET_TO_TDX.items()}
 
@@ -90,7 +103,13 @@ class KlineWarehouse:
         self._duckdb = _require_duckdb()
         self._path = Path(db_path) if db_path is not None else default_warehouse_path()
         self._path.parent.mkdir(parents=True, exist_ok=True)
-        self._conn = self._duckdb.connect(str(self._path))
+        try:
+            self._conn = self._duckdb.connect(str(self._path))
+        except Exception as exc:
+            raise RuntimeError(
+                f"无法打开 K 线仓库 {self._path}：{exc}"
+                "（DuckDB 为单写者——请检查是否另有 easy-tdx 进程/CLI 正占用该文件）"
+            ) from exc
         self._conn.execute(_SCHEMA)
 
     # ── 基本属性 ─────────────────────────────────────────────────────────────
@@ -143,7 +162,7 @@ class KlineWarehouse:
             if c not in src.columns:
                 src[c] = float("nan")
 
-        now = datetime.now()
+        now = _shanghai_now()
         today = now.date()
         before_close = now.time() < _MARKET_CLOSE_CUTOFF
 
@@ -201,15 +220,44 @@ class KlineWarehouse:
         updated = len(rows) - inserted
         return (inserted, updated)
 
-    def promote_provisional(self) -> int:
-        """把「日期已过」的 provisional 行转正（收盘后的临时值已被次日增量覆盖）。"""
-        today = datetime.now().date()
+    def promote_provisional(
+        self,
+        market: str | None = None,
+        code: str | None = None,
+        before: datetime | pd.Timestamp | None = None,
+    ) -> int:
+        """把 provisional 行转正为 completed，返回转正行数。
+
+        两种用法：
+
+        - **同步流程（推荐）**：``promote_provisional(market, code,
+          before=max_dt)``——在拉取成功并 upsert 之后调用，只转正「本次成功
+          拉到的数据已覆盖」（datetime <= before）的行；拉取失败/为空时不
+          调用，盘中临时值不会被洗成 completed。
+        - **无参维护**：仅转正日期早于沪市今日的行（历史遗留清理）。
+
+        Args:
+            market: 限定市场（None = 全仓库）。
+            code: 限定标的（None = 全市场）。
+            before: 只转正 datetime <= 该时刻的行；None = 日期早于沪市今日。
+        """
+        conds = ["status = 'provisional'"]
+        params: list[Any] = []
+        if before is not None:
+            conds.append("datetime <= ?")
+            params.append(pd.Timestamp(before).to_pydatetime())
+        else:
+            conds.append("CAST(datetime AS DATE) < ?")
+            params.append(_shanghai_now().date())
+        if market is not None:
+            conds.append("market = ?")
+            params.append(market.upper())
+        if code is not None:
+            conds.append("code = ?")
+            params.append(code)
         cur = self._conn.execute(
-            """
-            UPDATE klines SET status = 'completed'
-            WHERE status = 'provisional' AND CAST(datetime AS DATE) < ?
-            """,
-            [today],
+            f"UPDATE klines SET status = 'completed' WHERE {' AND '.join(conds)}",
+            params,
         )
         return int(cur.fetchone()[0]) if cur.description else 0
 
@@ -347,7 +395,7 @@ class KlineWarehouse:
         ).df()
 
         issues: list[dict[str, Any]] = []
-        today = date.today()
+        today = _shanghai_now().date()
         stale: list[dict[str, Any]] = []
         total_provisional = 0
 
@@ -409,7 +457,7 @@ class KlineWarehouse:
                 "symbols_with_issues": len({i["symbol"] for i in issues}),
                 "stale_symbols": stale[:20],
                 "provisional_rows": total_provisional,
-                "checked_at": datetime.now().isoformat(timespec="seconds"),
+                "checked_at": _shanghai_now().isoformat(timespec="seconds"),
             },
         }
 

src/easy_tdx/warehouse/sync.py

@@ -4,8 +4,10 @@
 
 - **首同步全量**：仓库无该标的数据时按 ``max_bars``（默认 8000 根）拉取；
 - **增量补缺**：已有数据时只拉最近 ``tail_bars``（默认 15 根）覆盖——
-  覆盖同日 bar（收盘价修正 / provisional 转正），不动更早历史；
-- 同步前自动 :meth:`promote_provisional`（过期临时行转正）。
+  覆盖同日 bar（收盘价修正 / provisional 转正），不动更早历史；尾部窗口
+  覆盖不到上次同步点（超过 tail_bars 个交易日未同步）时自动改全量重拉；
+- provisional 转正在每标的**拉取成功后**进行，且只转正本次数据已覆盖的行
+  （拉取失败/为空的标的，盘中临时值保持 provisional，不会被洗成 completed）。
 
 客户端只需具备 ``get_stock_kline(market:int, code, period, start, count,
 adjust)`` 签名（``MacClient`` / ``AsyncMacClient`` 均可，本同步器只用同步
@@ -18,6 +20,8 @@ import logging
 from collections.abc import Callable
 from typing import Any
 
+import pandas as pd
+
 from easy_tdx.warehouse.store import MARKET_TO_TDX, KlineWarehouse
 
 logger = logging.getLogger(__name__)
@@ -34,6 +38,22 @@ def _period_name(period: str) -> str:
     return period.upper()
 
 
+def _df_time_column(df: pd.DataFrame) -> str | None:
+    """取 df 的时间列名（兼容 ``datetime`` / ``date`` 两种客户端输出）。"""
+    for col in ("datetime", "date"):
+        if col in df.columns:
+            return col
+    return None
+
+
+def _df_max_datetime(df: pd.DataFrame) -> pd.Timestamp | None:
+    """df 内最大 bar 时间（无时间列返回 None）。"""
+    col = _df_time_column(df)
+    if col is None:
+        return None
+    return pd.to_datetime(df[col]).max()
+
+
 class WarehouseSyncer:
     """把客户端行情增量同步进仓库。
 
@@ -69,6 +89,53 @@ class WarehouseSyncer:
         self._tail_bars = max(int(tail_bars), 5)
         self._adjust = adjust
 
+    def _fetch(
+        self,
+        market: str,
+        code: str,
+        period: str,
+        count: int,
+    ) -> pd.DataFrame:
+        return self._client.get_stock_kline(
+            MARKET_TO_TDX[market.upper()],
+            code,
+            period=period,
+            start=0,
+            count=count,
+            adjust=self._adjust,
+        )
+
+    def _refetch_full_on_gap(
+        self,
+        df: pd.DataFrame,
+        existing_last: pd.Timestamp | None,
+        market: str,
+        code: str,
+        period: str,
+        symbol: str,
+    ) -> pd.DataFrame:
+        """增量尾部覆盖不到上次同步点时全量重拉，消除静默缺口。
+
+        尾部只拉 ``tail_bars`` 根：若超过 tail_bars 个交易日未同步，窗口
+        首 bar 会晚于仓库最新 bar，中间日期不补就永久缺失——检测到该形态
+        （首 bar 时间 > existing_last）即改全量重拉一次。
+        """
+        if existing_last is None or df is None or len(df) == 0:
+            return df
+        col = _df_time_column(df)
+        if col is None:
+            return df
+        first_dt = pd.to_datetime(df[col]).min()
+        if first_dt <= existing_last:
+            return df
+        logger.warning(
+            "仓库同步 %s：增量尾部（首根 %s）覆盖不到上次同步点 %s，存在缺口，改全量重拉",
+            symbol,
+            first_dt,
+            existing_last,
+        )
+        return self._fetch(market, code, period, self._max_bars)
+
     def sync_symbol(
         self,
         market: str,
@@ -80,17 +147,17 @@ class WarehouseSyncer:
         try:
             existing_last = self._wh.last_datetime(market, code, period)
             count = self._tail_bars if existing_last is not None else self._max_bars
-            df = self._client.get_stock_kline(
-                MARKET_TO_TDX[market.upper()],
-                code,
-                period=period,
-                start=0,
-                count=count,
-                adjust=self._adjust,
-            )
+            df = self._fetch(market, code, period, count)
+            df = self._refetch_full_on_gap(df, existing_last, market, code, period, symbol)
             if df is None or len(df) == 0:
                 return {"symbol": symbol, "added": 0, "updated": 0, "skipped": 1, "error": None}
             added, updated = self._wh.upsert_bars(market, code, df, period=period)
+            # provisional 转正：只在拉取成功后进行，且只转正本次数据已覆盖
+            # （datetime <= 拉取最大时间）的行——拉取失败/为空时盘中临时值
+            # 不会被洗成 completed。
+            max_dt = _df_max_datetime(df)
+            if max_dt is not None:
+                self._wh.promote_provisional(market=market, code=code, before=max_dt)
             return {
                 "symbol": symbol,
                 "added": added,
@@ -118,7 +185,6 @@ class WarehouseSyncer:
         Returns:
             ``{"total", "ok", "added", "updated", "skipped", "failed", "details"}``。
         """
-        self._wh.promote_provisional()
         p = _period_name(period)
 
         parsed: list[tuple[str, str]] = []

src/easy_tdx/web/backtest_schemas.py

@@ -134,8 +134,14 @@ class OptimizeBacktestRequest(BaseModel):
     strategy: str = Field(..., description="策略名")
     cash: float = Field(default=1_000_000.0, gt=0)
     commission: float = Field(default=0.0003, ge=0, le=0.01)
+    min_commission: float = Field(default=5.0, ge=0)
+    stamp_tax: float = Field(default=0.001, ge=0, le=0.01)
     slippage: float = Field(default=0.0, ge=0, le=0.05)
     execution: Literal["next_open", "next_close"] = Field(default="next_open")
+    auto_fees: bool = Field(
+        default=False,
+        description="按标的品种自动解析费率（ETF/可转债免印花税等）；显式非默认费率优先",
+    )
     workers: int = Field(
         default=1,
         ge=0,
@@ -304,10 +310,14 @@ class BacktestResultResponse(BaseModel):
 
 
 class TaskSubmitResponse(BaseModel):
-    """后台任务提交响应。"""
+    """后台任务提交响应。
+
+    ``status`` 透传提交瞬间的真实状态：通常是 pending/running；极快任务在
+    拿到响应前可能已 done/failed——如实上报，前端轮询一次即见分晓。
+    """
 
     task_id: str
-    status: Literal["pending", "running"]
+    status: Literal["pending", "running", "done", "failed"]
 
 
 class TaskStateResponse(BaseModel):

src/easy_tdx/web/routers/backtest.py

@@ -40,6 +40,10 @@ from easy_tdx.web.task_runner import get_runner
 
 router = APIRouter(tags=["backtest"])
 
+# 标准 TdxClient 单次 get_security_bars 取数上限（协议约束，服务器对更大
+# 请求静默截断）。所有按标的取数路径都必须经 _fetch_bars_paged 翻页。
+_BARS_PAGE_SIZE = 800
+
 
 # ── 策略枚举 ───────────────────────────────────────────────────────────────────
 
@@ -104,10 +108,9 @@ async def run_backtest_async(
     # 3. 提交后台任务
     runner = get_runner()
     task_id = runner.submit(lambda: _run_backtest(df, snapshot), description=description)
+    # 提交瞬间通常是 pending/running；极快任务可能已 done/failed，如实上报
     state = runner.get(task_id)
-    # 提交瞬间任务应是 pending/running；极端情况下线程已跑完则报实际状态
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 @router.get("/backtest/tasks", response_model=TaskListResponse)
@@ -252,8 +255,7 @@ async def run_portfolio_backtest_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 # ── 多策略组合回测（资金分仓） ───────────────────────────────────────────────
@@ -285,8 +287,7 @@ async def run_multi_strategy_backtest_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 @router.post(
@@ -316,8 +317,7 @@ async def run_multi_strategy_walkforward_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 @router.post(
@@ -348,8 +348,7 @@ async def run_multi_strategy_evaluate_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 @router.post("/backtest/optimize/run/async", response_model=TaskSubmitResponse, status_code=202)
@@ -388,8 +387,7 @@ async def run_optimize_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 # ── 一键寻优所有策略 ───────────────────────────────────────────────────────────
@@ -429,8 +427,7 @@ async def run_optimize_all_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 # ── 信号雷达（一键扫描已保存策略）────────────────────────────────────────────
@@ -467,8 +464,7 @@ async def run_signal_scan_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 # ── Walk-Forward / 一条龙评估（v1.25 防过拟合链）──────────────────────────────
@@ -494,8 +490,7 @@ async def run_walkforward_async(
         lambda: _run_walkforward(df, snapshot, n_windows), description=description
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 @router.post("/backtest/evaluate/run/async", response_model=TaskSubmitResponse, status_code=202)
@@ -516,8 +511,7 @@ async def run_evaluate_async(
     runner = get_runner()
     task_id = runner.submit(lambda: _run_evaluate(df, snapshot), description=description)
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 # ── 组合级 Walk-Forward / 一条龙评估（对齐单标的防过拟合链）──────────────────
@@ -550,8 +544,7 @@ async def run_portfolio_walkforward_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 @router.post(
@@ -581,8 +574,7 @@ async def run_portfolio_evaluate_async(
         description=description,
     )
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 async def _resolve_df(client: Any, req: BacktestRequest) -> pd.DataFrame:
@@ -606,19 +598,10 @@ async def run_multiseed_async(
     平均收益）。结果含 per_seed_positive_ratio 稳定性列，通过
     GET /backtest/tasks/{task_id} 轮询。
     """
-    from easy_tdx.web.convert import category_from_str, market_from_str
-
     stock_dfs: dict[str, pd.DataFrame] = {}
     for symbol in req.stocks:
-        market_str, code = symbol.split(":", 1)
         try:
-            page = await client.get_security_bars(
-                market_from_str(market_str),
-                code,
-                category_from_str(req.category),
-                0,
-                req.count,
-            )
+            page = await _fetch_bars_paged(client, symbol, req.category, req.count)
         except Exception:  # noqa: BLE001 — 单标的失败跳过
             continue
         if len(page) >= 30:
@@ -631,8 +614,7 @@ async def run_multiseed_async(
     runner = get_runner()
     task_id = runner.submit(lambda: _run_multiseed(stock_dfs, snapshot), description=description)
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 def _run_multiseed(stock_dfs: dict[str, pd.DataFrame], req: MultiSeedRequest) -> dict[str, Any]:
@@ -781,15 +763,10 @@ async def run_rotation_async(
     可选槽内止盈止损。打分支持内置动量（``score="momentum"`` + ``period``）
     或通达信公式数值输出（``score="formula"`` + ``formula_text`` + ``score_col``）。
     """
-    from easy_tdx.web.convert import category_from_str, market_from_str
-
     stock_dfs: dict[str, pd.DataFrame] = {}
     for symbol in req.stocks:
-        market_str, code = symbol.split(":", 1)
         try:
-            page = await client.get_security_bars(
-                market_from_str(market_str), code, category_from_str(req.category), 0, req.count
-            )
+            page = await _fetch_bars_paged(client, symbol, req.category, req.count)
         except Exception:  # noqa: BLE001 — 单标的失败跳过
             continue
         if page is not None and len(page) >= 30:
@@ -802,8 +779,7 @@ async def run_rotation_async(
     runner = get_runner()
     task_id = runner.submit(lambda: _run_rotation(stock_dfs, snapshot), description=description)
     state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
-    return TaskSubmitResponse(task_id=task_id, status=status)
+    return TaskSubmitResponse(task_id=task_id, status=state.status)
 
 
 def _run_rotation(
@@ -916,18 +892,44 @@ def _normalize_bars_dt(df: pd.DataFrame) -> pd.DataFrame:
     return df
 
 
-async def _fetch_bars(client: Any, symbol: str, category: str, count: int) -> pd.DataFrame:
-    """按标的取 K 线（async，必须在 event loop 内调用）。"""
+async def _fetch_bars_paged(client: Any, symbol: str, category: str, count: int) -> pd.DataFrame:
+    """按 800/页翻页取最多 ``count`` 根 K 线，返回时间升序 DataFrame。
+
+    TDX 协议单次 get_security_bars 最多返回 800 根：count>800 的单次调用会被
+    服务器静默截断（multiseed / rotation / formula 曾各自单页取数，悄悄少
+    数据）。本辅助按 start=0,800,1600… 翻页拼接，页间按时间升序排序；
+    末页不足 800 根视为数据起点，提前停止。列结构与 get_security_bars
+    原始输出一致（日线 ``date`` / 分钟 ``datetime``），不做改名/类型规整。
+    """
     from easy_tdx.web.convert import category_from_str, market_from_str
 
     market_str, code = symbol.split(":", 1)
-    df = await client.get_security_bars(
-        market_from_str(market_str),
-        code,
-        category_from_str(category),
-        0,
-        count,
-    )
+    market = market_from_str(market_str)
+    cat = category_from_str(category)
+    frames: list[pd.DataFrame] = []
+    fetched = 0
+    while fetched < count:
+        page_size = min(_BARS_PAGE_SIZE, count - fetched)
+        page_df = await client.get_security_bars(market, code, cat, fetched, page_size)
+        if page_df is None or len(page_df) == 0:
+            break
+        frames.append(page_df)
+        fetched += len(page_df)
+        if len(page_df) < page_size:
+            break  # 数据起点
+    if not frames:
+        return pd.DataFrame()
+    df = pd.concat(frames, ignore_index=True)
+    dt_col = "datetime" if "datetime" in df.columns else "date"
+    if dt_col in df.columns:
+        # 页间天然逆序（page0=最新一页），拼接后按时间升序
+        df = df.sort_values(dt_col).reset_index(drop=True)
+    return df
+
+
+async def _fetch_bars(client: Any, symbol: str, category: str, count: int) -> pd.DataFrame:
+    """按标的取 K 线（async，必须在 event loop 内调用）。"""
+    df = await _fetch_bars_paged(client, symbol, category, count)
     if len(df) == 0:
         raise ValueError(f"标的 {symbol} 未取到任何 K 线数据")
     return _normalize_bars_dt(df)
@@ -1182,31 +1184,67 @@ def _run_multi_strategy_evaluate(
     )
 
 
+def _resolve_effective_fees(
+    auto_fees: bool,
+    symbol: str | None,
+    commission: float,
+    min_commission: float,
+    stamp_tax: float,
+) -> tuple[float, float, float]:
+    """auto_fees 品种费率解析（与 BacktestEngine 同款口径）。
+
+    显式非默认值优先（调用方有意覆盖），默认值按品种费率表替换（如
+    ETF/可转债免印花税）。ParamGridOptimizer 无 auto_fees 参数，寻优端点
+    在 web 层预解析成具体费率再传入，保证与单标的回测同口径。
+    """
+    if not auto_fees or not symbol:
+        return commission, min_commission, stamp_tax
+    from easy_tdx.backtest.fees import resolve_fee_model
+
+    fee = resolve_fee_model(symbol)
+    if commission == 0.0003:
+        commission = fee.commission
+    if min_commission == 5.0:
+        min_commission = fee.min_commission
+    if stamp_tax == 0.001:
+        stamp_tax = fee.stamp_tax
+    return commission, min_commission, stamp_tax
+
+
 def _run_optimize(df: pd.DataFrame, req: OptimizeBacktestRequest) -> dict[str, Any]:
     """执行参数网格寻优并返回清洗后的结果字典（后台线程内调用）。"""
     from easy_tdx.backtest.benchmark import run_buy_hold_benchmark
     from easy_tdx.backtest.optimizer import ParamGridOptimizer
 
+    commission, min_commission, stamp_tax = _resolve_effective_fees(
+        req.auto_fees, req.symbol, req.commission, req.min_commission, req.stamp_tax
+    )
     optimizer = ParamGridOptimizer(
         strategy_name=req.strategy,
         param_grid=req.param_grid,
         df=df,
         cash=req.cash,
-        commission=req.commission,
+        commission=commission,
+        min_commission=min_commission,
+        stamp_tax=stamp_tax,
         slippage=req.slippage,
         execution=req.execution,
         workers=req.workers,
     )
     result = optimizer.run()
     out = result.to_dict()
-    # 买入持有基准（同区间/同费率/同资金，与一条龙评估同口径），
-    # 供前端在最优结果旁直观对比「策略 vs 买入不动」。
+    # 买入持有基准（同区间/同费率/同资金，与一条龙评估同口径），供前端在
+    # 最优结果旁直观对比「策略 vs 买入不动」。
     out["buy_hold"] = run_buy_hold_benchmark(
         df,
         cash=req.cash,
-        commission=req.commission,
+        commission=commission,
+        min_commission=min_commission,
+        stamp_tax=stamp_tax,
         slippage=req.slippage,
         execution=req.execution,
+        symbol=req.symbol,
+        auto_fees=req.auto_fees,
     )
     return out
 

src/easy_tdx/web/routers/bars.py

@@ -218,13 +218,24 @@ async def _fetch_120m(
 
 
 async def _baostock_last_resort(
-    market: str, code: str, category: str, start: int, count: int, adjust: str
+    market: str,
+    code: str,
+    category: str,
+    start: int,
+    count: int,
+    adjust: str,
+    is_index: bool = False,
 ) -> tuple[pd.DataFrame | None, str | None]:
     """TDX 全部路径失败/为空后的最后一级兜底：baostock（仅日线及以上）。
 
     未安装 baostock / 设置了 EASY_TDX_BAOSTOCK=0 / 周期不适用 / 查询失败
-    一律返回 ``(None, None)``——兜底源自身的任何失败都不影响原错误语义。
-    baostock 客户端阻塞且非线程安全：丢线程池执行，模块内部持锁串行。
+    一律返回 ``(None, None)``——兜底源自身的任何失败（含新版 fetch_bars 对
+    真故障抛出的 RuntimeError）都按"兜底不可用"处理，调用方继续维持原
+    TDX 错误语义。baostock 客户端阻塞且非线程安全：丢线程池执行，模块内部
+    持锁串行。
+
+    Args:
+        is_index: 标的是指数（/bars/index 兜底传 True，vol 股→手对齐契约）。
     """
     from easy_tdx.sources import baostock as baostock_source
 
@@ -232,14 +243,28 @@ async def _baostock_last_resort(
         return None, None
     try:
         df = await asyncio.to_thread(
-            baostock_source.fetch_bars, market, code, category, start, count, adjust
+            baostock_source.fetch_bars,
+            market,
+            code,
+            category,
+            start,
+            count,
+            adjust,
+            is_index,
         )
     except Exception as exc:  # noqa: BLE001 — 兜底失败不改变原错误路径
         _logger.warning("/bars baostock 兜底异常 (%s%s): %s", market, code, exc)
         return None, None
     if df is None or df.empty:
         return None, None
-    _logger.info("/bars 已启用 baostock 兜底 (%s%s %s，%d 根)", market, code, category, len(df))
+    _logger.info(
+        "/bars 已启用 baostock 兜底 (%s%s %s%s，%d 根)",
+        market,
+        code,
+        category,
+        "，指数" if is_index else "",
+        len(df),
+    )
     return df, "baostock"
 
 
@@ -339,7 +364,9 @@ async def security_bars(
             _logger.warning("/bars 标准 TdxClient 获取失败 (%s%s): %s", market, code, exc)
 
     if df is None or df.empty:
-        bdf, bsource = await _baostock_last_resort(market, code, category, start, count, adjust)
+        # 周期先归一成枚举名（"4"→DAY）：baostock 频率查表只认名称，
+        # 数字串直接透传会让兜底静默失效
+        bdf, bsource = await _baostock_last_resort(market, code, cat.name, start, count, adjust)
         if bdf is not None:
             df, source = bdf, bsource
 
@@ -391,7 +418,9 @@ async def index_bars(
         _logger.warning("/bars/index TdxClient 获取失败 (%s%s): %s", market, code, exc)
 
     if df is None or df.empty:
-        bdf, bsource = await _baostock_last_resort(market, code, category, start, count, "QFQ")
+        bdf, bsource = await _baostock_last_resort(
+            market, code, category_from_str(category).name, start, count, "QFQ", is_index=True
+        )
         if bdf is not None:
             df, source = bdf, bsource
 

src/easy_tdx/web/routers/board_mac.py

@@ -12,6 +12,7 @@ import pandas as pd
 from fastapi import APIRouter, Depends, Query
 
 from easy_tdx.mac.enums import Adjust, Period
+from easy_tdx.realtime.session import SHANGHAI_TZ, is_trading_time
 from easy_tdx.web.convert import (
     board_sort_from_str,
     board_type_from_str,
@@ -20,7 +21,7 @@ from easy_tdx.web.convert import (
     sort_type_from_str,
 )
 from easy_tdx.web.deps import get_mac_client
-from easy_tdx.web.schemas import DataFrameResponse, DictResponse
+from easy_tdx.web.schemas import DataFrameResponse, DictResponse, _json_safe
 
 _logger = logging.getLogger(__name__)
 
@@ -37,9 +38,9 @@ _OVERVIEW_METRIC_FIELDS: dict[str, str] = {
     "YTD": "chg_ytd",
 }
 _OVERVIEW_TTL = 15.0
-# (board_type, metrics) -> (monotonic 截止时间, payload)。无锁：并发重复拉取
-# 无害（AsyncMacClient 连接内本就串行），省去跨事件循环的锁生命周期问题。
-_overview_cache: dict[tuple[str, tuple[str, ...]], tuple[float, dict[str, Any]]] = {}
+# (board_type, metrics, count) -> (monotonic 截止时间, payload)。无锁：并发
+# 重复拉取无害（AsyncMacClient 连接内本就串行），省去跨事件循环的锁生命周期问题。
+_overview_cache: dict[tuple[str, tuple[str, ...], int], tuple[float, dict[str, Any]]] = {}
 
 # 可在单测中 monkeypatch 以控制 TTL 判定
 _now = time.monotonic
@@ -181,7 +182,7 @@ async def board_overview(
         valid = ", ".join(_OVERVIEW_METRIC_FIELDS)
         raise ValueError(f"无效指标 '{','.join(invalid)}'，可选值: {valid}")
 
-    cache_key = (bt.name, tuple(sort_names))
+    cache_key = (bt.name, tuple(sort_names), count)  # count 影响 payload 行数，必须入键
     cached = _overview_cache.get(cache_key)
     if cached is not None and _now() < cached[0]:
         return DictResponse.from_dict(cached[1])
@@ -231,7 +232,12 @@ async def board_overview(
                 row.setdefault(field, None)
             rows.append(row)
 
-    payload = {"board_type": bt.name, "ts": int(time.time()), "count": len(rows), "rows": rows}
+    # 坏值（NaN/inf，如无足够历史板块的 sort_value）在入缓存前清洗成 null：
+    # 带 NaN 的 payload 一旦入缓存，15s TTL 内每次命中都会在 JSON 序列化时
+    # 500（Starlette allow_nan=False）。
+    payload = _json_safe(
+        {"board_type": bt.name, "ts": int(time.time()), "count": len(rows), "rows": rows}
+    )
     _overview_cache[cache_key] = (_now() + _OVERVIEW_TTL, payload)
     return DictResponse.from_dict(payload)
 
@@ -263,8 +269,8 @@ _hotspot_builds: dict[str, dict[str, Any]] = {}
 
 
 def _today_str() -> str:
-    """当日日历日（缓存失效键；单测可 monkeypatch）。"""
-    return datetime.now().strftime("%Y-%m-%d")
+    """当日日历日（沪市时区，与主机时区无关；缓存失效键；单测可 monkeypatch）。"""
+    return datetime.now(SHANGHAI_TZ).strftime("%Y-%m-%d")
 
 
 async def _hotspot_build(board_key: str, bt: Any, client: Any) -> None:
@@ -500,8 +506,6 @@ async def board_hotspot(
     )
     rows_out = rows_out[:_HOTSPOT_MAX_ROWS]
 
-    from easy_tdx.realtime.session import is_trading_time
-
     payload: dict[str, Any] = {
         "status": "ready",
         "board_type": bt.name,

src/easy_tdx/web/routers/formula.py

@@ -131,7 +131,11 @@ async def run_formula_backtest_async(
         lambda: _run_formula_backtest(df, snapshot),
         description=f"公式回测 | {snapshot.symbol or '内联数据'}",
     )
-    return {"task_id": task_id, "status": "running"}
+    try:
+        status = runner.get(task_id).status
+    except KeyError:  # 极端：状态尚未可查时按提交默认态上报
+        status = "running"
+    return {"task_id": task_id, "status": status}
 
 
 @router.post("/formula/screen/run/async", status_code=202)
@@ -167,7 +171,11 @@ async def run_formula_screen_async(
         lambda: _run_formula_screen(bars, compiled, snapshot.signal_col),
         description=f"公式选股 | {len(bars)}只标的",
     )
-    return {"task_id": task_id, "status": "running"}
+    try:
+        status = runner.get(task_id).status
+    except KeyError:  # 极端：状态尚未可查时按提交默认态上报
+        status = "running"
+    return {"task_id": task_id, "status": status}
 
 
 # ── 内部实现 ───────────────────────────────────────────────────────────────────
@@ -187,12 +195,11 @@ async def _resolve_df(client: Any, req: FormulaComputeRequest) -> Any:
             df["datetime"] = pd.to_datetime(df["datetime"], errors="coerce")
         return df
     if req.symbol is not None:
-        from easy_tdx.web.convert import category_from_str, market_from_str
+        # 经 _fetch_bars_paged 按 800/页翻页取全量（协议单次上限 800，
+        # 单次调用 count>800 会被服务器静默截断，指标计算窗口悄悄变短）
+        from easy_tdx.web.routers.backtest import _fetch_bars_paged
 
-        market_str, code = req.symbol.split(":", 1)
-        df = await client.get_security_bars(
-            market_from_str(market_str), code, category_from_str(req.category), 0, req.count
-        )
+        df = await _fetch_bars_paged(client, req.symbol, req.category, req.count)
         if df is None or len(df) == 0:
             raise ValueError(f"标的 {req.symbol} 未取到 K 线数据")
         return df

src/easy_tdx/web/routers/llm.py

@@ -226,8 +226,10 @@ async def llm_chat_async(req: LlmChatRequest) -> TaskSubmitResponse:
 
     runner = get_runner()
     task_id = runner.submit(_run, description=desc)
-    state = runner.get(task_id)
-    status: Any = state.status if state.status in ("pending", "running") else "running"
+    try:
+        status = runner.get(task_id).status
+    except KeyError:  # 极端：状态尚未可查时按提交默认态上报
+        status = "running"
     return TaskSubmitResponse(task_id=task_id, status=status)
 
 

src/easy_tdx/web/routers/market.py

@@ -21,11 +21,13 @@ from easy_tdx.web.schemas import (
 
 router = APIRouter(tags=["market"])
 
-# 涨停生态结果缓存（vipdoc 盘中随通达信客户端落盘更新，60s 足够新鲜）
-_limitup_cache: tuple[float, dict[str, Any]] | None = None
+# 涨停生态结果缓存（vipdoc 盘中随通达信客户端落盘更新，60s 足够新鲜）。
+# 键 = effective vipdoc（显式参数 > 已存设置 > None）：不同数据目录的扫描
+# 结果必须互不串台。
+_limitup_cache: dict[str | None, tuple[float, dict[str, Any]]] = {}
 _LIMITUP_TTL = 60.0
-# 涨停逐日历史缓存（历史数据不变，10 分钟；按 days 分键）
-_limitup_history_cache: dict[int, tuple[float, dict[str, Any]]] = {}
+# 涨停逐日历史缓存（历史数据不变，10 分钟；键 = (days, effective vipdoc)）
+_limitup_history_cache: dict[tuple[int, str | None], tuple[float, dict[str, Any]]] = {}
 
 
 def _df_response(df: Any) -> DataFrameResponse:
@@ -148,8 +150,8 @@ async def set_vipdoc_setting(req: dict[str, Any]) -> dict[str, Any]:
         get_app_settings_store().delete(_VIPDOC_KEY)
         resolved = None
     # 路径变更后旧扫描结果作废
-    global _limitup_cache, _limitup_history_cache
-    _limitup_cache = None
+    global _limitup_cache
+    _limitup_cache.clear()
     _limitup_history_cache.clear()
     return {"stored": path, "resolved": resolved}
 
@@ -165,12 +167,11 @@ async def limitup_ecology(
     涨停判定按代码段：主板 10%（含 5% 疑似 ST 标记）、创业板/科创板 20%；
     .day 文件无名称，name 由前端经批量报价补齐。
     """
-    global _limitup_cache
-    now = time.monotonic()
-    if _limitup_cache is not None and now - _limitup_cache[0] < _LIMITUP_TTL:
-        return DictResponse.from_dict(_limitup_cache[1])
-
     effective = _effective_vipdoc(vipdoc)
+    now = time.monotonic()
+    cached = _limitup_cache.get(effective)
+    if cached is not None and now - cached[0] < _LIMITUP_TTL:
+        return DictResponse.from_dict(cached[1])
 
     def _scan() -> dict[str, Any]:
         from easy_tdx.screen.limitup import compute_limitup_ecology
@@ -194,7 +195,7 @@ async def limitup_ecology(
         }
 
     payload = await asyncio.to_thread(_scan)
-    _limitup_cache = (now, payload)
+    _limitup_cache[effective] = (now, payload)
     return DictResponse.from_dict(payload)
 
 
@@ -244,20 +245,20 @@ async def limitup_history(
 
     全市场扫描约需数十秒，结果缓存 10 分钟。日期覆盖受 vipdoc 数据范围限制。
     """
-    global _limitup_history_cache
+    effective = _effective_vipdoc(vipdoc)
     now = time.monotonic()
-    cached = _limitup_history_cache.get(days)
+    cached = _limitup_history_cache.get((days, effective))
     if cached is not None and now - cached[0] < 600:
         return DictResponse.from_dict(cached[1])
 
     def _scan() -> dict[str, Any]:
         from easy_tdx.screen.limitup import compute_limitup_history
 
-        rows = compute_limitup_history(_effective_vipdoc(vipdoc), days=days)
+        rows = compute_limitup_history(effective, days=days)
         return {"count": len(rows), "days": rows}
 
     payload = await asyncio.to_thread(_scan)
-    _limitup_history_cache[days] = (now, payload)
+    _limitup_history_cache[(days, effective)] = (now, payload)
     return DictResponse.from_dict(payload)
 
 

src/easy_tdx/web/routers/server.py

@@ -8,15 +8,20 @@
 from __future__ import annotations
 
 import asyncio
+from typing import Annotated
 
 from fastapi import APIRouter, Request
-from pydantic import BaseModel
+from pydantic import BaseModel, Field, StringConstraints
 
 from easy_tdx.config import get_best_host, get_known_hosts, get_port, save_best_host
 from easy_tdx.transport.sync import ping_all
 
 router = APIRouter(tags=["server"])
 
+# 单项 host 长度上限（域名合法最大 253 字符）；列表项数上限防被当作
+# 无限制的内网扫描跳板。
+HostStr = Annotated[str, StringConstraints(max_length=253)]
+
 
 # --------------------------------------------------------------------------- #
 # Schemas
@@ -41,10 +46,16 @@ class HostListResponse(BaseModel):
 
 
 class ServerTestRequest(BaseModel):
-    """POST /server/test 的请求。"""
+    """POST /server/test 的请求体。
 
-    hosts: list[str] | None = None  # None = 测全部候选
-    timeout: float = 5.0
+    ``timeout`` 限 0.5~30s（to_thread 内的同步 ping 无中断手段，无上界的
+    超时会长期占住线程池线程）；``hosts`` 限 50 项、单项 ≤253 字符。
+    """
+
+    hosts: list[HostStr] | None = Field(
+        default=None, max_length=50, description="待测主机列表；None = 测全部候选"
+    )
+    timeout: float = Field(default=5.0, ge=0.5, le=30.0, description="单主机连接超时（秒）")
 
 
 class ServerSwitchRequest(BaseModel):

src/easy_tdx/web/routers/watchlist.py

@@ -3,18 +3,22 @@
 from __future__ import annotations
 
 from fastapi import APIRouter, HTTPException, Query
+from fastapi import Path as PathParam
 from pydantic import BaseModel, Field
 
 from easy_tdx.web.watchlist_store import get_watchlist_store
 
 router = APIRouter(tags=["watchlist"])
 
+# 6 位数字代码（自选会被 QuoteStreamer 拿去轮询，非数字代码产生无效请求）
+_CODE_PATTERN = r"^\d{6}$"
+
 
 class WatchItemAdd(BaseModel):
     """加入自选请求。name 由前端从行情数据带过来。"""
 
     market: str = Field(..., pattern=r"^(SZ|SH|BJ)$")
-    code: str = Field(..., min_length=6, max_length=6)
+    code: str = Field(..., pattern=_CODE_PATTERN)
     name: str = Field("", max_length=64)
     group: str = Field("默认", max_length=32)
 
@@ -41,7 +45,10 @@ async def add_watch_item(req: WatchItemAdd) -> dict[str, object]:
 
 
 @router.delete("/watchlist/{market}/{code}", response_model=dict[str, object])
-async def remove_watch_item(market: str, code: str) -> dict[str, object]:
+async def remove_watch_item(
+    market: str,
+    code: str = PathParam(..., pattern=_CODE_PATTERN, description="6位数字代码"),
+) -> dict[str, object]:
     """移除自选。"""
     if market.upper() not in {"SZ", "SH", "BJ"}:
         raise HTTPException(status_code=400, detail=f"非法市场: {market}")

src/easy_tdx/web/schemas.py

@@ -2,11 +2,38 @@
 
 from __future__ import annotations
 
+import math
 from enum import IntEnum
 from typing import Any
 
 from pydantic import BaseModel, Field
 
+
+def _json_safe(v: Any) -> Any:
+    """递归把值清洗为 JSON 原生类型：NaN/±inf → None、datetime → ISO 串、
+    numpy 标量 → Python 原生、容器逐项处理。
+
+    Starlette 的 JSONResponse 以 ``allow_nan=False`` 序列化，任何 NaN/inf
+    漏出去都会让整个响应 500（v1.32 实测：/board-mac/overview 某行
+    sort_value=NaN → 全端点 500 且带毒 payload 入 15s 缓存）。所有
+    DictResponse / 缓存写入路径都应先过本函数。
+    """
+    # bool 是 int 子类，须先判
+    if v is None or isinstance(v, bool | str | int):
+        return v
+    if isinstance(v, float):
+        return None if (math.isnan(v) or math.isinf(v)) else v
+    if hasattr(v, "isoformat"):  # datetime/date/pd.Timestamp
+        return v.isoformat()
+    if hasattr(v, "item"):  # numpy 标量（含 np.float32 NaN）
+        return _json_safe(v.item())
+    if isinstance(v, dict):
+        return {str(k): _json_safe(val) for k, val in v.items()}
+    if isinstance(v, list | tuple):
+        return [_json_safe(item) for item in v]
+    return v
+
+
 # ---------------------------------------------------------------------------
 # Enums — mirror easy_tdx.models.enums but as string-based for REST clarity
 # ---------------------------------------------------------------------------
@@ -143,19 +170,15 @@ class DictResponse(BaseModel):
 
     @classmethod
     def from_dict(cls, d: dict[str, Any]) -> DictResponse:
-        """序列化 dict，将其中的 DataFrame 转为 records 格式。"""
+        """序列化 dict：DataFrame 转 records，值递归清洗（NaN/inf → null 等）。"""
         import pandas as pd
 
         cleaned: dict[str, Any] = {}
         for k, v in d.items():
             if isinstance(v, pd.DataFrame):
                 cleaned[k] = DataFrameResponse.from_dataframe(v).data
-            elif hasattr(v, "isoformat"):
-                cleaned[k] = v.isoformat()
-            elif hasattr(v, "item"):
-                cleaned[k] = v.item()
             else:
-                cleaned[k] = v
+                cleaned[k] = _json_safe(v)
         return cls(data=cleaned)
 
 

src/easy_tdx/web/sentiment_sampler.py

@@ -19,7 +19,7 @@ import logging
 from datetime import datetime
 from typing import Any
 
-from easy_tdx.realtime.session import is_trading_time
+from easy_tdx.realtime.session import SHANGHAI_TZ, is_trading_time
 from easy_tdx.web.sentiment_store import SentimentStore, get_sentiment_store
 
 logger = logging.getLogger(__name__)
@@ -81,7 +81,9 @@ class SentimentSampler:
         if df is None or df.empty:
             raise RuntimeError("get_market_stat 返回空数据")
         row = df.iloc[0]
-        now = datetime.now()
+        # (date, minute) 键按沪市时区取"现在"：与 is_trading_time 的时段判断
+        # 同一参照系，主机时区非 UTC+8（海外服务器）时不会整体错位
+        now = datetime.now(SHANGHAI_TZ)
         self._store.insert(
             {
                 "date": now.year * 10000 + now.month * 100 + now.day,
@@ -141,7 +143,7 @@ class FundFlowSampler:
         logger.info("FundFlowSampler 启动（间隔 %ss，交易日 14:45 后每日一条）", self._interval)
         while True:
             try:
-                now = datetime.now()
+                now = datetime.now(SHANGHAI_TZ)
                 if is_trading_time(now) and (now.hour * 100 + now.minute) >= 1445:
                     await self._sample_once()
             except asyncio.CancelledError:
@@ -153,7 +155,7 @@ class FundFlowSampler:
     async def _sample_once(self) -> None:
         from easy_tdx.mac.enums import BoardType
 
-        today = int(datetime.now().strftime("%Y%m%d"))
+        today = int(datetime.now(SHANGHAI_TZ).strftime("%Y%m%d"))
         if self._store.latest_fund_date() == today:
             return  # 当日已采样
         df = await self._client.get_board_ranking(

src/easy_tdx/web/sentiment_store.py

@@ -29,6 +29,18 @@ def _config_dir() -> Path:
     return Path(os.environ.get("EASY_TDX_CONFIG_DIR", str(Path.home() / ".easy_tdx")))
 
 
+def _real_or_zero(v: Any) -> float:
+    """数值兜底：None/NaN/inf → 0.0（REAL NOT NULL 列不吃 NaN——SQLite 会把
+    NaN 绑定成 NULL 而触发约束冲突，整条写入失败）。"""
+    try:
+        f = float(v)
+    except (TypeError, ValueError):
+        return 0.0
+    if f != f or f in (float("inf"), float("-inf")):
+        return 0.0
+    return f
+
+
 class SentimentStore:
     """情绪采样 SQLite 存储。"""
 
@@ -116,7 +128,11 @@ class SentimentStore:
             conn.close()
 
     def upsert_fund_day(self, date: int, boards: list[dict[str, Any]]) -> None:
-        """覆盖写入某日行业主力净流入排行（rank 按列表顺序 1 起）。"""
+        """覆盖写入某日行业主力净流入排行（rank 按列表顺序 1 起）。
+
+        ``main_net`` 的 NaN（上游行情缺失时 pandas 的空值口径）会被 SQLite 存成
+        NULL 而触发 NOT NULL 约束冲突 → 整日采样失败；这里统一落成 0.0。
+        """
         with _write_lock:
             conn = self._connect()
             try:
@@ -124,7 +140,13 @@ class SentimentStore:
                 conn.executemany(
                     "INSERT INTO board_fund (date, rank, code, name, main_net) VALUES (?,?,?,?,?)",
                     [
-                        (int(date), i + 1, str(b["code"]), str(b["name"]), float(b["main_net"]))
+                        (
+                            int(date),
+                            i + 1,
+                            str(b["code"]),
+                            str(b["name"]),
+                            _real_or_zero(b["main_net"]),
+                        )
                         for i, b in enumerate(boards)
                     ],
                 )
@@ -155,7 +177,9 @@ class SentimentStore:
                         "rank": int(r["rank"]),
                         "code": str(r["code"]),
                         "name": str(r["name"]),
-                        "main_net": float(r["main_net"]),
+                        # 手工编辑/旧库可能存有 NULL（SQLite 把 NaN 存成 NULL），
+                        # float(None) 会 TypeError，兜成 0.0
+                        "main_net": float(r["main_net"] or 0.0),
                     }
                 )
             return list(grouped.values())

src/easy_tdx/web/task_runner.py

@@ -18,9 +18,9 @@
 - ``_run`` 写状态时**不假设** ``self._tasks[task_id]`` 仍在表中——并发淘汰
   可能在任务运行期间移除其条目。``move_to_end`` 用 try/except 容忍，状态
   写到本地 ``state`` 引用（即使被淘汰也无害，GC 回收）。
-- ``_evict_if_needed_locked`` 跳过 ``running`` 状态的任务——正在执行的任务
-  恰好是 OrderedDict 头部（完成时才 move_to_end），盲目 FIFO 淘汰会优先
-  杀掉在途任务。淘汰改用「最旧的 non-running 条目」。
+- ``_evict_if_needed_locked`` 只淘汰 ``done/failed`` 终态条目——pending 尚未
+  起跑（淘汰会产生 worker 跳过 + 查询侧恢复成永久 pending 的"幽灵任务"），
+  running 在途；无终态可淘汰时宁可持续超限。
 """
 
 from __future__ import annotations
@@ -156,11 +156,15 @@ class BacktestTaskRunner:
         with self._lock:
             memory_items = list(self._tasks.values())
         seen = {s.task_id for s in memory_items}
-        # 磁盘侧多取一些（覆盖内存 LRU 已淘汰的），再合并排序
+        # 磁盘侧多取一些（覆盖内存 LRU 已淘汰的），再合并排序；列表页不需要
+        # result，懒加载跳过 result_json 的 SELECT 与解析（几百条大结果时
+        # 可观省内存与事件循环停顿）
         try:
             disk_items = [
                 self._dict_to_state(d)
-                for d in get_task_store().list_recent(limit=limit + len(seen))
+                for d in get_task_store().list_recent(
+                    limit=limit + len(seen), include_results=False
+                )
                 if d["task_id"] not in seen
             ]
         except Exception:  # noqa: BLE001 — 持久化故障不阻断列表查询
@@ -284,20 +288,20 @@ class BacktestTaskRunner:
         )
 
     def _evict_if_needed_locked(self) -> None:
-        """超过上限时丢弃最旧的 non-running 任务（调用方需持锁）。
+        """超过上限时丢弃最旧的终态（done/failed）任务（调用方需持锁）。
 
-        running 任务不会被淘汰（它们恰在 OrderedDict 头部，但盲淘汰会杀在途任务）。
-        只淘汰 pending/done/failed 中最旧者。
+        pending/running 一律不淘汰：淘汰尚未起跑的 pending 会造成"幽灵任务"——
+        worker 线程随后取不到状态直接跳过，磁盘遗留的 pending 行会被查询
+        恢复成永远 pending。全为在途任务时宁可不淘汰（超限跳过）。
         """
         while len(self._tasks) > self._max_results:
-            # 找第一个 non-running 条目淘汰；若无则停止（全在 running，不强制淘汰）
             evict_id: str | None = None
             for tid, st in self._tasks.items():
-                if st.status != "running":
+                if st.status in ("done", "failed"):
                     evict_id = tid
                     break
             if evict_id is None:
-                break  # 全部 running，暂时无法淘汰
+                break  # 无终态条目可淘汰
             self._tasks.pop(evict_id, None)
 
 

src/easy_tdx/web/task_store.py

@@ -196,14 +196,26 @@ class TaskStore:
             conn.close()
         return self._row_to_dict(row) if row is not None else None
 
-    def list_recent(self, limit: int = 20) -> list[dict[str, Any]]:
-        """按 created_at 倒序列出最近 N 条任务摘要（含 result，供详情直取）。"""
+    def list_recent(self, limit: int = 20, *, include_results: bool = True) -> list[dict[str, Any]]:
+        """按 created_at 倒序列出最近 N 条任务摘要。
+
+        Args:
+            limit: 最多返回条数。
+            include_results: True（默认）= 含 result（详情直取）；False = 不
+                SELECT/解析 ``result_json``——列表页只展示摘要，用它避免把
+                几百条大结果 JSON 拖进内存逐条解析（result 字段为 None）。
+        """
         conn = self._connect()
         try:
+            cols = (
+                "*"
+                if include_results
+                else "task_id, status, description, created_at, started_at, "
+                "finished_at, error, NULL"
+            )
             cur = conn.execute(
-                """
-                SELECT task_id, status, description, created_at, started_at,
-                       finished_at, error, result_json
+                f"""
+                SELECT {cols}
                 FROM backtest_tasks
                 ORDER BY created_at DESC, task_id DESC
                 LIMIT ?
@@ -314,7 +326,7 @@ class _NullTaskStore(TaskStore):
     def load(self, task_id: str) -> dict[str, Any] | None:  # noqa: ARG002
         return None
 
-    def list_recent(self, limit: int = 20) -> list[dict[str, Any]]:  # noqa: ARG002
+    def list_recent(self, limit: int = 20, *, include_results: bool = True) -> list[dict[str, Any]]:  # noqa: ARG002
         return []
 
     def delete(self, task_id: str) -> bool:  # noqa: ARG002

tests/unit/test_ai_llm.py

@@ -160,6 +160,309 @@ class TestClient:
         assert result["ok"] is False and "API Key" in result["error"]
 
 
+class TestApiUrlSchemeGuard:
+    """api_url SSRF 防线：仅允许 http/https、禁止携带 userinfo。
+
+    背景：_post_json 用 urllib 直连用户可配的 api_url，无 scheme 白名单时
+    ``file:///...`` 可读本地文件（llm.json 内含明文 key，且格式异常分支会
+    回显响应前 300 字节）、``ftp://`` 与内网 http 可被当跳板。
+    """
+
+    def test_file_scheme_rejected(self):
+        with pytest.raises(ValueError, match="http"):
+            resolve_config(
+                LlmConfig(provider="custom", api_url="file:///C:/Users/x/llm.json", model="m")
+            )
+
+    def test_ftp_scheme_rejected(self):
+        with pytest.raises(ValueError, match="http"):
+            resolve_config(LlmConfig(provider="custom", api_url="ftp://internal-host/x", model="m"))
+
+    def test_missing_scheme_rejected(self):
+        with pytest.raises(ValueError, match="http"):
+            resolve_config(LlmConfig(provider="custom", api_url="api.deepseek.com/v1", model="m"))
+
+    def test_userinfo_rejected(self):
+        with pytest.raises(ValueError, match="user:pass"):
+            resolve_config(
+                LlmConfig(provider="custom", api_url="https://user:pass@api.x.com/v1", model="m")
+            )
+
+    def test_http_https_case_insensitive_allowed(self):
+        r = resolve_config(LlmConfig(provider="custom", api_url="HTTPS://Api.X.com/v1", model="m"))
+        assert r.api_url == "HTTPS://Api.X.com/v1"
+        r2 = resolve_config(
+            LlmConfig(provider="custom", api_url="http://gw.local:8000/v1", model="m")
+        )
+        assert r2.api_url == "http://gw.local:8000/v1"
+
+    def test_preset_urls_still_resolve(self, config_dir):
+        save_config(LlmConfig(provider="deepseek", api_key="sk-x-1234567890"))
+        r = resolve_config()
+        assert r.api_url == "https://api.deepseek.com/v1"
+
+
+class TestHttpPostHardening:
+    """HTTP 层加固：错误不回显原始 body、响应体大小上限。"""
+
+    def _raise_http_error(self, body: bytes, code: int = 401):
+        import io
+        import urllib.error
+
+        def fake_urlopen(req, timeout):
+            raise urllib.error.HTTPError(
+                req.full_url, code, "Unauthorized", hdrs=None, fp=io.BytesIO(body)
+            )
+
+        return fake_urlopen
+
+    def test_http_error_extracts_provider_message_only(self, monkeypatch):
+        """错误响应只回显 provider 的 error.message，不回显原始 body 其他内容。"""
+        import json as _json
+
+        body = _json.dumps(
+            {"error": {"message": "Invalid API key", "internal_hint": "SECRET-STACK"}}
+        ).encode()
+        monkeypatch.setattr(llm_mod.urllib.request, "urlopen", self._raise_http_error(body))
+        with pytest.raises(LlmError) as ei:
+            llm_mod._post_json("https://x/v1/chat/completions", {}, {"m": 1}, 5.0)
+        assert ei.value.status == 401
+        assert "Invalid API key" in str(ei.value)
+        assert "SECRET-STACK" not in str(ei.value)
+
+    def test_http_error_non_json_body_is_generic(self, monkeypatch):
+        """非 JSON 错误页不给原始内容，只给通用 HTTP 状态描述。"""
+        body = b"<html><h1>gateway exploded with internal detail</h1></html>"
+        monkeypatch.setattr(llm_mod.urllib.request, "urlopen", self._raise_http_error(body))
+        with pytest.raises(LlmError) as ei:
+            llm_mod._post_json("https://x/v1/chat/completions", {}, {"m": 1}, 5.0)
+        assert "gateway exploded" not in str(ei.value)
+        assert "401" in str(ei.value)
+
+    def test_http_error_string_error_field_still_shown(self, monkeypatch):
+        """error 为字符串的网关（如 {"error":"bad key"}）仍展示该消息。"""
+        monkeypatch.setattr(
+            llm_mod.urllib.request, "urlopen", self._raise_http_error(b'{"error":"bad key"}')
+        )
+        with pytest.raises(LlmError, match="bad key") as ei:
+            llm_mod._post_json("https://x/v1/chat/completions", {}, {"m": 1}, 5.0)
+        assert ei.value.status == 401
+
+    def test_response_body_size_capped(self, config_dir, monkeypatch):
+        """超过 2MB 的响应体中止解析（防异常网关撑爆内存），报可操作错误。"""
+
+        class _FakeResp:
+            def __init__(self, payload: bytes) -> None:
+                self._buf = payload
+
+            def read(self, n: int = -1) -> bytes:
+                if n < 0:
+                    data, self._buf = self._buf, b""
+                    return data
+                data, self._buf = self._buf[:n], self._buf[n:]
+                return data
+
+            def __enter__(self) -> _FakeResp:
+                return self
+
+            def __exit__(self, *exc: object) -> bool:
+                return False
+
+        big = b"x" * (llm_mod._MAX_RESPONSE_BYTES + 1)
+        monkeypatch.setattr(llm_mod.urllib.request, "urlopen", lambda req, timeout: _FakeResp(big))
+        with pytest.raises(LlmError, match="过大|上限"):
+            llm_mod._post_json("https://x/v1/chat/completions", {}, {"m": 1}, 5.0)
+
+    def test_normal_response_within_cap_parses(self, config_dir, monkeypatch):
+        class _FakeResp:
+            def __init__(self, payload: bytes) -> None:
+                self._buf = payload
+
+            def read(self, n: int = -1) -> bytes:
+                if n < 0:
+                    data, self._buf = self._buf, b""
+                    return data
+                data, self._buf = self._buf[:n], self._buf[n:]
+                return data
+
+            def __enter__(self) -> _FakeResp:
+                return self
+
+            def __exit__(self, *exc: object) -> bool:
+                return False
+
+        payload = b'{"choices": [{"message": {"content": "OK"}}]}'
+        monkeypatch.setattr(
+            llm_mod.urllib.request, "urlopen", lambda req, timeout: _FakeResp(payload)
+        )
+        data = llm_mod._post_json("https://x/v1/chat/completions", {}, {"m": 1}, 5.0)
+        assert data["choices"][0]["message"]["content"] == "OK"
+
+
+class TestSaveConfigAtomic:
+    def test_replace_failure_preserves_old_file(self, config_dir, monkeypatch):
+        """os.replace 失败（磁盘满等）时旧配置原样保留，不留临时文件。"""
+        save_config(LlmConfig(provider="deepseek", api_key="sk-old-1234567890"))
+
+        def boom(src, dst):
+            raise OSError("disk full")
+
+        monkeypatch.setattr(llm_mod.os, "replace", boom)
+        with pytest.raises(OSError):
+            save_config(LlmConfig(provider="kimi", api_key="sk-new-9999999999"))
+
+        assert load_config().api_key == "sk-old-1234567890"  # 旧配置未被破坏
+        leftovers = [p.name for p in config_dir.iterdir() if p.name != "llm.json"]
+        assert leftovers == []  # 失败的临时文件已清理
+
+
+class TestLoadConfigFieldDefense:
+    """手工编辑 llm.json 的脏字段不得打挂 load_config（全部 /llm/* 依赖它）。"""
+
+    def test_null_fields_fall_back_to_defaults(self, config_dir):
+        import json as _json
+
+        (config_dir / "llm.json").write_text(
+            _json.dumps(
+                {
+                    "provider": None,
+                    "api_url": None,
+                    "api_key": None,
+                    "model": None,
+                    "temperature": None,
+                    "max_tokens": None,
+                    "timeout": None,
+                    "system_prompt": None,
+                }
+            ),
+            encoding="utf-8",
+        )
+        cfg = load_config()  # 旧码：float(None) TypeError
+        assert cfg.provider == "deepseek"
+        assert cfg.api_url == "" and cfg.api_key == "" and cfg.model == ""
+        assert cfg.temperature == 0.3
+        assert cfg.max_tokens == 16000
+        assert cfg.timeout == 180.0
+        assert cfg.system_prompt == LlmConfig.system_prompt
+
+    def test_wrong_types_fall_back_with_warning(self, config_dir, caplog):
+        import json as _json
+
+        (config_dir / "llm.json").write_text(
+            _json.dumps(
+                {
+                    "temperature": "abc",
+                    "max_tokens": "fast",
+                    "timeout": [],
+                    "provider": 123,
+                    "system_prompt": 456,
+                }
+            ),
+            encoding="utf-8",
+        )
+        with caplog.at_level("WARNING", logger="easy_tdx.ai.llm"):
+            cfg = load_config()  # 旧码：float("abc") ValueError
+        assert cfg.temperature == 0.3
+        assert cfg.max_tokens == 16000
+        assert cfg.timeout == 180.0
+        assert cfg.provider == "deepseek"  # 非字符串 provider 回退默认
+        assert cfg.system_prompt == LlmConfig.system_prompt
+        assert any("temperature" in r.message for r in caplog.records)
+
+    def test_non_finite_and_out_of_range_fall_back(self, config_dir):
+        import json as _json
+
+        (config_dir / "llm.json").write_text(
+            _json.dumps({"temperature": 1e999, "timeout": -5, "max_tokens": 0}),  # 1e999→inf
+            encoding="utf-8",
+        )
+        cfg = load_config()  # 旧码：inf temperature 会一路写进请求 payload
+        assert cfg.temperature == 0.3
+        assert cfg.timeout == 180.0
+        assert cfg.max_tokens == 16000
+
+    def test_string_numbers_leniently_coerced(self, config_dir):
+        import json as _json
+
+        (config_dir / "llm.json").write_text(
+            _json.dumps({"temperature": "0.7", "max_tokens": "8192.9", "timeout": "60"}),
+            encoding="utf-8",
+        )
+        cfg = load_config()
+        assert cfg.temperature == 0.7
+        assert cfg.max_tokens == 8192
+        assert cfg.timeout == 60.0
+
+
+class TestAnthropicRobustness:
+    """anthropic 协议与 openai 口径对齐：绝不静默返回空正文。"""
+
+    def _client(self) -> LlmClient:
+        return LlmClient(LlmConfig(provider="claude", api_key="sk-ant-123456789"))
+
+    def test_thinking_only_blocks_raise_actionable(self, config_dir, monkeypatch):
+        """仅 thinking 块（max_tokens 被思考耗尽）→ 可操作错误，而非空串成功。"""
+
+        def fake_post(url, headers, payload, timeout):
+            return {
+                "content": [{"type": "thinking", "thinking": "思考" * 200}],
+                "stop_reason": "max_tokens",
+            }
+
+        monkeypatch.setattr(llm_mod, "_post_json", fake_post)
+        with pytest.raises(LlmError, match="思考链"):
+            asyncio.run(self._client().chat("hi"))
+
+    def test_content_as_plain_string_accepted(self, config_dir, monkeypatch):
+        """部分网关把 content 放字符串而非块列表——正常取正文。"""
+
+        def fake_post(url, headers, payload, timeout):
+            return {"content": "纯字符串回复"}
+
+        monkeypatch.setattr(llm_mod, "_post_json", fake_post)
+        assert asyncio.run(self._client().chat("hi")) == "纯字符串回复"
+
+    def test_mixed_blocks_text_extracted(self, config_dir, monkeypatch):
+        def fake_post(url, headers, payload, timeout):
+            return {
+                "content": [
+                    {"type": "thinking", "thinking": "思考"},
+                    {"type": "text", "text": "正文"},
+                ],
+                "stop_reason": "end_turn",
+            }
+
+        monkeypatch.setattr(llm_mod, "_post_json", fake_post)
+        assert asyncio.run(self._client().chat("hi")) == "正文"
+
+    def test_missing_content_raises_llm_error(self, config_dir, monkeypatch):
+        """content 缺失 → LlmError（旧码 AttributeError 裸 500）。"""
+
+        def fake_post(url, headers, payload, timeout):
+            return {"stop_reason": "end_turn"}
+
+        monkeypatch.setattr(llm_mod, "_post_json", fake_post)
+        with pytest.raises(LlmError, match="格式异常"):
+            asyncio.run(self._client().chat("hi"))
+
+    def test_empty_blocks_generic_error_without_raw_echo(self, config_dir, monkeypatch):
+        def fake_post(url, headers, payload, timeout):
+            return {"content": [{"type": "tool_use", "id": "tool_1", "secret": "S3CR3T"}]}
+
+        monkeypatch.setattr(llm_mod, "_post_json", fake_post)
+        with pytest.raises(LlmError, match="content 为空") as ei:
+            asyncio.run(self._client().chat("hi"))
+        assert "S3CR3T" not in str(ei.value)  # 不回显原始响应体
+
+    def test_empty_string_content_with_max_tokens_stop(self, config_dir, monkeypatch):
+        def fake_post(url, headers, payload, timeout):
+            return {"content": "", "stop_reason": "max_tokens"}
+
+        monkeypatch.setattr(llm_mod, "_post_json", fake_post)
+        with pytest.raises(LlmError, match="截断"):
+            asyncio.run(self._client().chat("hi"))
+
+
 def test_provider_presets_cover_major_vendors():
     vendors = [
         "deepseek",

tests/unit/test_backtest_cli.py

@@ -141,3 +141,67 @@ class TestPortfolioCLIFlags:
         assert "--evaluate" in result.output
         assert "--wf" in result.output
         assert "--auto-fees" in result.output
+
+
+class TestIgnoredFlagWarnings:
+    """静默忽略的旗标组合必须显式告警（stderr），不得无提示吞掉。"""
+
+    def test_combo_with_wf_warns(self):
+        """--combo-strategies + --wf：旧码静默忽略，新码应告警并继续报策略文件错误。"""
+        from easy_tdx.backtest.cli import backtest
+
+        runner = CliRunner()
+        result = runner.invoke(
+            backtest,
+            [
+                "SZ",
+                "000001",
+                "--combo-strategies",
+                "nope_a.py,nope_b.py",
+                "--wf",
+            ],
+        )
+        assert "已忽略" in result.output
+        assert result.exit_code != 0  # 随后仍因策略文件不存在报错（无网络依赖）
+
+    def test_combo_with_evaluate_warns(self):
+        from easy_tdx.backtest.cli import backtest
+
+        runner = CliRunner()
+        result = runner.invoke(
+            backtest,
+            [
+                "SZ",
+                "000001",
+                "--combo-strategies",
+                "nope_a.py,nope_b.py",
+                "--evaluate",
+            ],
+        )
+        assert "已忽略" in result.output
+
+    def test_single_with_wf_no_warning(self):
+        """单策略 + --wf 是合法组合，不告警（在策略校验失败前无「已忽略」字样）。"""
+        from easy_tdx.backtest.cli import backtest
+
+        runner = CliRunner()
+        result = runner.invoke(backtest, ["SZ", "000001", "--wf"])
+        assert "已忽略" not in result.output
+
+    def test_optimize_all_with_param_warns(self):
+        """--all + --param：旧码静默忽略自定义网格，新码应告警。"""
+        from easy_tdx.backtest.cli import optimize
+
+        runner = CliRunner()
+        # 用非法市场名在联网取数前中断，仅验证告警已发出
+        result = runner.invoke(optimize, ["XX", "000001", "--all", "--param", "fast=5,10"])
+        assert "忽略" in result.output
+
+    def test_optimize_single_with_param_no_warning(self):
+        from easy_tdx.backtest.cli import optimize
+
+        runner = CliRunner()
+        result = runner.invoke(
+            optimize, ["XX", "000001", "--strategy", "ma_cross", "--param", "fast=5,10"]
+        )
+        assert "已忽略" not in result.output

tests/unit/test_backtest_fitness_benchmark.py

@@ -197,6 +197,31 @@ def test_evaluate_strategy_auto_fees_for_etf():
     assert report["config"]["auto_fees"] is True
 
 
+def test_evaluate_portfolio_auto_fees_fitness_per_stock_symbol():
+    """组合体检逐段应按各标的品种解析费率（auto_fees 与主回测同口径）。
+
+    回归：evaluate_portfolio 的 per-stock FitnessEngine 漏传 symbol，
+    auto_fees 不生效——ETF 组合的三段体检被按股票口径错收印花税（卖方 0.001），
+    与组合回测主路径（PortfolioBacktestEngine 逐标的 resolve_fee_model）不一致。
+    期望：报告里的三段体检与「正确传入 symbol 的独立体检」逐段一致。
+    """
+    from easy_tdx.backtest.benchmark import evaluate_portfolio
+    from easy_tdx.backtest.portfolio_engine import StockData
+
+    df = _df(240, drift=0.002)
+    stocks = [StockData("159915", "SZ", df)]  # ETF：法定免印花税
+    report = evaluate_portfolio(
+        _CycleTrader, stocks, total_cash=100_000, auto_fees=True, n_windows=3
+    )
+
+    expected = FitnessEngine(symbol="SZ159915", strategy=_CycleTrader, auto_fees=True).evaluate(df)
+    actual_returns = [seg["total_return"] for seg in report["fitness"]["segments"]]
+    expected_returns = [seg.total_return for seg in expected.segments]
+    assert len(actual_returns) == 3
+    for actual, exp in zip(actual_returns, expected_returns):
+        assert actual == pytest.approx(exp, abs=1e-9)
+
+
 # ── evaluate_portfolio（v1.31 组合级一条龙）───────────────────────────────────
 def _stocks_for_portfolio() -> list[Any]:
     from easy_tdx.backtest.portfolio_engine import StockData

tests/unit/test_backtest_rotation.py

@@ -157,3 +157,93 @@ def test_momentum_score_helper():
     score = momentum_score(10)(df)
     assert score > 0
     assert momentum_score(10)(_stock(5)) == 0.0  # 数据不足 → 0
+
+
+# ── 回归：停牌/初始调仓/历史不足（审查修复） ─────────────────────────────────
+
+
+def _bar_frame(dates: pd.DatetimeIndex, prices: list[float]) -> pd.DataFrame:
+    closes = np.asarray(prices, dtype=float)
+    return pd.DataFrame(
+        {
+            "datetime": dates[: len(closes)],
+            "open": closes * 0.999,
+            "high": closes * 1.01,
+            "low": closes * 0.98,
+            "close": closes,
+            "vol": 1e6,
+        }
+    )
+
+
+def test_rotation_suspension_defers_fill_to_resume_open():
+    """停牌日挂单顺延：成交日=复牌日、成交价=复牌开盘（旧码在停牌日按停牌前价格成交）。"""
+    dates = pd.date_range("2024-01-01", periods=12, freq="D")
+    a = _bar_frame(dates, [10 + 0.05 * i for i in range(12)])
+    # B：01-01..01-07 有 bar（01-07 收盘崩盘跌出排名），01-08 停牌（下标 7 无 bar），
+    # 01-09 复牌开盘 -30%（下标 8 = 4.2）
+    b_prices = [10, 10.1, 10.2, 10.3, 10.4, 10.5, 6.0, 4.9, 4.2, 4.3, 4.3, 4.3]
+    b = pd.DataFrame(
+        [
+            {
+                "datetime": dates[i],
+                "open": p * 0.999,
+                "high": p * 1.01,
+                "low": p * 0.98,
+                "close": p,
+                "vol": 1e6,
+            }
+            for i, p in enumerate(b_prices)
+            if i != 7  # 01-08 停牌，无 bar
+        ]
+    )
+
+    engine = RotationEngine(
+        {"SH:600001": a, "SZ:000002": b},
+        momentum_score(2),
+        slots=1,
+        refresh="daily",
+        keep_rank=1,
+    )
+    res = engine.run()
+
+    sells_b = [t for t in res.trades if t["symbol"] == "SZ:000002" and t["direction"] == "SELL"]
+    assert len(sells_b) == 1
+    sell = sells_b[0]
+    assert sell["datetime"] == "2024-01-09"  # 旧码记 2024-01-08（停牌日）
+    assert sell["price"] == pytest.approx(4.2 * 0.999)  # 旧码记 6.0 * 0.999（停牌前开盘）
+    # 复牌前净值按最后已知收盘估值，不应把持仓价值清零
+    eq_by_date = {r["datetime"]: r for r in res.equity_curve}
+    assert eq_by_date["2024-01-08"]["position_value"] > 0
+
+
+def test_rotation_day0_counts_as_first_rebalance():
+    """day0 即为首个调仓日（排名只用 ≤day0 数据，次日开盘执行），不再人为空仓一天。"""
+    pool = _pool({f"SH:60000{i}": 0.002 for i in range(5)}, n=40)
+    res = RotationEngine(pool, momentum_score(5), slots=3, refresh="weekly").run()
+    # 旧码首个调仓日是下一 ISO 周 2024-01-08
+    assert res.rebalance_dates[0] == "2024-01-01"
+    # next_open 语义：任何成交不早于第二个交易日（day0 信号次日执行）
+    if res.trades:
+        assert min(t["datetime"] for t in res.trades) > res.rebalance_dates[0]
+
+
+def test_rotation_new_listing_not_bought_on_zero_score():
+    """历史不足（<5 根）从买入候选剔除：次新股 0 分不得排在负动量标的之前被买入。"""
+    n = 60
+    dates = pd.date_range("2024-01-01", periods=n, freq="B")
+    declining = 100.0 * np.cumprod(np.full(n, 1.0 - 0.005))
+    a = _bar_frame(dates, list(declining))  # 长历史持续阴跌，动量为负
+    b = _bar_frame(dates, [10.0, 10.0, 10.0])  # 末段才上市，全程 idx<5
+
+    res = RotationEngine(
+        {"SH:600001": a, "SZ:000002": b},
+        momentum_score(5),
+        slots=1,
+        refresh="daily",
+    ).run()
+
+    buys_b = [t for t in res.trades if t["symbol"] == "SZ:000002" and t["direction"] == "BUY"]
+    assert buys_b == []  # 旧码 B 以 0 分登顶被买入
+    # A 作为唯一有效候选被正常买入
+    assert any(t["symbol"] == "SH:600001" and t["direction"] == "BUY" for t in res.trades)

tests/unit/test_backtest_walkforward.py

@@ -174,3 +174,100 @@ def test_wf_auto_fes_passed_through():
     assert wf_engine._engine_kwargs["auto_fees"] is True
     wf = wf_engine.run(_trend_df(300))
     assert len(wf.windows) == 3
+
+
+# ── 回归：窗口绩效口径 / 聚合方向 / 切窗下限 / 失败日志 / int 日期 ────────────
+
+
+def test_wf_window_metrics_exclude_context_bars():
+    """上下文只做指标预热：窗口绩效指标不随 context_bars 变化。
+
+    旧码把 context 恒定现金段一并喂给 PerformanceAnalyzer，sharpe/年化/波动
+    被稀释（同窗 total_return 相同而 sharpe 相差近一倍）。
+    """
+    df = _trend_df(500)
+    wf0 = WalkForwardEngine(_BuyFirstBar, n_windows=5, warmup_ratio=0.3, context_bars=0).run(df)
+    wf60 = WalkForwardEngine(_BuyFirstBar, n_windows=5, warmup_ratio=0.3, context_bars=60).run(df)
+    assert len(wf0.windows) == len(wf60.windows) == 5
+    for w0, w60 in zip(wf0.windows, wf60.windows):
+        assert w0.total_return == pytest.approx(w60.total_return)
+        assert w0.sharpe == pytest.approx(w60.sharpe)
+        assert w0.max_drawdown == pytest.approx(w60.max_drawdown)
+        assert w0.performance["annual_return"] == pytest.approx(w60.performance["annual_return"])
+        assert w0.performance["volatility"] == pytest.approx(w60.performance["volatility"])
+
+
+def test_wf_worst_drawdown_is_max_not_min():
+    """worst_drawdown 应取各窗最深回撤（max）；旧码 min 取成最浅回撤。"""
+    from easy_tdx.backtest.walkforward import WalkForwardResult, WalkForwardWindow
+
+    result = WalkForwardResult(n_windows=3, warmup_ratio=0.3)
+    for i, dd in enumerate((0.05, 0.40, 0.11)):
+        result.windows.append(
+            WalkForwardWindow(
+                index=i,
+                start="2024-01-01",
+                end="2024-02-01",
+                bars=20,
+                total_return=0.01,
+                sharpe=1.0,
+                max_drawdown=dd,
+                total_trades=2,
+                win_rate=0.5,
+            )
+        )
+    WalkForwardEngine._aggregate(result)
+    assert result.worst_drawdown == pytest.approx(0.40)
+
+
+def test_wf_windows_below_min_bars_skipped():
+    """单窗实际 bar 数 < 20 时跳过（与 docstring「每窗 ≥ 20 根」口径一致）。"""
+    wf = WalkForwardEngine(_BuyFirstBar, n_windows=9).run(_trend_df(220))
+    assert wf.windows == []
+
+
+class _BoomStrategy(Strategy):
+    """init 即抛错：单窗失败应记 warning 而非静默跳过。"""
+
+    def init(self) -> None:
+        raise RuntimeError("boom")
+
+    def next(self) -> None:
+        pass
+
+
+def test_wf_window_failure_logs_warning(caplog):
+    """单窗回测异常记 warning（含窗号与异常摘要），不拖垮整组。"""
+    import logging
+
+    with caplog.at_level(logging.WARNING, logger="easy_tdx.backtest.walkforward"):
+        wf = WalkForwardEngine(_BoomStrategy, n_windows=3).run(_trend_df(300))
+    assert wf.windows == []
+    msgs = [r.getMessage() for r in caplog.records if r.levelno >= logging.WARNING]
+    assert any("第 0 窗" in m and "boom" in m for m in msgs), msgs
+
+
+def test_wf_int_yyyymmdd_date_column_window_labels():
+    """datetime 为 int YYYYMMDD（TDX 日线原样）时窗口起止日期正确。
+
+    旧码 pd.Timestamp(int) 按纳秒换算，窗口日期全变 1970-01-01。
+    """
+    n = 300
+    dates = pd.date_range("2023-01-02", periods=n, freq="B")
+    close = 10.0 * np.linspace(1.0, 2.0, n)
+    df = pd.DataFrame(
+        {
+            "datetime": dates.strftime("%Y%m%d").astype(int),
+            "open": close * 0.999,
+            "high": close * 1.01,
+            "low": close * 0.99,
+            "close": close,
+            "vol": 1000.0,
+        }
+    )
+    wf = WalkForwardEngine(_BuyFirstBar, n_windows=3, context_bars=10).run(df)
+    assert len(wf.windows) == 3
+    eval_start = int(n * 0.3)
+    assert wf.windows[0].start == dates[eval_start].strftime("%Y-%m-%d")
+    assert wf.windows[0].end == dates[eval_start + (n - eval_start) // 3 - 1].strftime("%Y-%m-%d")
+    assert not wf.windows[0].start.startswith("1970")

tests/unit/test_baostock_source.py

@@ -46,16 +46,28 @@ def _fake_rows(n: int, end: str = "2026-09-04") -> list[list[str]]:
     return [[d, "10.0", "11.0", "9.5", "10.5", "100000", "1050000.0", "1"] for d in dates]
 
 
+def _weekly_rows(n: int, end: str = "2026-09-04") -> list[list[str]]:
+    """n 个周线行（无 tradestatus 列，与真实 W/M 返回一致）。"""
+    dates = pd.date_range(end=end, periods=n, freq="W-FRI").strftime("%Y-%m-%d")
+    return [[d, "10.0", "11.0", "9.5", "10.5", "500000", "5250000.0"] for d in dates]
+
+
 def _install_fake_bs(
     rows: list[list[str]] | None,
     captured: dict,
     *,
     query_error: bool = False,
+    login_error: bool = False,
 ) -> types.ModuleType:
     mod = types.ModuleType("baostock")
 
     def _login():  # type: ignore[no-untyped-def]
         captured["login"] = captured.get("login", 0) + 1
+        if login_error:
+            result = _FakeLoginResult()
+            result.error_code = "10001"
+            result.error_msg = "用户登录失败"
+            return result
         return _FakeLoginResult()
 
     mod.login = _login  # type: ignore[attr-defined]
@@ -166,12 +178,83 @@ def test_unsupported_inputs(fake_bs):
     assert "calls" not in fake_bs
 
 
-def test_query_error_returns_none(fake_bs):
-    """baostock 查询失败：返回 None 且不向上抛（兜底失败不改变原错误路径）。"""
+def test_query_error_raises_runtimeerror_and_logs(fake_bs, caplog):
+    """baostock 查询失败（error_code≠0）：记 warning 并抛 RuntimeError。
+
+    回归：旧实现吞掉所有异常静默返回 None——`--source baostock` 显式使用时
+    故障被伪装成"无数据"（sync 记 skipped 而非 failed）。auto 兜底路径
+    （web/routers/bars.py）以 except Exception 包裹调用，不受影响。
+    """
     _install_fake_bs([], fake_bs, query_error=True)
     from easy_tdx.sources import baostock as bs_source
 
-    assert bs_source.fetch_bars("SH", "600519", "DAY", 0, 5, "QFQ") is None
+    with pytest.raises(RuntimeError, match="baostock 拉取失败"):
+        bs_source.fetch_bars("SH", "600519", "DAY", 0, 5, "QFQ")
+    assert "baostock 拉取失败" in caplog.text
+
+
+def test_login_failure_raises_runtimeerror(fake_bs, monkeypatch: pytest.MonkeyPatch):
+    """baostock 登录失败：同样 warning + RuntimeError（不再静默）。"""
+    _install_fake_bs([], fake_bs, login_error=True)
+    from easy_tdx.sources import baostock as bs_source
+
+    with pytest.raises(RuntimeError, match="拉取失败"):
+        bs_source.fetch_bars("SZ", "000001", "DAY", 0, 5, "QFQ")
+
+
+def test_weekly_monthly_fields_exclude_tradestatus(fake_bs):
+    """W/M 请求不传 tradestatus（baostock 实测 error_code=10004012 报错），
+    日线保留。"""
+    from easy_tdx.sources import baostock as bs_source
+
+    weekly_rows = _weekly_rows(6)
+    _install_fake_bs(weekly_rows, fake_bs)
+    df = bs_source.fetch_bars("SH", "600519", "WEEK", 0, 5, "QFQ")
+    assert df is not None and len(df) == 5
+    assert fake_bs["frequency"] == "w"
+    assert "tradestatus" not in fake_bs["fields"]
+    assert list(df.columns) == ["date", "open", "close", "high", "low", "vol", "amount"]
+
+    _install_fake_bs(_weekly_rows(6), fake_bs)
+    df = bs_source.fetch_bars("SH", "600519", "MONTH", 0, 5, "QFQ")
+    assert df is not None and len(df) == 5
+    assert fake_bs["frequency"] == "m"
+    assert "tradestatus" not in fake_bs["fields"]
+
+    # 日线仍保留 tradestatus（停牌剔除依赖它）
+    _install_fake_bs(_fake_rows(6), fake_bs)
+    bs_source.fetch_bars("SH", "600519", "DAY", 0, 5, "QFQ")
+    assert "tradestatus" in fake_bs["fields"]
+
+
+def test_weekly_suspension_dropped_by_volume(fake_bs):
+    """WEEK 无 tradestatus 列时，停牌/无成交周（volume=0）按 vol>0 兜底剔除。"""
+    from easy_tdx.sources import baostock as bs_source
+
+    rows = _weekly_rows(6)
+    rows[2][5] = "0"  # volume=0 的停牌周
+    _install_fake_bs(rows, fake_bs)
+    df = bs_source.fetch_bars("SZ", "000001", "WEEK", 0, 10, "QFQ")
+    assert df is not None and len(df) == 5
+    assert (df["vol"] > 0).all()
+
+
+def test_index_volume_converted_to_lots(fake_bs):
+    """is_index=True：指数 vol 股→手（÷100），对齐 /bars/index 契约。
+
+    实测 sh.000001 2026-09-04：baostock volume=53,728,616,100（股），
+    ÷100 = 537,286,161 手（TDX 指数日线口径为手）。
+    """
+    from easy_tdx.sources import baostock as bs_source
+
+    df = bs_source.fetch_bars("SH", "000001", "DAY", 0, 5, "NONE", is_index=True)
+    assert df is not None
+    assert (df["vol"] == 1000.0).all()  # 100000 股 ÷100 = 1000 手
+
+    # 默认（个股路径）不换算
+    _install_fake_bs(_fake_rows(6), fake_bs)
+    df = bs_source.fetch_bars("SH", "600519", "DAY", 0, 5, "QFQ")
+    assert (df["vol"] == 100000).all()
 
 
 # ---------------------------------------------------------------------------

tests/unit/test_board_mac_hotspot.py

@@ -379,3 +379,39 @@ def test_hotspot_correlation_building_passthrough():
     assert body["status"] in ("building", "error", "ready")  # 单机假客户端极快时可能已完成
     if body["status"] == "building":
         assert 0.0 <= body["progress"] <= 1.0
+
+
+# ── 时区统一（v1.32.6）：日历日一律取沪市时区，与主机时区无关 ────────────────
+
+
+# 模块导入时捕获真实实现（autouse fixture 会把 _today_str 换成钉死的 lambda）
+_board_mac_mod = pytest.importorskip("easy_tdx.web.routers.board_mac")
+_REAL_TODAY_STR = _board_mac_mod._today_str
+
+
+def test_today_str_uses_shanghai_tz(monkeypatch):
+    """_today_str 必须用 SHANGHAI_TZ 取"今天"（旧实现用主机本地时区）。
+
+    海外机器（如 UTC-5）上北京时间 09-06 02:00 时本地还是 09-05，
+    旧实现会把热点矩阵的"今日"判定错一天。
+    """
+    from datetime import datetime
+
+    pytest.importorskip("fastapi")
+    from easy_tdx.realtime.session import SHANGHAI_TZ
+    from easy_tdx.web.routers import board_mac
+
+    # 恢复被 autouse fixture 钉住的真实现
+    monkeypatch.setattr(board_mac, "_today_str", _REAL_TODAY_STR)
+
+    captured: dict = {}
+
+    class _FakeDatetime:
+        @classmethod
+        def now(cls, tz=None):
+            captured["tz"] = tz
+            return datetime(2026, 9, 6, 2, 0, tzinfo=tz) if tz else datetime(2026, 9, 6, 2, 0)
+
+    monkeypatch.setattr(board_mac, "datetime", _FakeDatetime)
+    assert board_mac._today_str() == "2026-09-06"
+    assert captured["tz"] is SHANGHAI_TZ

tests/unit/test_board_mac_overview.py

@@ -219,3 +219,75 @@ def test_overview_zero_pre_close_change_pct_null():
     row = resp.json()["data"]["rows"][0]
     assert row["change_pct"] is None
     assert row["leader_change_pct"] is None
+
+
+def test_overview_cache_key_includes_count():
+    """缓存键须含 count：不同 count 的请求在 TTL 内不互相命中。
+
+    旧实现缓存键只有 (board_type, metrics)，先到的小 count 请求会把大 count
+    的响应"污染"成少数行（15s TTL 内）。
+    """
+    pytest.importorskip("fastapi")
+    from fastapi.testclient import TestClient
+
+    frames = {
+        "CHANGE_PCT": _board_df(
+            [
+                _board_row("881001", "软件服务", 5000.0, 4900.0),
+                _board_row("881002", "半导体", 3000.0, 2950.0),
+            ]
+        ),
+    }
+
+    class _CountingFake(_FakeOverviewMacClient):
+        """尊重 count 参数（与真实客户端一致地截断行数）。"""
+
+        async def get_board_list(self, board_type=None, count=10000, sort_column=None):
+            df = await super().get_board_list(
+                board_type=board_type, count=count, sort_column=sort_column
+            )
+            return df.head(count) if df is not None else df
+
+    fake = _CountingFake(frames)
+    with TestClient(_overview_app(fake)) as client:
+        r_small = client.get("/api/v1/board-mac/overview", params={"board_type": "HY", "count": 1})
+        assert r_small.status_code == 200
+        assert r_small.json()["data"]["count"] == 1
+
+        r_big = client.get("/api/v1/board-mac/overview", params={"board_type": "HY", "count": 2})
+        assert r_big.status_code == 200
+        # 不允许命中 count=1 的缓存
+        assert r_big.json()["data"]["count"] == 2
+    assert fake.calls.count("CHANGE_PCT") == 2  # 两个 count 各拉一次
+
+
+def test_overview_nan_payload_cleaned_before_cache():
+    """坏值（NaN）行不产生 500，且写入缓存前已清洗（缓存里不留 NaN）。
+
+    旧实现：sort_value=NaN → payload 带 NaN → Starlette allow_nan=False
+    序列化 500，且带毒 payload 先入 15s 缓存，TTL 内持续 500。
+    """
+    pytest.importorskip("fastapi")
+    from fastapi.testclient import TestClient
+
+    from easy_tdx.web.routers import board_mac
+
+    nan = float("nan")
+    frames = {
+        "CHANGE_PCT": _board_df([_board_row("881001", "软件服务", 5000.0, 4900.0)]),
+        "SPEED": _board_df([_board_row("881001", "软件服务", 5000.0, 4900.0, sort_value=nan)]),
+    }
+    fake = _FakeOverviewMacClient(frames)
+    with TestClient(_overview_app(fake)) as client:
+        r1 = _get_overview(client)
+        assert r1.status_code == 200
+        assert r1.json()["data"]["rows"][0]["speed"] is None
+
+        # 坏 payload 不得入缓存：缓存里的 speed 应已是 None
+        cached = board_mac._overview_cache[("HY", ("SPEED", "CHANGE_20D"), 2000)][1]
+        assert cached["rows"][0]["speed"] is None
+
+        r2 = _get_overview(client)  # 命中缓存也不再 500
+        assert r2.status_code == 200
+        assert r2.json()["data"]["rows"][0]["speed"] is None
+    assert fake.calls.count("SPEED") == 1

tests/unit/test_ccpm.py

@@ -177,6 +177,38 @@ def test_parse_xml_error_page_raises() -> None:
         parse_xml("404 page，非 XML 内容")
 
 
+def test_parse_xml_structure_change_raises() -> None:
+    """XML 合法但 0 个 <data> 节点且存在其他子结构：抛 CcpmError（官网改版信号）。
+
+    回归：旧实现静默返回空表——改版后 CLI/Web 层展示"无数据"而非报错，
+    改版长期无人察觉。正常发布日必有 <data>；真实无数据日走 302。
+    """
+    from easy_tdx.ccpm import CcpmError, parse_xml
+
+    changed = (
+        '<?xml version="1.0" encoding="UTF-8"?>'
+        "<positionRank><record><instrumentid>IF2609</instrumentid></record></positionRank>"
+    )
+    with pytest.raises(CcpmError, match="结构可能已变更"):
+        parse_xml(changed)
+
+
+def test_parse_xml_empty_root_is_no_data() -> None:
+    """空 <positionRank/>（无任何子节点）：按当日无数据处理，返回空表。"""
+    from easy_tdx.ccpm import parse_xml
+
+    assert parse_xml('<?xml version="1.0" encoding="UTF-8"?><positionRank></positionRank>') == []
+
+
+def test_get_rank_structure_change_raises(isolated_config, monkeypatch) -> None:
+    """结构变更经 get_rank 透传为 CcpmError（不缓存、不返回空表伪装成功）。"""
+    from easy_tdx.ccpm import CcpmClient, CcpmError
+
+    _mock_fetch(monkeypatch, sample="<positionRank><unknown>1</unknown></positionRank>")
+    with pytest.raises(CcpmError, match="结构可能已变更"):
+        CcpmClient().get_rank("IF", "2026-09-02")
+
+
 # ---------------------------------------------------------------------------
 # 日期归一化
 # ---------------------------------------------------------------------------

tests/unit/test_cli_symbol_and_sync.py

@@ -0,0 +1,251 @@
+"""CLI 参数校验与退出码测试（cmd_warehouse / cmd_formula，#审查修复）。
+
+覆盖：
+- ``市场:代码`` 解析辅助：缺冒号/空段 → click.BadParameter（而非裸 ValueError）；
+- ``warehouse sync``：failed>0 时 exit 1（对齐 ccpm 口径）、summary 带 source 标注、
+  ``--period`` Choice 限定、``--source baostock`` 不支持分钟周期时参数层报错；
+- ``warehouse check``：先做「只支持一个标的」校验再解析，单标的缺冒号也报错；
+- ``formula screen``：缺冒号标的前置报错（不再裸 traceback）。
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+import pytest
+from click.testing import CliRunner
+
+from easy_tdx.cli.cmd_formula import _parse_symbol as _parse_symbol_formula
+from easy_tdx.cli.cmd_warehouse import (
+    _BAOSTOCK_PERIODS,
+    _PERIOD_CHOICES,
+    warehouse_check,
+    warehouse_sync,
+)
+from easy_tdx.cli.cmd_warehouse import (
+    _parse_symbol as _parse_symbol_warehouse,
+)
+
+
+class TestParseSymbol:
+    @pytest.mark.parametrize("parse", [_parse_symbol_formula, _parse_symbol_warehouse])
+    def test_valid(self, parse):
+        assert parse("SH:600519") == ("SH", "600519")
+        assert parse(" sz:000001 ") == ("SZ", "000001")
+
+    @pytest.mark.parametrize("parse", [_parse_symbol_formula, _parse_symbol_warehouse])
+    @pytest.mark.parametrize(
+        "bad", ["SH600519", "SH:", ":600519", ":", "SH 600519", "SH:600519:extra"]
+    )
+    def test_malformed_raises_bad_parameter(self, parse, bad):
+        # 旧码：sym.split(":", 1) 裸 ValueError（"SH:600519:extra" 旧码能过但语义错，也收紧）
+        from click import BadParameter
+
+        with pytest.raises(BadParameter, match="市场:代码"):
+            parse(bad)
+
+
+class _FakeWarehouse:
+    """context-manager 形假的 KlineWarehouse（cmd 只当透传对象用）。"""
+
+    def __init__(self, *a: Any, **k: Any) -> None:
+        pass
+
+    def __enter__(self) -> _FakeWarehouse:
+        return self
+
+    def __exit__(self, *exc: object) -> bool:
+        return False
+
+    def health_check(self, market: str | None = None, code: str | None = None) -> dict[str, Any]:
+        return {"issues": [], "market": market, "code": code}
+
+
+class _FakeSyncer:
+    """可编程结果假 WarehouseSyncer（cmd_warehouse 从 easy_tdx.warehouse 导入它）。"""
+
+    result: dict[str, Any] = {}
+
+    def __init__(self, *a: Any, **k: Any) -> None:
+        pass
+
+    def sync(self, symbols: Any, period: str, progress: Any = None) -> dict[str, Any]:
+        if progress is not None:
+            progress(1, len(symbols), str(symbols[0]))
+        return dict(self.result)
+
+
+@pytest.fixture()
+def patched_warehouse(monkeypatch):
+    """打桩 cmd_warehouse 的全部外部依赖（仓库 / TDX 客户端 / 同步器）。"""
+    import easy_tdx.cli.cmd_warehouse as cw
+    import easy_tdx.cli.conn as conn_mod
+    import easy_tdx.warehouse as wh_pkg
+
+    class _FakeMacClient:
+        def __enter__(self) -> _FakeMacClient:
+            return self
+
+        def __exit__(self, *exc: object) -> bool:
+            return False
+
+    monkeypatch.setattr(cw, "_require_warehouse", lambda db_path: _FakeWarehouse())
+    monkeypatch.setattr(conn_mod, "get_mac_client", lambda: _FakeMacClient())
+    monkeypatch.setattr(wh_pkg, "WarehouseSyncer", _FakeSyncer)
+    return cw
+
+
+class TestWarehouseSync:
+    def _invoke(self, *args: str):
+        return CliRunner().invoke(warehouse_sync, list(args), catch_exceptions=False)
+
+    def test_failed_symbols_exit_1(self, patched_warehouse):
+        """有标的失败 → exit 1（旧码：failed 只进 summary，命令仍 exit 0）。"""
+        _FakeSyncer.result = {
+            "total": 2,
+            "ok": 1,
+            "added": 3,
+            "updated": 0,
+            "skipped": 0,
+            "failed": 1,
+            "details": [
+                {"symbol": "SH:600519", "added": 3, "updated": 0, "skipped": 0, "error": None},
+                {"symbol": "SZ:000001", "added": 0, "updated": 0, "skipped": 0, "error": "boom"},
+            ],
+        }
+        result = self._invoke("--symbols", "SH:600519,SZ:000001", "--source", "tdx")
+        assert result.exit_code == 1, result.output
+
+    def test_all_ok_exit_0_and_source_in_summary(self, patched_warehouse):
+        """全部成功 → exit 0，summary JSON 带 source 标注（与 /bars 响应呼应）。"""
+        _FakeSyncer.result = {
+            "total": 1,
+            "ok": 1,
+            "added": 3,
+            "updated": 0,
+            "skipped": 0,
+            "failed": 0,
+            "details": [
+                {"symbol": "SH:600519", "added": 3, "updated": 0, "skipped": 0, "error": None},
+            ],
+        }
+        result = self._invoke("--symbols", "SH:600519", "--source", "tdx")
+        assert result.exit_code == 0, result.output
+        payload = json.loads(result.stdout)  # stdout 仅 summary JSON（进度/错误在 stderr）
+        assert payload["source"] == "tdx"
+        assert payload["ok"] == 1
+
+    def test_malformed_symbol_no_bare_traceback(self, patched_warehouse):
+        """缺冒号标的 → 友好 BadParameter（exit 2），不发网络请求不裸崩。"""
+        result = CliRunner().invoke(warehouse_sync, ["--symbols", "SH600519", "--source", "tdx"])
+        assert result.exit_code == 2
+        assert "市场:代码" in result.output
+
+    def test_period_choice_rejects_unknown(self, patched_warehouse):
+        result = CliRunner().invoke(warehouse_sync, ["--symbols", "SH:600519", "--period", "WEEKN"])
+        assert result.exit_code == 2
+
+    def test_baostock_rejects_intraday_period(self, patched_warehouse):
+        """--source baostock + 分钟周期 → 参数层直接报错（旧码：静默空转 exit 0）。"""
+        for period in _PERIOD_CHOICES:
+            if period not in _BAOSTOCK_PERIODS:
+                result = CliRunner().invoke(
+                    warehouse_sync,
+                    ["--symbols", "SH:600519", "--source", "baostock", "--period", period],
+                )
+                assert result.exit_code == 2, (period, result.output)
+                assert "baostock" in result.output
+
+    def test_baostock_accepts_daily(self, patched_warehouse, monkeypatch):
+        """--source baostock + DAILY 正常放行（不走 TDX 客户端）。"""
+        _FakeSyncer.result = {
+            "total": 1,
+            "ok": 1,
+            "added": 0,
+            "updated": 0,
+            "skipped": 0,
+            "failed": 0,
+            "details": [],
+        }
+
+        # baostock 路径不经过 get_mac_client——若被调用说明走错分支
+        import easy_tdx.cli.conn as conn_mod
+
+        def _no_tdx():
+            raise AssertionError("baostock source 不应触碰 TDX 客户端")
+
+        monkeypatch.setattr(conn_mod, "get_mac_client", _no_tdx)
+        result = self._invoke("--symbols", "SH:600519", "--source", "baostock")
+        assert result.exit_code == 0, result.output
+        assert json.loads(result.stdout)["source"] == "baostock"
+
+
+class TestWarehouseCheck:
+    def test_multiple_symbols_rejected_before_parse(self, monkeypatch):
+        """多标的先报「只支持一个」，不再先 split 崩溃（旧码顺序颠倒）。"""
+        import easy_tdx.cli.cmd_warehouse as cw
+
+        monkeypatch.setattr(cw, "_require_warehouse", lambda db_path: _FakeWarehouse())
+        result = CliRunner().invoke(
+            warehouse_check,
+            ["--symbols", "SH600519,SZ:000001"],  # 旧码：含冒号绕过校验 → split 裸崩
+            catch_exceptions=False,
+        )
+        assert result.exit_code == 1
+        assert "只支持一个标的" in result.output
+
+    def test_single_symbol_missing_colon_rejected(self, monkeypatch):
+        import easy_tdx.cli.cmd_warehouse as cw
+
+        monkeypatch.setattr(cw, "_require_warehouse", lambda db_path: _FakeWarehouse())
+        result = CliRunner().invoke(
+            warehouse_check,
+            ["--symbols", "SH600519"],  # 旧码：split(":", 1) 裸 ValueError
+            catch_exceptions=False,
+        )
+        assert result.exit_code == 2
+        assert "市场:代码" in result.output
+
+    def test_single_valid_symbol_passes_market_code(self, monkeypatch):
+        import easy_tdx.cli.cmd_warehouse as cw
+
+        monkeypatch.setattr(cw, "_require_warehouse", lambda db_path: _FakeWarehouse())
+        result = CliRunner().invoke(
+            warehouse_check,
+            ["--symbols", "SH:600519"],
+            catch_exceptions=False,
+        )
+        assert result.exit_code == 0, result.output
+        payload = json.loads(result.output)
+        assert payload["market"] == "SH" and payload["code"] == "600519"
+
+    def test_issues_still_exit_0_by_design(self, monkeypatch):
+        """自检发现 issues → 正常输出并 exit 0（自检结果本身是正常输出，保持原口径）。"""
+        import easy_tdx.cli.cmd_warehouse as cw
+
+        class _Wh(_FakeWarehouse):
+            def health_check(self, market=None, code=None):
+                return {"issues": ["gap"]}
+
+        monkeypatch.setattr(cw, "_require_warehouse", lambda db_path: _Wh())
+        result = CliRunner().invoke(
+            warehouse_check, ["--symbols", "SH:600519"], catch_exceptions=False
+        )
+        assert result.exit_code == 0
+        assert "gap" in result.output
+
+
+class TestFormulaScreenSymbolValidation:
+    def test_malformed_symbol_fails_fast(self):
+        """缺冒号标的前置报错（旧码：循环里裸 ValueError traceback）。"""
+        from easy_tdx.cli.cmd_formula import formula_screen
+
+        result = CliRunner().invoke(
+            formula_screen,
+            ["--symbols", "SH:600519,SH600036", "--formula", "金叉: CROSS(MA(C,5), MA(C,20));"],
+            catch_exceptions=False,
+        )
+        assert result.exit_code == 2
+        assert "市场:代码" in result.output
+        assert "SH600036" in result.output  # 报错指出坏标的

tests/unit/test_formula.py

@@ -212,3 +212,81 @@ def test_compiled_formula_is_dataclass_safe():
     assert np.allclose(
         f.compute(_df(20)).columns["值"], f2.compute(_df(20)).columns["值"], equal_nan=True
     )
+
+
+# ── 回归：归类收严 / FILTER 副作用 / 递归上限 / REF 负移位（审查修复） ────────
+
+
+def test_ratio_output_classified_as_value_not_signal():
+    """0~1 区间的价格比率是数值列，不是信号列（旧码按 [0,1] 值域兜底误判）。"""
+    res = compile_formula("比率: C / HHV(C, 20);").compute(_df(40))
+    assert res.signals == []
+    assert res.values == ["比率"]
+
+
+def test_normalized_oscillator_classified_as_value():
+    """归一化振荡器（RSI/100）是数值列，不是信号列。"""
+    res = compile_formula("强度: RSI(C, 14) / 100;").compute(_df(40))
+    assert res.signals == []
+    assert res.values == ["强度"]
+
+
+def test_binary_zero_one_values_still_signal():
+    """真正的 0/1 两值输出仍兜底归信号列。"""
+    res = compile_formula("X: IF(C > MA(C, 5), 1, 0);").compute(_df(40))
+    assert res.signals == ["X"]
+
+
+def test_filter_does_not_pollute_series():
+    """FILTER(C, N) 不改写输入序列：同公式后续 MA(C, 2) 与未过滤一致。"""
+    df = _df(30)
+    res = compile_formula("A: FILTER(C, 2); B: MA(C, 2);").compute(df)
+    close = pd.to_numeric(df["close"]).to_numpy(dtype=float)
+    expected = pd.Series(close).rolling(2).mean().to_numpy()
+    np.testing.assert_allclose(res.columns["B"], expected, equal_nan=True)
+
+
+def test_deep_paren_nesting_formula_error():
+    """超深括号嵌套抛 FormulaError（嵌套过深），而非 RecursionError 逃逸。"""
+    text = "X: " + "(" * 5000 + "C" + ")" * 5000 + ";"
+    with pytest.raises(FormulaError, match="嵌套过深"):
+        compile_formula(text)
+
+
+def test_deep_unary_chain_formula_error():
+    """超长一元运算符链同样受深度上限保护。"""
+    with pytest.raises(FormulaError, match="嵌套过深"):
+        compile_formula("X: " + "!" * 5000 + "C;")
+
+
+def test_moderate_nesting_still_compiles():
+    """常规嵌套深度不受上限影响。"""
+    res = compile_formula("X: -(-(-(C + 1) * 2) + 3);").compute(_df(10))
+    assert res.columns["X"].shape == (10,)
+
+
+def test_ref_negative_shift_banned():
+    """REF 负移位（未来函数）显式 FormulaError，不再依赖 float 类型巧合。"""
+    with pytest.raises(FormulaError, match="负移位"):
+        compile_formula("X: REF(C, -1);").compute(_df(30))
+
+
+def test_ref_negative_via_expression_banned():
+    """负移位经表达式算出（如 0-1）同样被禁。"""
+    with pytest.raises(FormulaError, match="负移位"):
+        compile_formula("X: REF(C, 0 - 1);").compute(_df(30))
+
+
+def test_ref_positive_still_works():
+    res = compile_formula("X: REF(C, 1);").compute(_df(30))
+    assert np.isnan(res.columns["X"][0])
+    assert res.columns["X"][1] == pytest.approx(float(pd.to_numeric(_df(30)["close"]).iloc[0]))
+
+
+def test_mytt_internal_negative_ref_unaffected():
+    """MyTT 库内直调（ICHIMOKU 迟行带）不经公式白名单，负移位仍可用。"""
+    from easy_tdx.MyTT import REF
+
+    close = np.arange(5, dtype=float)
+    out = REF(close, -1)
+    assert out[0] == pytest.approx(1.0)

tests/unit/test_formula_integration.py

@@ -191,3 +191,13 @@ def test_rest_formula_screen_async_task():
     # symbols 路径需要行情连接——离线环境预期 400/500（无 mock client）
     # 这里只验证请求校验（symbols 非空）不炸
     assert r.status_code in (400, 500, 202)
+
+
+def test_pick_signal_columns_ignores_ratio_value_column():
+    """0~1 值域的比率列归类为数值输出后，不再被自动挑成买卖信号列。"""
+    _, result = attach_formula_columns(
+        _df(60), compile_formula("比率: C / HHV(C, 20);\n强弱: C > MA(C, 5);")
+    )
+    buy, sell = pick_signal_columns(result)
+    assert buy == "强弱"  # 旧码 signals 含「比率」且排在首位，被误选为买入列
+    assert sell is None

tests/unit/test_grading_scoring.py

@@ -205,6 +205,44 @@ def test_combined_metrics_dd_duration_unclosed_counts_to_end():
     assert m.max_dd_duration == 20
 
 
+def test_combined_metrics_missing_drawdown_pct_row_continues_state():
+    """缺行 drawdown_pct 视为状态延续，不得当作创新高截断水下期。
+
+    回归：compute_combined_metrics 的 drawdown_pct 分支曾把缺行/None 取 0
+    （=创新高），把真实水下段从中间截断，max_dd_duration 被低估（3 → 2）。
+    """
+    eq = [
+        {"total": 100.0, "datetime": "2024-01-01", "drawdown_pct": 0.0},
+        {"total": 50.0, "datetime": "2024-01-02", "drawdown_pct": 0.5},
+        {"total": 50.0, "datetime": "2024-01-03"},  # 缺 drawdown_pct → 沿用水下
+        {"total": 50.0, "datetime": "2024-01-04", "drawdown_pct": 0.5},
+    ]
+    m = compute_combined_metrics(eq)
+    # idx0 之后一直未创新高（缺行延续 idx1 的水下状态）→ 计到末点 = 3
+    assert m.max_dd_duration == 3
+
+
+def test_combined_metrics_nan_drawdown_pct_continues_state():
+    """NaN drawdown_pct 沿用上一根状态：峰值后的 NaN 仍按峰值处理。
+
+    回归：NaN 行既不算峰值也不截断，last_peak 停在上一根真峰值，导致
+    水下期被多算一根（2 → 1）。
+    """
+    eq = [
+        {"total": 100.0, "datetime": "2024-01-01", "drawdown_pct": 0.0},
+        {"total": 110.0, "datetime": "2024-01-02", "drawdown_pct": 0.0},
+        {"total": 110.0, "datetime": "2024-01-03", "drawdown_pct": float("nan")},
+        {
+            "total": 109.0,
+            "datetime": "2024-01-04",
+            "drawdown_pct": (110.0 - 109.0) / 110.0,
+        },
+    ]
+    m = compute_combined_metrics(eq)
+    # NaN 行沿用 idx1 的峰值状态 → 最后一次创新高为 idx2 → 水下 1 根
+    assert m.max_dd_duration == 1
+
+
 def test_combined_metrics_insufficient_points():
     m = compute_combined_metrics([{"total": 100.0}])
     assert m.n_points == 1

tests/unit/test_limitup_ecology.py

@@ -128,6 +128,67 @@ def test_limitup_empty_vipdoc(tmp_path):
     assert eco.summary()["limit_up_count"] == 0
 
 
+# ── 涨跌停价舍入（回归：浮点 floor(x*100+0.5) 在半分边界错 1 分）──────────────
+
+
+def test_limit_price_matches_exchange_rounding_all_range():
+    """_limit_price 与交易所 ROUND_HALF_UP 对 1.00~600.00 全价位零差异。
+
+    旧实现（float 乘后 floor）在 ±10% 档 67/318 个价位、±5% 档 90/884 个
+    价位算低 1 分（如 prev=1.15：涨停价应 1.27，旧算 1.26）。
+    """
+    from decimal import ROUND_HALF_UP, Decimal
+
+    from easy_tdx.screen.limitup import _limit_price
+
+    for pct in (10, 5, 20, -10, -5, -20):
+        for cents in range(100, 60001):
+            prev = Decimal(cents).scaleb(-2)
+            expected = (
+                int(
+                    (Decimal(cents) * (100 + pct) / 100).quantize(
+                        Decimal("1"), rounding=ROUND_HALF_UP
+                    )
+                )
+                / 100
+            )
+            got = _limit_price(float(prev), pct)
+            assert got == expected, (prev, pct, got, expected)
+
+
+def test_exchange_boundary_prices_detected(tmp_path):
+    """半分边界价位的真实涨跌停不因浮点舍入漏判。
+
+    选点依据：.day 读回（raw×0.01）的浮点误差会抵消部分边界，33.05→36.36
+    与 2.65→2.39 是经读回仿真验证后旧实现（floor 浮点版）仍漏判的价位。
+    """
+    from easy_tdx.screen.limitup import compute_limitup_ecology
+
+    # prev=33.05 → 交易所涨停价 36.36（旧实现误算 36.35 → 漏判涨停）
+    _write_stock(tmp_path, "sh", "600901", [33.05, 36.36])
+    # prev=2.65 → 交易所跌停价 2.39（旧实现误算 2.38 → 漏判跌停）
+    _write_stock(tmp_path, "sz", "000902", [2.65, 2.39])
+
+    eco = compute_limitup_ecology(tmp_path)
+    up = {e.code: e for e in eco.limit_up}
+    down = {e.code: e for e in eco.limit_down}
+    assert "600901" in up, f"33.05→36.36 应判涨停，实际 limit_up={up}"
+    assert up["600901"].streak == 1
+    assert "000902" in down, f"2.65→2.39 应判跌停，实际 limit_down={down}"
+
+
+def test_history_boundary_prices_counted(tmp_path):
+    """历史回补同样按交易所口径计涨跌停（33.05→36.36 / 2.65→2.39）。"""
+    from easy_tdx.screen.limitup import compute_limitup_history
+
+    _write_stock(tmp_path, "sh", "600901", [33.05, 36.36])
+    _write_stock(tmp_path, "sz", "000902", [2.65, 2.39])
+
+    rows = {r["date"]: r for r in compute_limitup_history(tmp_path, days=5)}
+    assert rows[20260802]["limit_up"] == 1
+    assert rows[20260802]["limit_down"] == 1
+
+
 def test_limitup_endpoint_and_cache(vipdoc, monkeypatch):
     """端点返回 DictResponse 包装；60s 内命中缓存（扫描只跑一次）。"""
     pytest.importorskip("fastapi")
@@ -164,3 +225,42 @@ def test_limitup_endpoint_and_cache(vipdoc, monkeypatch):
         assert r2.json()["data"] == d1
 
     assert calls["n"] == 1  # 第二次命中缓存
+
+
+def test_limitup_endpoint_cache_key_includes_vipdoc(vipdoc, tmp_path, monkeypatch):
+    """缓存键须含 vipdoc：不同 vipdoc 的请求在 TTL 内不互相命中。
+
+    旧实现 _limitup_cache 是单值缓存，先到的 vipdoc=A 结果会被 vipdoc=B
+    的请求在 60s TTL 内复用。
+    """
+    pytest.importorskip("fastapi")
+    from fastapi import FastAPI
+    from fastapi.testclient import TestClient
+
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import market as market_mod
+
+    other = tmp_path / "vipdoc_other"
+    (other / "sh" / "lday").mkdir(parents=True)
+    (other / "sh" / "lday" / "sh600100.day").write_bytes(
+        _day(20260801, 9.95, 11.0, 9.9, 11.0) + _day(20260802, 11.0, 12.1, 10.9, 12.1)
+    )
+
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(market_mod.router, prefix="/api/v1")
+    app.state.tdx_client = object()
+
+    with TestClient(app) as client:
+        r1 = client.get("/api/v1/limitup-ecology", params={"vipdoc": str(vipdoc)})
+        assert r1.status_code == 200
+        r2 = client.get("/api/v1/limitup-ecology", params={"vipdoc": str(other)})
+        assert r2.status_code == 200
+
+    # 同 vipdoc 的第二次请求才命中缓存；不同 vipdoc 必须各自扫描
+    with TestClient(app) as client:
+        client.get("/api/v1/limitup-ecology", params={"vipdoc": str(vipdoc)})
+        client.get("/api/v1/limitup-ecology", params={"vipdoc": str(vipdoc)})
+        d = client.get("/api/v1/limitup-ecology", params={"vipdoc": str(other)}).json()["data"]
+        # other 目录只有 600100 一只 2 连板，不含 vipdoc 目录的 3 连板数据
+        assert d["summary"]["limit_up_count"] == 1

tests/unit/test_multi_strategy.py

@@ -13,6 +13,7 @@ from __future__ import annotations
 
 import numpy as np
 import pandas as pd
+import pytest
 
 from easy_tdx.backtest.multi_strategy_engine import (
     MultiStrategyEngine,
@@ -206,6 +207,31 @@ class TestMultiStrategyEngine:
         # 合并曲线长度应至少覆盖两个范围的最晚结束日（并集）
         assert len(result.combined_equity) >= 60
 
+    def test_total_return_capital_weighted_with_disjoint_dates(self) -> None:
+        """晚起步槽位建仓前应按初始资金趴账（合并曲线首值=总投入资金）。
+
+        回归：旧实现对日期并集的前导缺口填 0——B 槽位起步前贡献 0 而非其
+        分得的 50 万，合并曲线首值 = 50 万 < 总资金 100 万，total_return 被虚增。
+        正确口径：前导缺口用每列首个有效值（=初始资金）回填（bfill）。
+        """
+        df_a = _make_df(60, seed=1, start="2024-01-01")
+        df_b = _make_df(60, seed=2, start="2024-03-01")
+        slots = [
+            StrategySlot("A", "SH:601088", SimpleBuyStrategy(), df_a),
+            StrategySlot("B", "SZ:000001", SimpleBuyStrategy(), df_b),
+        ]
+        result = MultiStrategyEngine(slots, total_cash=1_000_000).run()
+
+        # 合并曲线首值 = 总投入资金（旧实现 = 500000，缺晚起步槽位的资金）
+        assert result.combined_equity["total"].iloc[0] == pytest.approx(1_000_000.0)
+
+        # total_return == 各槽位资金加权真实收益
+        weighted = sum(
+            0.5 * res.performance.get("total_return", 0.0)
+            for res in result.individual_results.values()
+        )
+        assert result.total_performance["total_return"] == pytest.approx(weighted, abs=1e-9)
+
     def test_empty_strategies_returns_empty_result(self) -> None:
         """空策略列表应返回空结果，不抛异常。"""
         engine = MultiStrategyEngine([], total_cash=1_000_000)

tests/unit/test_mytt.py

@@ -557,3 +557,44 @@ class TestBBPandBBW:
         quiet = 100 + rng.standard_normal(60) * 0.1
         wild = 100 + rng.standard_normal(60) * 5.0
         assert MyTT.BBW(wild)[-1] > MyTT.BBW(quiet)[-1]
+
+
+class TestFilter:
+    """FILTER 无副作用（审查修复：曾原地改写输入序列）。"""
+
+    def test_filter_does_not_mutate_input(self):
+        x = np.array([1.0, 0.0, 1.0, 1.0, 1.0, 0.0, 1.0])
+        snapshot = x.copy()
+        out = MyTT.FILTER(x, 2)
+        np.testing.assert_array_equal(x, snapshot)  # 旧码把 x 原地置零，失败
+        # x[0]=1 触发 → 后 2 根置零；x[3]=1 触发 → 后 2 根置零；x[6]=1
+        np.testing.assert_array_equal(out, [1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0])
+
+    def test_filter_accepts_bool_input(self):
+        s = np.array([True, False, True, True])
+        out = MyTT.FILTER(s, 1)
+        np.testing.assert_array_equal(s, [True, False, True, True])
+        assert out.dtype == bool
+        # s[0]=True 触发 → s[1] 置零；s[2]=True 触发 → s[3] 置零（s[2] 本身保留）
+        np.testing.assert_array_equal(out, [True, False, True, False])
+
+    def test_filter_formula_series_not_polluted(self):
+        """公式场景：FILTER(C, N) 后 C 仍是原收盘序列。"""
+        from easy_tdx.formula import compile_formula
+
+        n = 30
+        close = 10.0 * np.linspace(1.0, 2.0, n)
+        df = pd.DataFrame(
+            {
+                "datetime": pd.date_range("2024-01-01", periods=n),
+                "open": close,
+                "high": close,
+                "low": close,
+                "close": close,
+                "vol": np.ones(n),
+            }
+        )
+        res = compile_formula("A: FILTER(C, 2); B: MA(C, 2);").compute(df)
+        np.testing.assert_allclose(
+            res.columns["B"], pd.Series(close).rolling(2).mean().to_numpy(), equal_nan=True
+        )

tests/unit/test_portfolio_engine.py

@@ -290,6 +290,51 @@ class TestPortfolioFullMetrics:
         )
         assert result.total_performance["total_return"] == pytest.approx(weighted, abs=1e-9)
 
+    def test_total_return_capital_weighted_with_uneven_start_dates(self) -> None:
+        """晚上市标的建仓前应按初始资金趴账（合并曲线首值=总投入资金）。
+
+        回归：旧实现 ``_build_combined_equity`` 对日期并集的前导缺口填 0——
+        晚上市标的上市前贡献 0 而非其分得的初始资金，合并曲线首值 < 总投入，
+        total_return 被系统性虚增（100 根 + 晚 60 根起步的组合实测虚增约 10 倍）。
+        正确口径（与组合 Walk-Forward 的 ffill().bfill() 一致）：前导缺口用
+        每列首个有效值回填——资金在组合起点即已分配，建仓前趴账。
+        """
+        n = 100
+        close = np.linspace(10.0, 12.0, n)
+        dates = pd.bdate_range("2024-01-02", periods=n)
+
+        def _mk(cnt: int) -> pd.DataFrame:
+            c = close[-cnt:]
+            return pd.DataFrame(
+                {
+                    "datetime": dates[-cnt:],
+                    "open": c,
+                    "high": c,
+                    "low": c,
+                    "close": c,
+                    "vol": 1e6,
+                    "amount": c * 1e6,
+                }
+            )
+
+        stocks = [
+            StockData("000001", "SZ", _mk(n)),  # 全程 100 根
+            StockData("600000", "SH", _mk(40)),  # 同涨势、晚 60 根起步
+        ]
+        result = PortfolioBacktestEngine(
+            strategy=SimpleBuyStrategy, stocks=stocks, total_cash=200000
+        ).run()
+
+        # 1) 合并曲线首值 = 总投入资金（旧实现 = 100000，缺晚上市标的的资金）
+        assert result.combined_equity["total"].iloc[0] == pytest.approx(200000.0)
+
+        # 2) total_return == 各标的资金加权真实收益
+        weighted = sum(
+            0.5 * res.performance.get("total_return", 0.0)
+            for res in result.individual_results.values()
+        )
+        assert result.total_performance["total_return"] == pytest.approx(weighted, abs=1e-9)
+
     def test_to_dict_contains_trades(self) -> None:
         """to_dict 应包含组合层成交表（REST/AI 解读消费）。"""
         stocks = [StockData("000001", "SZ", _make_df(100, seed=42))]

tests/unit/test_portfolio_walkforward.py

@@ -182,3 +182,30 @@ def test_multi_strategy_wf_empty_slots() -> None:
 
     wf = MultiStrategyWalkForwardEngine(strategies=[], n_windows=3).run()
     assert wf.windows == []
+
+
+def test_combo_slot_failure_logs_warning(caplog):
+    """单槽位回测异常记 warning（含槽位标识），不拖垮整窗（旧码静默 continue）。"""
+    import logging
+
+    from easy_tdx.backtest.multi_strategy_engine import StrategySlot
+    from easy_tdx.backtest.walkforward import MultiStrategyWalkForwardEngine
+
+    class Boom(Strategy):
+        def init(self) -> None:
+            raise RuntimeError("slot-boom")
+
+        def next(self) -> None:
+            pass
+
+    slots = [
+        StrategySlot(
+            label="正常", symbol="SH:601088", strategy=PeriodicStrategy(), df=_make_df(400, seed=42)
+        ),
+        StrategySlot(label="炸裂", symbol="SZ:000001", strategy=Boom(), df=_make_df(400, seed=99)),
+    ]
+    with caplog.at_level(logging.WARNING, logger="easy_tdx.backtest.walkforward"):
+        wf = MultiStrategyWalkForwardEngine(strategies=slots, n_windows=3).run()
+    assert len(wf.windows) == 3  # 正常槽位照常出窗
+    assert wf.total_trades > 0
+    assert any("炸裂" in r.getMessage() for r in caplog.records if r.levelno >= logging.WARNING)

tests/unit/test_qfq_crosscheck.py

@@ -116,6 +116,13 @@ def test_detect_gap_uses_chinext_threshold() -> None:
     assert detect_ex_dividend_gaps(df2, "300750") == ["2010-01-06"]
 
 
+def test_detect_gap_ignores_nonfinite_ratio() -> None:
+    """前收缺失（NaN）导致的非有限比率不计为除权跳空（首根前收缺失常见）。"""
+    df = _kline([float("nan")] + [10.0] * 5, opens=[5.0] + [10.0] * 5)
+    # 旧实现把 ratio=NaN 也当跳空 → 误报 ["2010-01-02"]
+    assert detect_ex_dividend_gaps(df, "600000") == []
+
+
 # --------------------------------------------------------------------------- #
 # 已知案例回归（合成）
 # --------------------------------------------------------------------------- #

tests/unit/test_realtime_feed.py

@@ -332,3 +332,23 @@ class TestStopFlag:
 
         await asyncio.wait_for(feed._run_sync_loop(client, None), timeout=2.0)
         assert client.calls == []
+
+    async def test_restart_after_stop_runs_again(self) -> None:
+        """start→stop→start：停止请求一次性消费，实例可再次启动。
+
+        回归：旧实现 _stop_requested 只置位不复位——stop 后再次 run_async
+        会静默立即返回（假启动），同一实例永久失效。
+        """
+        bus = EventBus()
+        client = AsyncMockClient([_sample_quotes_df()])
+        feed = RealtimeDataFeed(bus=bus, symbols=[(0, "000001")], sessions=(), interval=0.1)
+
+        await feed.run_async(client, max_iterations=1)
+        assert len(client.calls) == 1
+
+        feed.stop()
+        await feed.run_async(client, max_iterations=1)  # 消费停止请求：启动即退出
+        assert len(client.calls) == 1
+
+        await feed.run_async(client, max_iterations=1)  # 再次启动应正常轮询
+        assert len(client.calls) == 2

tests/unit/test_sentiment.py

@@ -7,6 +7,7 @@ sentiment_store 用 EASY_TDX_CONFIG_DIR 指向临时目录；limitup 历史复
 from __future__ import annotations
 
 import asyncio
+import pathlib
 
 import pytest
 
@@ -110,7 +111,11 @@ def test_sampler_inserts_store_rows(store):
 
 @pytest.fixture
 def vipdoc_factory(tmp_path):
-    """按 {文件名: {dates, closes}} 合成 vipdoc 目录的工厂。"""
+    """按 {文件名: {dates, closes}} 合成 vipdoc 目录的工厂。
+
+    ``factory(specs, root=None)``：root 缺省写 tmp_path；同一测试需要多个
+    独立 vipdoc 目录时传不同 root。
+    """
     from easy_tdx.offline.daily_bar import _DAILY_FMT
 
     def _day(date: int, close: float) -> bytes:
@@ -125,14 +130,15 @@ def vipdoc_factory(tmp_path):
             0,
         )
 
-    def factory(specs: dict[str, dict]) -> object:
+    def factory(specs: dict[str, dict], root=None) -> object:
+        base = pathlib.Path(root) if root is not None else tmp_path
         for filename, spec in specs.items():
             exchange = filename[:2]
-            lday = tmp_path / exchange / "lday"
+            lday = base / exchange / "lday"
             lday.mkdir(parents=True, exist_ok=True)
             data = b"".join(_day(d, c) for d, c in zip(spec["dates"], spec["closes"]))
             (lday / f"{filename}.day").write_bytes(data)
-        return tmp_path
+        return base
 
     return factory
 
@@ -198,3 +204,173 @@ def test_limitup_history_endpoint_cache(vipdoc_factory, monkeypatch):
         assert body["days"][0] == {"date": 20260802, "limit_up": 1, "limit_down": 0}
         client.get("/api/v1/market/limitup-history", params={"days": 10, "vipdoc": str(v)})
     assert calls["n"] == 1  # 缓存命中
+
+
+def test_board_fund_history_null_main_net_no_type_error(tmp_path):
+    """历史遗留的 main_net NULL 行不得让 /market/board-fund/history 抛 TypeError。
+
+    正式 schema 的 main_net 是 REAL NOT NULL（NaN 会被 SQLite 存成 NULL 而
+    被 NOT NULL 拒绝），但手工编辑/旧版本库可能存在 NULL 行——读侧须兜底。
+    旧实现：float(None) 抛 TypeError。
+    """
+    import sqlite3
+
+    db = tmp_path / "legacy_sentiment.db"
+    conn = sqlite3.connect(db)
+    conn.executescript(
+        """
+        CREATE TABLE samples (date INTEGER NOT NULL, minute INTEGER NOT NULL, ts INTEGER NOT NULL,
+            up_count INTEGER NOT NULL, down_count INTEGER NOT NULL, neutral_count INTEGER NOT NULL,
+            total_count INTEGER NOT NULL, limit_up_count INTEGER NOT NULL,
+            limit_down_count INTEGER NOT NULL, total_amount REAL NOT NULL,
+            PRIMARY KEY (date, minute));
+        CREATE TABLE board_fund (date INTEGER NOT NULL, rank INTEGER NOT NULL,
+            code TEXT NOT NULL, name TEXT NOT NULL, main_net REAL,
+            PRIMARY KEY (date, rank));
+        """
+    )
+    conn.execute(
+        "INSERT INTO board_fund (date, rank, code, name, main_net) VALUES (?,?,?,?,?)",
+        (20260901, 1, "881001", "银行", None),
+    )
+    conn.commit()
+    conn.close()
+
+    from easy_tdx.web.sentiment_store import SentimentStore
+
+    store = SentimentStore(db_path=db)
+    assert store.list_fund_days(5) == [
+        {
+            "date": 20260901,
+            "boards": [{"rank": 1, "code": "881001", "name": "银行", "main_net": 0.0}],
+        }
+    ]
+
+
+def test_upsert_fund_day_nan_main_net_stored_as_zero(store):
+    """写入口径：NaN 主力净流入落库为 0.0（REAL NOT NULL 列不吃 NaN）。"""
+    store.upsert_fund_day(
+        20260902,
+        [{"code": "881001", "name": "银行", "main_net": float("nan")}],
+    )
+    days = store.list_fund_days(5)
+    assert days[0]["boards"][0]["main_net"] == 0.0
+
+
+def test_limitup_history_cache_key_includes_vipdoc(vipdoc_factory, monkeypatch, tmp_path):
+    """limitup-history 缓存键须含 (days, vipdoc)，不同 vipdoc 不互相命中。"""
+    pytest.importorskip("fastapi")
+    from fastapi import FastAPI
+    from fastapi.testclient import TestClient
+
+    from easy_tdx.screen import limitup as limitup_mod
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import market as market_mod
+
+    v1 = vipdoc_factory({"sh600100": {"dates": [20260801, 20260802], "closes": [10.0, 11.0]}})
+    v2_dir = tmp_path / "vipdoc_v2"
+    (v2_dir / "sh" / "lday").mkdir(parents=True)
+    v2 = vipdoc_factory(
+        {
+            "sh600100": {"dates": [20260801, 20260802], "closes": [10.0, 11.0]},
+            "sz000200": {"dates": [20260801, 20260802], "closes": [10.0, 9.0]},
+        },
+        root=v2_dir,
+    )
+    calls = {"n": 0}
+    real = limitup_mod.compute_limitup_history
+
+    def counting(*a, **kw):
+        calls["n"] += 1
+        return real(*a, **kw)
+
+    monkeypatch.setattr(limitup_mod, "compute_limitup_history", counting)
+
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(market_mod.router, prefix="/api/v1")
+    app.state.tdx_client = object()
+
+    with TestClient(app) as client:
+        r1 = client.get("/api/v1/market/limitup-history", params={"days": 10, "vipdoc": str(v1)})
+        assert r1.status_code == 200
+        r2 = client.get("/api/v1/market/limitup-history", params={"days": 10, "vipdoc": str(v2)})
+        assert r2.status_code == 200
+        # v2 多一只跌停股 → 结果必须不同（不允许命中 v1 的缓存）
+        assert r2.json()["data"]["days"][0]["limit_down"] == 1
+    assert calls["n"] == 2
+
+
+# ── 采样器时区统一（v1.32.6）：日期/分钟键一律取沪市时区 ─────────────────────
+
+
+class _RecorderDatetime:
+    """记录 now(tz) 实参的 datetime 替身。"""
+
+    captured: dict = {}
+
+    @classmethod
+    def now(cls, tz=None):
+        cls.captured["tz"] = tz
+        from datetime import datetime as _dt
+
+        return _dt(2026, 9, 6, 2, 30, tzinfo=tz) if tz else _dt(2026, 9, 6, 2, 30)
+
+
+def test_sentiment_sampler_uses_shanghai_tz(store, monkeypatch):
+    """SentimentSampler._sample_once 的 (date, minute) 键必须取沪市时区。"""
+    import asyncio
+
+    import pandas as pd
+
+    from easy_tdx.realtime.session import SHANGHAI_TZ
+    from easy_tdx.web import sentiment_sampler as ss_mod
+
+    async def fake_stat():
+        return pd.DataFrame(
+            [
+                {
+                    "up_count": 2000,
+                    "down_count": 2000,
+                    "neutral_count": 100,
+                    "total_count": 4100,
+                    "limit_up_count": 50,
+                    "limit_down_count": 10,
+                    "total_amount": 8e11,
+                }
+            ]
+        )
+
+    monkeypatch.setattr(ss_mod, "datetime", _RecorderDatetime)
+    sampler = ss_mod.SentimentSampler(fake_stat, store=store)
+    asyncio.run(sampler._sample_once())
+
+    assert _RecorderDatetime.captured["tz"] is SHANGHAI_TZ
+    rows = store.day_samples(20260906)
+    assert len(rows) == 1 and rows[0]["minute"] == 230
+
+
+def test_fund_flow_sampler_uses_shanghai_tz(store, monkeypatch):
+    """FundFlowSampler._sample_once 的采样日期必须取沪市时区。"""
+    import asyncio
+
+    import pandas as pd
+
+    from easy_tdx.realtime.session import SHANGHAI_TZ
+    from easy_tdx.web import sentiment_sampler as ss_mod
+
+    class _FakeMac:
+        async def get_board_ranking(self, **kw):
+            return pd.DataFrame(
+                [
+                    {"code": "881001", "name": "银行", "main_net_amount": 1.2e9},
+                ]
+            )
+
+    monkeypatch.setattr(ss_mod, "datetime", _RecorderDatetime)
+    sampler = ss_mod.FundFlowSampler(_FakeMac(), store=store)
+    asyncio.run(sampler._sample_once())
+
+    assert _RecorderDatetime.captured["tz"] is SHANGHAI_TZ
+    days = store.list_fund_days(5)
+    assert days[0]["date"] == 20260906

tests/unit/test_task_store.py

@@ -257,3 +257,78 @@ def test_task_list_includes_persisted_history_after_new_app(persisted_env):
     client = _client()
     tasks = client.get("/api/v1/backtest/tasks?limit=50").json()["tasks"]
     assert any(t["task_id"] == task_id for t in tasks)
+
+
+# ── v1.32.6 修复：列表查询懒加载 result_json + 淘汰只移终态 ────────────────────
+
+
+def test_list_recent_without_results_skips_result_json(persisted_env):
+    """include_results=False 时不 SELECT/解析 result_json（列表页瘦身）。
+
+    旧签名无该参数：列表页会把每条任务的完整结果 JSON 拖进内存解析。
+    """
+    import sqlite3
+
+    from easy_tdx.web.task_store import TaskStore
+
+    store = TaskStore()
+    store.save(
+        task_id="big",
+        status="done",
+        created_at=2.0,
+        result={
+            "performance": {"total_return": 0.25},
+            "equity_curve": [{"i": i} for i in range(500)],
+        },
+    )
+    # 把 result_json 打坏：include_results=False 路径根本不读它 → 不受影响
+    with sqlite3.connect(store.path) as conn:
+        conn.execute("UPDATE backtest_tasks SET result_json = '{not-json' WHERE task_id='big'")
+        conn.commit()
+
+    rows = store.list_recent(limit=10, include_results=False)
+    assert len(rows) == 1
+    assert rows[0]["task_id"] == "big"
+    assert rows[0]["status"] == "done"
+    assert rows[0]["result"] is None
+
+    # 详情（load）仍取全量并走损坏降级路径
+    d = store.load("big")
+    assert d is not None and d["result"] is None
+
+
+def test_eviction_only_removes_terminal_states():
+    """超限淘汰只移 done/failed；pending/running 不淘汰（消灭 pending 幽灵）。
+
+    旧实现"淘汰最旧 non-running"会把最早提交、尚未起跑的 pending 条目淘汰，
+    其 worker 随后取不到状态直接跳过 → 磁盘遗留 pending 行被恢复成永久
+    pending 的幽灵任务。
+    """
+    from easy_tdx.web.task_runner import BacktestTaskRunner, TaskState
+
+    runner = BacktestTaskRunner(max_workers=1, max_results=2)
+    with runner._lock:
+        # 插入顺序即 LRU 序：pending 最旧、done 最新
+        runner._tasks["p1"] = TaskState(task_id="p1", status="pending")
+        runner._tasks["r1"] = TaskState(task_id="r1", status="running")
+        runner._tasks["d1"] = TaskState(task_id="d1", status="done")
+        runner._evict_if_needed_locked()
+
+    assert "p1" in runner._tasks, "pending 不得被淘汰"
+    assert "r1" in runner._tasks, "running 不得被淘汰"
+    assert "d1" not in runner._tasks, "超限时应淘汰最旧的终态条目"
+
+
+def test_eviction_pending_still_evicted_by_old_logic_regression_guard():
+    """回归对照：若恢复旧逻辑（淘汰首个 non-running），pending 会先被选中。
+
+    本测试钉死新语义——全部条目为 pending 时宁可不淘汰（超限跳过）。
+    """
+    from easy_tdx.web.task_runner import BacktestTaskRunner, TaskState
+
+    runner = BacktestTaskRunner(max_workers=1, max_results=1)
+    with runner._lock:
+        runner._tasks["p1"] = TaskState(task_id="p1", status="pending")
+        runner._tasks["p2"] = TaskState(task_id="p2", status="pending")
+        runner._evict_if_needed_locked()
+    assert len(runner._tasks) == 2  # 无终态可淘汰 → 跳过，不丢任务

tests/unit/test_unusual.py

@@ -17,6 +17,8 @@
 import struct
 from datetime import time
 
+import pytest
+
 from easy_tdx import UNUSUAL_TYPE_NAMES
 from easy_tdx.mac.commands.unusual import UnusualCmd, _describe_unusual
 
@@ -48,15 +50,20 @@ class TestDescribeUnusualKnownTypes:
 
     def test_type_0x04(self):
         # 真实样本：605365 立达信 2026-09-01 09:35:11
-        desc, val = _describe_unusual(0x04, bytes.fromhex("00b8d73d3d0000000000000000"))
+        desc, val = _describe_unusual(0x04, bytes.fromhex("00b8d73d3d0000000000000000"), 9)
         assert desc == "加速拉升"
         assert val == "4.63%"
 
     def test_unknown_type_fallback(self):
-        desc, val = _describe_unusual(0x42, bytes.fromhex("00" * 13))
+        desc, val = _describe_unusual(0x42, bytes.fromhex("00" * 13), 9)
         assert desc == "异动类型0x42"
         assert val == ""
 
+    def test_hour_is_required(self):
+        """hour 必传：缺省 9 会让 15:00 的 0x15 被误标「竞价」（回归防护）。"""
+        with pytest.raises(TypeError):
+            _describe_unusual(0x15, bytes.fromhex("00" * 13))
+
 
 class TestType0x15:
     """0x15 竞价/尾盘异动（Issue #62）。
@@ -67,19 +74,19 @@ class TestType0x15:
 
     def test_auction_drop(self):
         # 真实样本：600551 时代出版 09:25:00，v1=0x03 竞价下跌
-        desc, val = _describe_unusual(0x15, bytes.fromhex("030c9846bc003e1d4700000000"))
+        desc, val = _describe_unusual(0x15, bytes.fromhex("030c9846bc003e1d4700000000"), 9)
         assert desc == "竞价下跌"
         assert val == "-1.21%/40254手"
 
     def test_auction_rise(self):
         # 真实样本：600127 金健米业 09:25:00，v1=0x02 竞价拉升（尾段自 10.84 冲至 12.05）
-        desc, val = _describe_unusual(0x15, bytes.fromhex("0213d2cd3d00367b4700000000"))
+        desc, val = _describe_unusual(0x15, bytes.fromhex("0213d2cd3d00367b4700000000"), 9)
         assert desc == "竞价拉升"
         assert val == "10.05%/64310手"
 
     def test_auction_flat(self):
         # 真实样本：600410 华胜天成 09:25:01，v1=0x01 竞价平稳（尾段价格未动）
-        desc, val = _describe_unusual(0x15, bytes.fromhex("01000000000098a54500000000"))
+        desc, val = _describe_unusual(0x15, bytes.fromhex("01000000000098a54500000000"), 9)
         assert desc == "竞价平稳"
         assert val == "0.00%/5299手"
 
@@ -96,7 +103,7 @@ class TestType0x15:
         assert val == "-0.77%/171634手"
 
     def test_unknown_sub_type_falls_back(self):
-        desc, _ = _describe_unusual(0x15, struct.pack("<B2fI", 0x77, 0.0, 100.0, 0))
+        desc, _ = _describe_unusual(0x15, struct.pack("<B2fI", 0x77, 0.0, 100.0, 0), 9)
         assert desc == "竞价异动"
         desc, _ = _describe_unusual(0x15, struct.pack("<B2fI", 0x77, 0.0, 100.0, 0), 15)
         assert desc == "尾盘异动"
@@ -107,19 +114,19 @@ class TestType0x16:
 
     def test_strong_at_auction(self):
         # 真实样本：600551 时代出版 09:25:00，v2=+5.82% 与当日开盘涨幅精确一致
-        desc, val = _describe_unusual(0x16, bytes.fromhex("010f506e3dcb846e3d00000000"))
+        desc, val = _describe_unusual(0x16, bytes.fromhex("010f506e3dcb846e3d00000000"), 9)
         assert desc == "盘中强势"
         assert val == "5.82%"
 
     def test_weak_at_auction(self):
         # 真实样本：600683 京投发展 09:25:01，v1=0xFF（弱势 1 级），v2=-6.40%
-        desc, val = _describe_unusual(0x16, bytes.fromhex("ffc71d83bd690383bd00000000"))
+        desc, val = _describe_unusual(0x16, bytes.fromhex("ffc71d83bd690383bd00000000"), 9)
         assert desc == "盘中弱势"
         assert val == "-6.40%"
 
     def test_new_stock_no_limit(self):
         # 真实样本：601123 N马矿 09:25:00，新股无涨跌幅限制，v2=+245.86%
-        desc, val = _describe_unusual(0x16, bytes.fromhex("03775a1d404a5b1d4000000000"))
+        desc, val = _describe_unusual(0x16, bytes.fromhex("03775a1d404a5b1d4000000000"), 9)
         assert desc == "盘中强势"
         assert val == "245.86%"
 
@@ -134,13 +141,13 @@ class TestType0x13:
 
     def test_auction_test_buy(self):
         # 真实样本：603980 09:15:14，申报价 8.71 高于昨收 7.92（往上试）
-        desc, val = _describe_unusual(0x13, bytes.fromhex("00295c0b41006c354600000000"))
+        desc, val = _describe_unusual(0x13, bytes.fromhex("00295c0b41006c354600000000"), 9)
         assert desc == "竞价试买"
         assert val == "8.71/11611手"
 
     def test_auction_test_sell(self):
         # 真实样本：603900 09:15:17，申报价 6.46 低于昨收 7.17（往下试）
-        desc, val = _describe_unusual(0x13, bytes.fromhex("0152b8ce400000c94300000000"))
+        desc, val = _describe_unusual(0x13, bytes.fromhex("0152b8ce400000c94300000000"), 9)
         assert desc == "竞价试卖"
         assert val == "6.46/402手"
 
@@ -150,13 +157,13 @@ class TestType0x1D0x1E:
 
     def test_fast_rise(self):
         # 真实样本：605365 立达信 09:35:08
-        desc, val = _describe_unusual(0x1D, bytes.fromhex("009d50843c0000000000000000"))
+        desc, val = _describe_unusual(0x1D, bytes.fromhex("009d50843c0000000000000000"), 9)
         assert desc == "急速拉升"
         assert val == "1.62%"
 
     def test_fast_fall(self):
         # 真实样本：601123 N马矿 09:35:03
-        desc, val = _describe_unusual(0x1E, bytes.fromhex("019cd393bc0000000000000000"))
+        desc, val = _describe_unusual(0x1E, bytes.fromhex("019cd393bc0000000000000000"), 9)
         assert desc == "急速下跌"
         assert val == "-1.80%"
 
@@ -211,7 +218,7 @@ class TestTypeNames:
         """映射表中的类型不应落入"异动类型0x??"兜底分支。"""
         zeros = bytes.fromhex("00" * 13)
         for utype in UNUSUAL_TYPE_NAMES:
-            desc, _ = _describe_unusual(utype, zeros)
+            desc, _ = _describe_unusual(utype, zeros, 9)
             assert not desc.startswith("异动类型"), f"0x{utype:02X} 未实现解析分支"
 
     def test_top_level_export(self):

tests/unit/test_vipdoc_settings.py

@@ -62,7 +62,7 @@ def test_vipdoc_settings_endpoints(settings_env, tmp_path):
         r = client.put("/api/v1/settings/vipdoc", json={"path": str(real_dir)})
         assert r.status_code == 200
         assert r.json()["stored"] == str(real_dir)
-        assert market_mod._limitup_cache is None
+        assert market_mod._limitup_cache == {}
 
         # GET 回读
         assert client.get("/api/v1/settings/vipdoc").json()["stored"] == str(real_dir)

tests/unit/test_warehouse.py

@@ -2,6 +2,9 @@
 
 from __future__ import annotations
 
+import datetime as _dt
+import logging
+
 import numpy as np
 import pandas as pd
 import pytest
@@ -176,6 +179,131 @@ def test_promote_provisional(wh):
     assert len(wh.query("SH", "600519")) == 3  # 转正后可见
 
 
+def test_promote_provisional_scoped_to_market_code_and_before(wh):
+    """scoped 转正：只转正指定标的且 datetime <= before 的 provisional 行。"""
+
+    def _one(d: str) -> pd.DataFrame:
+        return pd.DataFrame(
+            {
+                "datetime": pd.date_range(d, periods=1),
+                "open": 10.0,
+                "high": 10.1,
+                "low": 9.9,
+                "close": 10.0,
+                "vol": 100.0,
+                "amount": 1000.0,
+            }
+        )
+
+    wh.upsert_bars("SH", "600519", _one("2024-01-05"), status="provisional")
+    wh.upsert_bars("SH", "600519", _one("2024-06-01"), status="provisional")
+    wh.upsert_bars("SZ", "000001", _one("2024-01-05"), status="provisional")
+
+    n = wh.promote_provisional(market="SH", code="600519", before=pd.Timestamp("2024-03-01"))
+    assert n == 1
+    out = wh.query("SH", "600519")  # 默认查询只含 completed
+    assert len(out) == 1
+    assert pd.Timestamp(out["datetime"].iloc[0]) == pd.Timestamp("2024-01-05")
+    all_rows = wh.query("SH", "600519", include_provisional=True)
+    assert len(all_rows) == 2  # 2024-06-01 行超出 before，保持 provisional
+
+
+def _fake_clock(
+    store_mod,  # noqa: ANN001 — monkeypatch 目标模块（未用）
+    *,
+    shanghai: tuple[int, int, int, int],
+    local: tuple[int, int, int, int],
+):
+    """伪造 store 模块时钟：now(tz)=沪时区正确墙钟；now()=本地误判墙钟。
+
+    模拟「UTC 主机」：沪市已 18:00（当日 bar 应为 completed），本地 naive
+    时钟却还是 10:00（旧实现会误标 provisional）。
+    """
+
+    class _FixedDT(_dt.datetime):
+        @classmethod
+        def now(cls, tz=None):  # type: ignore[override]
+            if tz is not None:
+                y, m, d, hh = shanghai
+                return _dt.datetime(y, m, d, hh, 0, tzinfo=tz)
+            y, m, d, hh = local
+            return _dt.datetime(y, m, d, hh, 0)
+
+    return _FixedDT
+
+
+def test_provisional_uses_shanghai_clock_not_local(wh, monkeypatch):
+    """provisional 判定按沪市墙钟：沪市 18:00（收盘后）当日 bar 必须 completed。
+
+    回归：旧实现用系统本地 now()——UTC 主机上沪市收盘时本地才 10:00，
+    当日 bar 被误标 provisional，默认查询隐藏当天数据。
+    """
+    import easy_tdx.warehouse.store as store_mod
+
+    monkeypatch.setattr(
+        store_mod,
+        "datetime",
+        _fake_clock(store_mod, shanghai=(2026, 9, 7, 18), local=(2026, 9, 7, 10)),
+    )
+
+    dates = pd.date_range(
+        pd.Timestamp("2026-09-07") - pd.Timedelta(days=10), periods=11, freq="D"
+    ).tolist()  # 2026-08-28 .. 2026-09-07（末根 = 沪市「当日」）
+    df = pd.DataFrame(
+        {
+            "datetime": dates,
+            "open": 10.0,
+            "high": 10.1,
+            "low": 9.9,
+            "close": 10.0,
+            "vol": 100.0,
+            "amount": 1000.0,
+        }
+    )
+    wh.upsert_bars("SH", "600519", df)
+    # 沪市已收盘：全部 11 根都应为 completed（旧实现：当日根 provisional）
+    assert len(wh.query("SH", "600519")) == 11
+
+
+def test_promote_provisional_uses_shanghai_date(wh, monkeypatch):
+    """无参转正的「今日」边界按沪市日期：沪市已过 0 点即转正昨日临时行。"""
+    import easy_tdx.warehouse.store as store_mod
+
+    monkeypatch.setattr(
+        store_mod,
+        "datetime",
+        _fake_clock(store_mod, shanghai=(2026, 9, 8, 0), local=(2026, 9, 7, 16)),
+    )
+
+    old = pd.DataFrame(
+        {
+            "datetime": pd.date_range("2026-09-07", periods=1),
+            "open": 10.0,
+            "high": 10.1,
+            "low": 9.9,
+            "close": 10.0,
+            "vol": 100.0,
+            "amount": 1000.0,
+        }
+    )
+    wh.upsert_bars("SH", "600519", old, status="provisional")
+    # 沪市日期已是 9/8 → 9/7 的临时行应转正（旧实现按本地 9/7 → n=0）
+    assert wh.promote_provisional() == 1
+    assert len(wh.query("SH", "600519")) == 1
+
+
+def test_open_conflict_clear_error(tmp_path, monkeypatch):
+    """仓库文件被其他进程占用：给可操作的中文错误而非裸 duckdb 异常。"""
+    import duckdb as duckdb_mod
+
+    def _raise(*args, **kwargs):  # type: ignore[no-untyped-def]
+        raise duckdb_mod.IOException("Could not set lock on file")
+
+    monkeypatch.setattr(duckdb_mod, "connect", _raise)
+    with pytest.raises(RuntimeError, match="占用"):
+        KlineWarehouse(tmp_path / "lock.duckdb")
+
+
 # ── 健康自检 ─────────────────────────────────────────────────────────────────
 
 
@@ -291,6 +419,115 @@ def test_sync_progress_callback(tmp_path):
         warehouse.close()
 
 
+class _ScriptedClient:
+    """按调用序返回预置 DataFrame 的假客户端（末帧可重复）。"""
+
+    def __init__(self, frames: list[pd.DataFrame]) -> None:
+        self._frames = frames
+        self.calls: list[int] = []
+
+    def get_stock_kline(self, market, code, period="DAILY", start=0, count=800, adjust="NONE"):
+        self.calls.append(count)
+        idx = min(len(self.calls) - 1, len(self._frames) - 1)
+        return self._frames[idx].copy()
+
+
+def test_sync_refetch_full_when_tail_gap(tmp_path, caplog):
+    """增量尾部覆盖不到上次同步点（首 bar 晚于 existing_last）→ 全量重拉补缺。
+
+    回归：旧实现固定只拉 tail_bars 根——超过 15 个交易日未同步的标的，
+    中间日期永不补齐且无任何告警。
+    """
+    warehouse = KlineWarehouse(tmp_path / "gap.duckdb")
+    try:
+        source_full = _bars(130)  # 2024-01-01 起 130 个工作日
+        initial = source_full.iloc[:100]  # 首同步窗口（末根 idx99）
+        stale_tail = source_full.iloc[115:]  # 增量窗口：首根 idx115 > idx99 → 有缺口
+        client = _ScriptedClient([initial, stale_tail, source_full])
+        syncer = WarehouseSyncer(client, warehouse, max_bars=800, tail_bars=15)
+
+        with caplog.at_level(logging.WARNING, logger="easy_tdx.warehouse.sync"):
+            syncer.sync(["SH:600519"])
+            syncer.sync(["SH:600519"])
+
+        assert client.calls == [800, 15, 800]  # 第二次 sync 触发了全量重拉
+        rows = warehouse.query("SH", "600519")
+        assert len(rows) == 130  # 无缺口
+        bridge = pd.Timestamp(source_full["datetime"].iloc[100])
+        dts = pd.to_datetime(rows["datetime"])
+        assert (dts == bridge).any()  # 缺口桥接 bar 已补上
+        assert "缺口" in caplog.text
+    finally:
+        warehouse.close()
+
+
+def test_sync_failure_keeps_provisional(tmp_path):
+    """拉取失败：不转正 provisional，盘中临时值不会被洗成 completed。"""
+    warehouse = KlineWarehouse(tmp_path / "keep.duckdb")
+    try:
+        old = pd.DataFrame(
+            {
+                "datetime": pd.date_range("2024-01-01", periods=3),
+                "open": 10.0,
+                "high": 10.1,
+                "low": 9.9,
+                "close": 10.0,
+                "vol": 100.0,
+                "amount": 1000.0,
+            }
+        )
+        warehouse.upsert_bars("SH", "600519", old, status="provisional")
+
+        class _BadClient:
+            def get_stock_kline(self, *a, **kw):
+                raise ConnectionError("断网")
+
+        s = WarehouseSyncer(_BadClient(), warehouse).sync(["SH:600519"])
+        assert s["failed"] == 1
+        # 仍为 provisional：默认查询不可见（旧实现 sync 前盲转正 → 可见）
+        assert len(warehouse.query("SH", "600519")) == 0
+        assert len(warehouse.query("SH", "600519", include_provisional=True)) == 3
+    finally:
+        warehouse.close()
+
+
+def test_sync_promotes_only_up_to_fetched_max(tmp_path):
+    """转正上界 = 本次成功拉到的最大 datetime：未覆盖到的行保持 provisional。"""
+    warehouse = KlineWarehouse(tmp_path / "bound.duckdb")
+    try:
+
+        def _one(d: str) -> pd.DataFrame:
+            return pd.DataFrame(
+                {
+                    "datetime": pd.date_range(d, periods=1),
+                    "open": 10.0,
+                    "high": 10.1,
+                    "low": 9.9,
+                    "close": 10.0,
+                    "vol": 100.0,
+                    "amount": 1000.0,
+                }
+            )
+
+        warehouse.upsert_bars("SH", "600519", _one("2024-01-05"), status="provisional")
+        warehouse.upsert_bars("SH", "600519", _one("2024-06-01"), status="provisional")
+
+        fetched = _bars(11, start="2024-01-10")  # 最大 datetime 2024-01-24
+        client = _ScriptedClient([fetched])
+        WarehouseSyncer(client, warehouse, tail_bars=15).sync(["SH:600519"])
+
+        completed = warehouse.query("SH", "600519")
+        # 01-05 行 <= 拉取上界 → 已转正；06-01 行超出上界 → 保持 provisional
+        assert len(completed) == 12
+        all_rows = warehouse.query("SH", "600519", include_provisional=True)
+        assert len(all_rows) == 13
+        stale = all_rows[all_rows["status"] == "provisional"]
+        assert len(stale) == 1
+        assert pd.Timestamp(stale["datetime"].iloc[0]) == pd.Timestamp("2024-06-01")
+    finally:
+        warehouse.close()
+
+
 def test_missing_duckdb_helpful_error(tmp_path, monkeypatch):
     """duckdb 未安装时给出安装指引（模拟 ImportError）。"""
     import builtins

tests/unit/test_watchlist_and_streamer.py

@@ -153,3 +153,52 @@ def test_is_trading_hours() -> None:
     assert _is_trading_hours(datetime(2026, 9, 1, 10, 0, tzinfo=tz)) is True  # 周二盘中
     assert _is_trading_hours(datetime(2026, 9, 1, 3, 0, tzinfo=tz)) is False  # 凌晨
     assert _is_trading_hours(datetime(2026, 9, 5, 10, 0, tzinfo=tz)) is False  # 周六
+
+
+# ── /watchlist 端点 code 格式校验（v1.32.6）─────────────────────────────────
+
+
+def _watch_app(monkeypatch, tmp_path):
+    from fastapi import FastAPI
+
+    from easy_tdx.web import watchlist_store as ws
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import watchlist as watchlist_mod
+
+    monkeypatch.setenv("EASY_TDX_CONFIG_DIR", str(tmp_path / "cfg"))
+    ws._store = None
+
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(watchlist_mod.router, prefix="/api/v1")
+    return app
+
+
+def test_watchlist_add_rejects_non_numeric_code(monkeypatch, tmp_path):
+    """code 非 6 位数字 → 422（旧实现可把 'abcdef' 存进自选并喂给轮询器）。"""
+    pytest.importorskip("fastapi")
+    from fastapi.testclient import TestClient
+
+    app = _watch_app(monkeypatch, tmp_path)
+    with TestClient(app) as client:
+        bad = client.post("/api/v1/watchlist", json={"market": "SZ", "code": "abcdef", "name": "x"})
+        assert bad.status_code == 422
+        short = client.post(
+            "/api/v1/watchlist", json={"market": "SZ", "code": "00001", "name": "x"}
+        )
+        assert short.status_code == 422
+        ok = client.post(
+            "/api/v1/watchlist", json={"market": "SZ", "code": "000001", "name": "平安银行"}
+        )
+        assert ok.status_code == 200
+
+
+def test_watchlist_remove_validates_code_format(monkeypatch, tmp_path):
+    """remove 路径 code 非 6 位数字 → 422，不触达存储。"""
+    pytest.importorskip("fastapi")
+    from fastapi.testclient import TestClient
+
+    app = _watch_app(monkeypatch, tmp_path)
+    with TestClient(app) as client:
+        resp = client.delete("/api/v1/watchlist/SZ/abc123")
+        assert resp.status_code == 422

tests/unit/test_web_api.py

@@ -599,3 +599,33 @@ def test_create_app_no_ui_mode():
     app_ui = create_app()
     mounts_ui = [r for r in app_ui.routes if type(r).__name__ == "Mount"]
     assert any(getattr(m, "name", "") == "web-ui" for m in mounts_ui)
+
+
+# ── /server/test 输入约束（v1.32.6：timeout 上界 + hosts 限长）───────────────
+
+
+def test_server_test_request_constraints():
+    """timeout 限 0.5~30s；hosts ≤50 项且单项 ≤253 字符。
+
+    旧实现 timeout 无上界（1e9 会把 to_thread 线程挂死）、hosts 不限长
+    （可当内网扫描跳板）。
+    """
+    pytest.importorskip("fastapi")
+    from pydantic import ValidationError
+
+    from easy_tdx.web.routers.server import ServerTestRequest
+
+    assert ServerTestRequest(hosts=None, timeout=5.0).timeout == 5.0
+    assert ServerTestRequest(hosts=["127.0.0.1"], timeout=0.5).timeout == 0.5
+
+    with pytest.raises(ValidationError):
+        ServerTestRequest(timeout=31.0)  # 超上界
+    with pytest.raises(ValidationError):
+        ServerTestRequest(timeout=0.1)  # 低于下界
+    with pytest.raises(ValidationError):
+        ServerTestRequest(hosts=[f"h{i}" for i in range(51)])  # 超 50 项
+    with pytest.raises(ValidationError):
+        ServerTestRequest(hosts=["x" * 254])  # 单项超 253 字符
+    # 边界可用
+    ok = ServerTestRequest(hosts=["h" * 253] * 50, timeout=30.0)
+    assert len(ok.hosts) == 50

tests/unit/test_web_backtest.py

@@ -349,37 +349,31 @@ def test_task_runner_captures_failure():
 
 
 def test_task_runner_lru_eviction():
-    """超过上限应丢弃最旧的非 running 任务。
+    """超限淘汰只移终态任务（v1.32.6：pending/running 不淘汰，防幽灵任务）。
 
-    注意：淘汰发生在 submit 时，淘汰对象是「当时最旧的非 running 任务」。
-    用 max_workers=1 串行执行时，哪个任务被淘汰取决于提交速度 vs 执行速度
-    的竞态（快机器上 t0 还在 running 会被跳过，慢机器上 t0 已完成会被淘汰）。
-    所以本测试不断言「特定 task_id 被淘汰」，而是验证：
-    (1) 存活的 non-running 任务数 ≤ max_results
-    (2) 最后提交的任务一定存活（它是最近的，不可能被 LRU 淘汰）
-    (3) 至少有 2 个任务被淘汰（5 提交 - 3 上限 = 2）
+    确定性设计：max_workers=1 串行。先提交 5 个并等全部 done（提交瞬间的
+    淘汰因全是 pending 而跳过——这正是新语义），随后提交第 6 个触发淘汰：
+    此时 5 个全是终态，按 LRU 淘到只剩 max_results=3（t3/t4/t5）。
     """
     from easy_tdx.web.task_runner import BacktestTaskRunner
 
     runner = BacktestTaskRunner(max_workers=1, max_results=3)
     ids = [runner.submit(lambda: {"i": i}, description=f"t{i}") for i in range(5)]
-    # 等待存活的任务全部完成（被淘汰的 peek 返回 None，跳过）
+    # 等 5 个任务全部完成
     for _ in range(200):
-        alive = [tid for tid in ids if runner.peek(tid) is not None]
-        if all(runner.peek(tid).status in ("done", "failed") for tid in alive):
+        states = [runner.peek(tid) for tid in ids]
+        if all(s is not None and s.status in ("done", "failed") for s in states):
             break
         time.sleep(0.02)
 
-    # 最后提交的任务一定存活（LRU 最近，不可能被淘汰）
-    assert runner.peek(ids[4]) is not None, "最后提交的任务不应被淘汰"
-
-    # 至少淘汰 2 个（5 提交 - max_results 3 = 2）
-    surviving = [tid for tid in ids if runner.peek(tid) is not None]
-    evicted = [tid for tid in ids if runner.peek(tid) is None]
-    assert len(evicted) >= 2, f"应至少淘汰 2 个任务，实际淘汰 {len(evicted)} 个"
-
-    # 存活任务数不超过 max_results（running 完成后）
-    assert len(surviving) <= 3, f"存活任务 {len(surviving)} 超过上限 3"
+    # 全部完成后提交第 6 个 → 淘汰最旧的 3 个 done（t0/t1/t2）
+    ids.append(runner.submit(lambda: {"i": 5}, description="t5"))
+    assert runner.peek(ids[5]) is not None
+    assert runner.peek(ids[0]) is None, "最旧的 done 应被淘汰"
+    assert runner.peek(ids[1]) is None
+    assert runner.peek(ids[2]) is None
+    for tid in ids[3:]:
+        assert runner.peek(tid) is not None, "最近的任务不应被淘汰"
 
     runner.shutdown()
 
@@ -1365,3 +1359,284 @@ def test_multi_strategy_evaluate_endpoint(client, monkeypatch):
     assert report["grade"]["scenario"] == "portfolio"
     assert report["fitness"]["total_checks"] == 8
     assert report["config"]["slots"] == ["双均线交叉@SH:601088", "MACD 金叉@SZ:000001"]
+
+
+# ── submit 响应真实状态（v1.32.6 修复：不再把 done/failed 谎报为 running）──────
+
+
+class _InstantDoneRunner:
+    """submit 即同步跑完的假 runner（模拟"拿到 future 前任务已完成"）。"""
+
+    def __init__(self, status: str = "done"):
+        self.status = status
+
+    def submit(self, func, *, description=""):
+        func()  # 同步执行完毕
+        return "tid-done"
+
+    def get(self, task_id):
+        from easy_tdx.web.task_runner import TaskState
+
+        return TaskState(
+            task_id=task_id,
+            status=self.status,  # type: ignore[arg-type]
+            result={
+                "performance": {},
+                "equity_curve": [],
+                "trades": [],
+                "positions": [],
+                "config": {},
+            },
+            finished_at=1.0,
+            started_at=0.0,
+            created_at=0.0,
+        )
+
+
+def test_task_submit_response_accepts_terminal_status():
+    """TaskSubmitResponse.status 应接受 done/failed（旧 Literal 只许 pending/running）。"""
+    from easy_tdx.web.backtest_schemas import TaskSubmitResponse
+
+    assert TaskSubmitResponse(task_id="x", status="done").status == "done"
+    assert TaskSubmitResponse(task_id="x", status="failed").status == "failed"
+    assert TaskSubmitResponse(task_id="x", status="pending").status == "pending"
+
+
+def test_async_submit_reports_real_done_status(monkeypatch):
+    """极快任务已 done 时，202 响应应透传真实状态 "done"（旧码谎报 "running"）。"""
+    from fastapi import FastAPI
+    from fastapi.testclient import TestClient
+
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import backtest as backtest_mod
+
+    monkeypatch.setattr(backtest_mod, "get_runner", lambda: _InstantDoneRunner("done"))
+
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(backtest_mod.router, prefix="/api/v1")
+    app.state.tdx_client = object()
+
+    with TestClient(app) as tc:
+        resp = tc.post(
+            "/api/v1/backtest/run/async",
+            json={
+                "strategy": "ma_cross",
+                "params": {"fast": 3, "slow": 6},
+                "ohlcv": [
+                    {
+                        "datetime": f"2024-01-{d:02d}",
+                        "open": 10.0,
+                        "high": 10.5,
+                        "low": 9.5,
+                        "close": 10.0,
+                        "vol": 1000.0,
+                        "amount": 10000.0,
+                    }
+                    for d in range(1, 6)
+                ],
+            },
+        )
+    assert resp.status_code == 202, resp.text
+    assert resp.json()["status"] == "done"
+
+
+def test_async_submit_reports_failed_status(monkeypatch):
+    """任务同步失败时响应应报 "failed" 而非 "running"。"""
+    from fastapi import FastAPI
+    from fastapi.testclient import TestClient
+
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import backtest as backtest_mod
+
+    class _FailingRunner(_InstantDoneRunner):
+        def submit(self, func, *, description=""):
+            try:
+                func()
+            except Exception:
+                pass
+            return "tid-fail"
+
+    monkeypatch.setattr(backtest_mod, "get_runner", lambda: _FailingRunner("failed"))
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(backtest_mod.router, prefix="/api/v1")
+    app.state.tdx_client = object()
+
+    with TestClient(app) as tc:
+        resp = tc.post(
+            "/api/v1/backtest/run/async",
+            json={
+                "strategy": "no_such_strategy",
+                "ohlcv": [
+                    {
+                        "datetime": "2024-01-01",
+                        "open": 10.0,
+                        "high": 10.5,
+                        "low": 9.5,
+                        "close": 10.0,
+                        "vol": 1000.0,
+                        "amount": 10000.0,
+                    },
+                    {
+                        "datetime": "2024-01-02",
+                        "open": 10.0,
+                        "high": 10.5,
+                        "low": 9.5,
+                        "close": 10.0,
+                        "vol": 1000.0,
+                        "amount": 10000.0,
+                    },
+                ],
+            },
+        )
+    assert resp.status_code == 202, resp.text
+    assert resp.json()["status"] == "failed"
+
+
+def test_formula_submit_reports_real_status(monkeypatch):
+    """formula 回测提交响应透传真实状态（旧码硬编码 "running"）。"""
+    from fastapi import FastAPI
+    from fastapi.testclient import TestClient
+
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import formula as formula_mod
+
+    monkeypatch.setattr(formula_mod, "get_runner", lambda: _InstantDoneRunner("done"))
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(formula_mod.router, prefix="/api/v1")
+    app.state.tdx_client = object()
+
+    ohlcv = [
+        {
+            "datetime": f"2024-01-{d:02d}",
+            "open": 10.0,
+            "high": 10.5,
+            "low": 9.5,
+            "close": 10.0,
+            "vol": 1000.0,
+        }
+        for d in range(1, 6)
+    ]
+    with TestClient(app) as tc:
+        resp = tc.post(
+            "/api/v1/formula/backtest/run/async",
+            json={"text": "CROSS(C, MA(C, 3));", "ohlcv": ohlcv},
+        )
+    assert resp.status_code == 202, resp.text
+    assert resp.json()["status"] == "done"
+
+
+# ── optimize 费率口径（stamp_tax / min_commission / auto_fees 透传）───────────
+
+
+def test_optimize_request_accepts_fee_fields():
+    """OptimizeBacktestRequest 应支持 stamp_tax/min_commission/auto_fees（镜像单标的）。"""
+    from easy_tdx.web.backtest_schemas import OptimizeBacktestRequest
+
+    req = OptimizeBacktestRequest(
+        strategy="ma_cross",
+        param_grid={"fast": [3]},
+        symbol="SZ:000001",
+        stamp_tax=0.0,
+        min_commission=1.0,
+        auto_fees=True,
+    )
+    assert req.stamp_tax == 0.0
+    assert req.min_commission == 1.0
+    assert req.auto_fees is True
+    with pytest.raises(ValueError):
+        OptimizeBacktestRequest(
+            strategy="ma_cross",
+            param_grid={"fast": [3]},
+            symbol="SZ:000001",
+            stamp_tax=0.5,  # > le=0.01
+        )
+
+
+def test_optimize_auto_fees_etf_matches_explicit_fee_backtest(sample_ohlcv):
+    """ETF + auto_fees：寻优结果的买入持有基准与"显式 ETF 费率"口径一致。
+
+    旧实现不透传 stamp_tax/min_commission（恒按股票默认 0.001/5.0），品种
+    口径无法生效。用可转债（佣金 0.0002 / 最低佣金 1.0 / 免印花税）验证：
+    寻优结果的买入持有基准与"显式可转债费率"同口径，且与旧股票默认口径
+    不同（_BuyAndHold 不卖出，印花税不进收益，差异来自佣金/最低佣金）。
+    """
+    from easy_tdx.backtest.benchmark import run_buy_hold_benchmark
+    from easy_tdx.web.backtest_schemas import OptimizeBacktestRequest
+    from easy_tdx.web.routers.backtest import _run_optimize
+
+    df = pd.DataFrame(sample_ohlcv)
+    df["datetime"] = pd.to_datetime(df["datetime"])
+    req = OptimizeBacktestRequest(
+        strategy="ma_cross",
+        param_grid={"fast": [3], "slow": [10]},
+        cash=1_000_000.0,
+        symbol="SH:110059",  # 可转债：佣金 0.0002 / 最低佣金 1.0 / 免印花税
+        auto_fees=True,
+    )
+    out = _run_optimize(df, req)
+
+    # 期望口径：可转债费率
+    expected = run_buy_hold_benchmark(
+        df,
+        cash=req.cash,
+        commission=0.0002,
+        min_commission=1.0,
+        slippage=req.slippage,
+        execution=req.execution,
+    )
+    legacy = run_buy_hold_benchmark(
+        df,
+        cash=req.cash,
+        commission=req.commission,  # 0.0003（股票默认）
+        min_commission=5.0,
+        slippage=req.slippage,
+        execution=req.execution,
+    )
+    assert out["buy_hold"] is not None
+    assert out["buy_hold"]["total_return"] == pytest.approx(expected["total_return"])
+    # 若与旧股票默认口径相同则说明 auto_fees 没生效
+    assert out["buy_hold"]["total_return"] != pytest.approx(legacy["total_return"])
+
+
+def test_optimize_passes_resolved_fees_to_optimizer(sample_ohlcv, monkeypatch):
+    """auto_fees 解析出的费率应透传给 ParamGridOptimizer（显式值优先）。"""
+    import easy_tdx.backtest.optimizer as opt_mod
+    from easy_tdx.web.backtest_schemas import OptimizeBacktestRequest
+    from easy_tdx.web.routers.backtest import _run_optimize
+
+    captured: dict = {}
+
+    class SpyOptimizer(opt_mod.ParamGridOptimizer):
+        def __init__(self, *args, **kwargs):
+            captured.update(kwargs)
+            super().__init__(*args, **kwargs)
+
+    monkeypatch.setattr(opt_mod, "ParamGridOptimizer", SpyOptimizer)
+
+    df = pd.DataFrame(sample_ohlcv)
+    df["datetime"] = pd.to_datetime(df["datetime"])
+    req = OptimizeBacktestRequest(
+        strategy="ma_cross",
+        param_grid={"fast": [3], "slow": [10]},
+        symbol="SH:510300",
+        auto_fees=True,
+    )
+    _run_optimize(df, req)
+    # ETF 口径：印花税解析为 0
+    assert captured.get("stamp_tax") == 0.0
+    assert captured.get("min_commission") == 5.0
+
+    # 显式非默认 stamp_tax 优先于品种默认
+    captured.clear()
+    req2 = OptimizeBacktestRequest(
+        strategy="ma_cross",
+        param_grid={"fast": [3], "slow": [10]},
+        symbol="SH:510300",
+        auto_fees=True,
+        stamp_tax=0.005,
+    )
+    _run_optimize(df, req2)
+    assert captured.get("stamp_tax") == 0.005

tests/unit/test_web_bars_fallback.py

@@ -0,0 +1,219 @@
+"""/bars、/bars/index 的 baostock 兜底集成测试（v1.32.6 修复项）。
+
+覆盖：
+- 数字周期字符串（category="4"）归一后也能走兜底（旧实现直接透传原串，
+  baostock 频率查表落空 → 兜底静默失效，维持原错误）；
+- 指数兜底必须传 is_index=True（baostock 指数 vol 股→手），个股路径不传；
+- fetch_bars 真故障抛 RuntimeError 时按"兜底不可用"处理，维持原 TDX 错误。
+"""
+
+from __future__ import annotations
+
+import sys
+import types
+
+import pandas as pd
+import pytest
+
+pytest.importorskip("fastapi")
+
+from fastapi.testclient import TestClient  # noqa: E402
+
+# ── 测试替身 ─────────────────────────────────────────────────────────────────
+
+
+class _RaisingMac:
+    async def get_stock_kline(self, *args, **kwargs):  # noqa: ANN002, ANN003
+        raise RuntimeError("MAC 连接失败")
+
+
+class _RaisingTdx:
+    async def get_security_bars(self, *args, **kwargs):  # noqa: ANN002, ANN003
+        raise RuntimeError("标准协议连接失败")
+
+    async def get_index_bars(self, *args, **kwargs):  # noqa: ANN002, ANN003
+        raise RuntimeError("标准协议连接失败")
+
+
+def _bars_app(mac_client, tdx_client):
+    from fastapi import FastAPI
+
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import bars
+
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(bars.router, prefix="/api/v1")
+    app.state.tdx_client = tdx_client
+    app.state.mac_client = mac_client
+    return app
+
+
+def _fallback_df(n: int = 5) -> pd.DataFrame:
+    dates = pd.bdate_range(end="2026-09-04", periods=n)
+    return pd.DataFrame(
+        {
+            "date": dates.normalize(),
+            "open": [10.0] * n,
+            "close": [10.5] * n,
+            "high": [11.0] * n,
+            "low": [9.5] * n,
+            "vol": [100000.0] * n,
+            "amount": [1050000.0] * n,
+        }
+    )
+
+
+def _install_fake_bs_module(monkeypatch: pytest.MonkeyPatch, rows: int = 10) -> dict:
+    """装一个最小可用的 baostock 模块替身，返回 captured 观测点。"""
+    from easy_tdx.sources import baostock as bs_source
+
+    captured: dict = {}
+
+    def _login():
+        lg = types.SimpleNamespace()
+        lg.error_code = "0"
+        lg.error_msg = "ok"
+        return lg
+
+    def query_history_k_data_plus(**kwargs):  # noqa: ANN003
+        captured.update(kwargs)
+        captured["calls"] = captured.get("calls", 0) + 1
+        data = [
+            [f"2026-08-{d:02d}", "10.0", "10.5", "11.0", "9.5", "100000", "1050000", "1"]
+            for d in range(1, rows + 1)
+        ]
+        rs = types.SimpleNamespace()
+        rs.error_code = "0"
+        rs.error_msg = "ok"
+        rs._rows = data
+        rs._i = 0
+
+        rs.next = lambda: rs._i < len(rs._rows)  # type: ignore[method-assign]
+        rs.get_row_data = lambda: rs._rows[rs._i]  # type: ignore[method-assign]
+
+        def _advance():
+            row = rs._rows[rs._i]
+            rs._i += 1
+            return row
+
+        rs.get_row_data = _advance  # type: ignore[method-assign]
+        return rs
+
+    mod = types.ModuleType("baostock")
+    mod.login = _login  # type: ignore[attr-defined]
+    mod.logout = lambda: None  # type: ignore[attr-defined]
+    mod.query_history_k_data_plus = query_history_k_data_plus  # type: ignore[attr-defined]
+
+    monkeypatch.setitem(sys.modules, "baostock", mod)
+    monkeypatch.delenv("EASY_TDX_BAOSTOCK", raising=False)
+    monkeypatch.setattr(bs_source, "_logged_in", False)
+    return captured
+
+
+# ── 项11：数字周期字符串归一后再兜底 ──────────────────────────────────────────
+
+
+def test_bars_numeric_category_still_falls_back(monkeypatch):
+    """category="4"（=DAY 的数字形式）TDX 全败时也应命中 baostock 兜底。
+
+    旧实现把原串 "4" 透传给 fetch_bars，_FREQ_BY_CATEGORY.get("4") 落空
+    返回 None → 兜底静默失效，客户端拿到 500。
+    """
+    captured = _install_fake_bs_module(monkeypatch)
+    with TestClient(
+        _bars_app(_RaisingMac(), _RaisingTdx()), raise_server_exceptions=False
+    ) as client:
+        resp = client.get(
+            "/api/v1/bars", params={"market": "SH", "code": "600519", "category": "4"}
+        )
+    assert resp.status_code == 200, resp.text
+    body = resp.json()
+    assert body["source"] == "baostock"
+    assert body["count"] > 0
+    assert captured["frequency"] == "d"  # 归一成 DAY 后映射到日线
+
+
+# ── 项12：is_index 传递与异常语义 ────────────────────────────────────────────
+
+
+def test_index_fallback_passes_is_index_true(monkeypatch):
+    """/bars/index 兜底必须带 is_index=True（指数 vol 股→手 ÷100）。"""
+    from easy_tdx.sources import baostock as bs_source
+
+    calls: dict = {}
+
+    def fake_fetch(market, code, category, start, count, adjust, is_index=False):
+        calls["is_index"] = is_index
+        return _fallback_df()
+
+    monkeypatch.setattr(bs_source, "is_enabled", lambda: True)
+    monkeypatch.setattr(bs_source, "fetch_bars", fake_fetch)
+
+    with TestClient(_bars_app(None, _RaisingTdx())) as client:
+        resp = client.get(
+            "/api/v1/bars/index", params={"market": "SH", "code": "000001", "category": "DAY"}
+        )
+    assert resp.status_code == 200
+    assert resp.json()["source"] == "baostock"
+    assert calls["is_index"] is True
+
+
+def test_bars_stock_fallback_keeps_is_index_false(monkeypatch):
+    """个股路径兜底 is_index=False（vol 保持股口径）。"""
+    from easy_tdx.sources import baostock as bs_source
+
+    calls: dict = {}
+
+    def fake_fetch(market, code, category, start, count, adjust, is_index=False):
+        calls["is_index"] = is_index
+        return _fallback_df()
+
+    monkeypatch.setattr(bs_source, "is_enabled", lambda: True)
+    monkeypatch.setattr(bs_source, "fetch_bars", fake_fetch)
+
+    with TestClient(_bars_app(_RaisingMac(), _RaisingTdx())) as client:
+        resp = client.get("/api/v1/bars", params={"market": "SH", "code": "600519"})
+    assert resp.status_code == 200
+    assert calls["is_index"] is False
+
+
+def test_bars_fallback_exception_keeps_original_tdx_error(monkeypatch):
+    """fetch_bars 真故障抛 RuntimeError → 按"兜底不可用"处理，重抛原 TDX 异常。
+
+    响应错误详情须是标准协议的失败原因，而非 baostock 的失败原因（baostock
+    的失败只记日志），且不返回空数据伪装成功。
+    """
+    from easy_tdx.sources import baostock as bs_source
+
+    def boom(*args, **kwargs):  # noqa: ANN002, ANN003
+        raise RuntimeError("baostock 拉取失败: 网络异常")
+
+    monkeypatch.setattr(bs_source, "is_enabled", lambda: True)
+    monkeypatch.setattr(bs_source, "fetch_bars", boom)
+
+    with TestClient(
+        _bars_app(_RaisingMac(), _RaisingTdx()), raise_server_exceptions=False
+    ) as client:
+        resp = client.get("/api/v1/bars", params={"market": "SH", "code": "600519"})
+    assert resp.status_code == 500
+    assert "标准协议连接失败" in resp.json()["detail"]
+    assert "baostock" not in resp.json()["detail"]
+
+
+def test_index_fallback_exception_keeps_original_tdx_error(monkeypatch):
+    """/bars/index 同语义：baostock 异常不吞掉原 TDX 错误。"""
+    from easy_tdx.sources import baostock as bs_source
+
+    def boom(*args, **kwargs):  # noqa: ANN002, ANN003
+        raise RuntimeError("baostock 拉取失败: 网络异常")
+
+    monkeypatch.setattr(bs_source, "is_enabled", lambda: True)
+    monkeypatch.setattr(bs_source, "fetch_bars", boom)
+
+    with TestClient(_bars_app(None, _RaisingTdx()), raise_server_exceptions=False) as client:
+        resp = client.get(
+            "/api/v1/bars/index", params={"market": "SH", "code": "000001", "category": "DAY"}
+        )
+    assert resp.status_code == 500
+    assert "标准协议连接失败" in resp.json()["detail"]

tests/unit/test_web_bars_paging.py

@@ -0,0 +1,172 @@
+"""count>800 的分页取数测试（离线）。
+
+TDX 协议单次 get_security_bars 最多返回 800 根：旧实现里 multiseed /
+rotation / formula 的单次调用在 count>800 时被服务器静默截断。本文件钉死
+"分页取全量 + 页序正确 + 数据起点提前停止"行为。
+"""
+
+from __future__ import annotations
+
+import asyncio
+
+import pandas as pd
+import pytest
+
+pytest.importorskip("fastapi")
+
+_PAGE_CAP = 800
+
+
+class _CappedBarsClient:
+    """模拟 TDX 服务器：单次最多返回 _PAGE_CAP 根，start 为回看偏移。"""
+
+    def __init__(self, total_bars: int = 3000):
+        self.total_bars = total_bars
+        self.calls: list[tuple[int, int]] = []  # (start, count)
+
+    def _make_page(self, start: int, n: int) -> pd.DataFrame:
+        """start 偏移处往前 n 根（升序页）；越过数据起点则截断为 0 根。"""
+        hi = self.total_bars - start  # 本页最旧一根的全局序号（0 起）
+        lo = max(0, hi - n)
+        if hi <= 0:
+            return pd.DataFrame()
+        dates = pd.date_range("2020-01-01", periods=self.total_bars, freq="B")
+        idx = dates[lo:hi]
+        return pd.DataFrame(
+            {
+                "date": idx,
+                "open": 10.0,
+                "high": 11.0,
+                "low": 9.0,
+                "close": 10.5,
+                "vol": 1000.0,
+                "amount": 10000.0,
+            }
+        )
+
+    async def get_security_bars(self, market, code, category, start, count, **kw):
+        self.calls.append((int(start), int(count)))
+        return self._make_page(int(start), int(count))
+
+
+# ── 共享分页辅助 ───────────────────────────────────────────────────────────────
+
+
+def test_fetch_bars_paged_requests_multiple_pages():
+    """count=2000 → 3 次请求（800/800/400），拼齐 2000 根且时间升序。"""
+    from easy_tdx.web.routers.backtest import _fetch_bars_paged
+
+    fake = _CappedBarsClient(total_bars=3000)
+    df = asyncio.run(_fetch_bars_paged(fake, "SZ:000001", "DAY", 2000))
+
+    assert fake.calls == [(0, 800), (800, 800), (1600, 400)]
+    assert len(df) == 2000
+    dates = pd.to_datetime(df["date"])
+    assert dates.is_monotonic_increasing  # 页序拼接后必须升序
+
+
+def test_fetch_bars_paged_stops_at_data_start():
+    """数据起点不足一页时提前停止，不多发请求。"""
+    from easy_tdx.web.routers.backtest import _fetch_bars_paged
+
+    fake = _CappedBarsClient(total_bars=1000)
+    df = asyncio.run(_fetch_bars_paged(fake, "SH:600519", "DAY", 2000))
+
+    # 第二页只回 200 根（不足一页）= 数据起点，循环不再发第三笔请求
+    assert fake.calls == [(0, 800), (800, 800)]
+    assert len(df) == 1000
+
+
+def test_fetch_bars_paged_small_count_single_call():
+    """count≤800 仍单页取齐（不多打请求）。"""
+    from easy_tdx.web.routers.backtest import _fetch_bars_paged
+
+    fake = _CappedBarsClient(total_bars=3000)
+    df = asyncio.run(_fetch_bars_paged(fake, "SZ:000001", "DAY", 250))
+    assert fake.calls == [(0, 250)]
+    assert len(df) == 250
+
+
+def test_fetch_bars_paged_empty_returns_empty_df():
+    from easy_tdx.web.routers.backtest import _fetch_bars_paged
+
+    fake = _CappedBarsClient(total_bars=0)
+    df = asyncio.run(_fetch_bars_paged(fake, "SZ:000001", "DAY", 800))
+    assert df.empty
+
+
+# ── multiseed / rotation 端点（取数在 handler 内完成，POST 返回即可断言）──────
+
+
+def _app_with(fake_client):
+    from fastapi import FastAPI
+
+    from easy_tdx.web.errors import register_exception_handlers
+    from easy_tdx.web.routers import backtest as backtest_mod
+
+    app = FastAPI()
+    register_exception_handlers(app)
+    app.include_router(backtest_mod.router, prefix="/api/v1")
+    app.state.tdx_client = fake_client
+    app.state.mac_client = None
+    app.state.ex_client = None
+    return app
+
+
+def test_multiseed_fetches_full_count_via_paging():
+    """multiseed count=900（>800）→ 每标的 2 次请求，不再被 800 截断。"""
+    pytest.importorskip("fastapi")
+    from fastapi.testclient import TestClient
+
+    fake = _CappedBarsClient(total_bars=3000)
+    with TestClient(_app_with(fake)) as client:
+        resp = client.post(
+            "/api/v1/backtest/multiseed/run/async",
+            json={
+                "strategy": "ma_cross",
+                "params": {"fast": 3, "slow": 6},
+                "stocks": ["SZ:000001", "SH:600519"],
+                "count": 900,
+            },
+        )
+        assert resp.status_code == 202, resp.text
+
+    # 2 标的 × 2 页
+    assert fake.calls == [(0, 800), (800, 100), (0, 800), (800, 100)]
+
+
+def test_rotation_fetches_full_count_via_paging():
+    """rotation count=900（>800）→ 每标的 2 次请求。"""
+    pytest.importorskip("fastapi")
+    from fastapi.testclient import TestClient
+
+    fake = _CappedBarsClient(total_bars=3000)
+    with TestClient(_app_with(fake)) as client:
+        resp = client.post(
+            "/api/v1/backtest/rotation/run/async",
+            json={
+                "stocks": ["SZ:000001", "SH:600519"],
+                "count": 900,
+            },
+        )
+        assert resp.status_code == 202, resp.text
+
+    assert fake.calls == [(0, 800), (800, 100), (0, 800), (800, 100)]
+
+
+# ── formula 取数路径 ──────────────────────────────────────────────────────────
+
+
+def test_formula_resolve_df_pages_full_count():
+    """formula _resolve_df symbol 路径 count=2000 → 3 页拼齐且升序。"""
+    from easy_tdx.web.routers.backtest import _fetch_bars_paged  # noqa: F401  需已存在
+    from easy_tdx.web.routers.formula import FormulaComputeRequest, _resolve_df
+
+    fake = _CappedBarsClient(total_bars=3000)
+    df = asyncio.run(
+        _resolve_df(fake, FormulaComputeRequest(text="C", symbol="SZ:000001", count=2000))
+    )
+    assert fake.calls == [(0, 800), (800, 800), (1600, 400)]
+    assert len(df) == 2000
+    dates = pd.to_datetime(df["date"])
+    assert dates.is_monotonic_increasing

web-ui/src/__tests__/api.test.ts

@@ -0,0 +1,108 @@
+/**
+ * api.ts 行为自检（Node 内置 test runner，fetch 打桩，无 DOM 依赖）。
+ *
+ * 运行：node --test src/__tests__/api.test.ts
+ *
+ * 覆盖：
+ * 1. fetchRankList 的 market 归一化——北交所(market=2)必须映射为 'BJ'
+ *    （与 fetchBoardMembers 同口径；错标 'SZ' 会导致榜单点开个股用错市场拉行情）。
+ * 2. runLlmChatWithPolling 的可选 AbortSignal——signal 中止后循环立即以
+ *    AbortError 退出，不再继续轮询（弹窗卸载清理依赖此行为）。
+ */
+
+import { test } from 'node:test'
+import assert from 'node:assert/strict'
+
+import { fetchRankList, runLlmChatWithPolling } from '../api.ts'
+import type { TaskState } from '../types.ts'
+
+/** 临时替换 globalThis.fetch，返回可编程响应序列。 */
+function stubFetch(handler: (url: string, init?: RequestInit) => unknown): {
+  calls: string[]
+  restore: () => void
+} {
+  const calls: string[] = []
+  const original = globalThis.fetch
+  globalThis.fetch = (async (input: RequestInfo | URL, init?: RequestInit) => {
+    const url = String(input instanceof URL ? input : input)
+    calls.push(url)
+    const body = handler(url, init)
+    return new Response(JSON.stringify(body), {
+      status: 200,
+      headers: { 'Content-Type': 'application/json' },
+    })
+  }) as typeof fetch
+  return {
+    calls,
+    restore: () => {
+      globalThis.fetch = original
+    },
+  }
+}
+
+// ── 1. fetchRankList 北交所映射 ───────────────────────────────────────────────
+
+test('fetchRankList: market=2（北交所）归一化为 BJ', async () => {
+  const stub = stubFetch(() => ({
+    data: [
+      { close: 10.5, pre_close: 9.55, market: 2, code: '920002', name: '测试北交所' },
+      { close: 8.0, pre_close: 8.8, market: 0, code: '000001', name: '平安银行' },
+      { close: 20.0, pre_close: 19.0, market: 1, code: '600519', name: '贵州茅台' },
+    ],
+  }))
+  try {
+    const rows = await fetchRankList('DESC', 3)
+    assert.equal(rows[0].market, 'BJ', 'market=2 应映射为 BJ（旧码错标 SZ）')
+    assert.equal(rows[1].market, 'SZ', 'market=0 应映射为 SZ')
+    assert.equal(rows[2].market, 'SH', 'market=1 应映射为 SH')
+    // 顺带核对涨跌幅口径：close/pre-1
+    assert.ok(Math.abs((rows[0].change_pct ?? 0) - ((10.5 / 9.55 - 1) * 100)) < 1e-6)
+  } finally {
+    stub.restore()
+  }
+})
+
+// ── 2. runLlmChatWithPolling 的 AbortSignal ──────────────────────────────────
+
+const runningState: TaskState = { status: 'running' } as unknown as TaskState
+
+test('runLlmChatWithPolling: signal 中止后立即退出并抛 AbortError', async () => {
+  const stub = stubFetch((url) => {
+    if (url.includes('/llm/chat/async')) return { task_id: 't1' }
+    return runningState
+  })
+  try {
+    const ctrl = new AbortController()
+    const promise = runLlmChatWithPolling(
+      'ping',
+      null,
+      () => ctrl.abort(), // 第一次轮询即中止
+      5, // intervalMs
+      30_000, // timeoutMs（远大于中止所需时间，确保超时兜底不先触发）
+      ctrl.signal,
+    )
+    await assert.rejects(promise, (e: unknown) => (e as Error).name === 'AbortError')
+    // 中止后不应继续轮询：任务查询次数应极少（≤3）
+    const polls = stub.calls.filter((u) => u.includes('/llm/chat/tasks/')).length
+    assert.ok(polls <= 3, `中止后应停止轮询，实际轮询 ${polls} 次`)
+  } finally {
+    stub.restore()
+  }
+})
+
+test('runLlmChatWithPolling: 未中止时正常返回 done', async () => {
+  let polls = 0
+  const stub = stubFetch((url) => {
+    if (url.includes('/llm/chat/async')) return { task_id: 't2' }
+    polls += 1
+    if (polls < 3) return runningState
+    return { status: 'done', result: { reply: 'ok', model: 'm', provider: 'p' } }
+  })
+  try {
+    const state = await runLlmChatWithPolling('ping', null, undefined, 1, 30_000)
+    assert.equal(state.status, 'done')
+    assert.equal((state.result as { reply?: string } | null)?.reply, 'ok')
+  } finally {
+    stub.restore()
+  }
+})

web-ui/src/api.ts

@@ -646,7 +646,9 @@ export async function fetchRankList(
     const r = { ...row } as Record<string, unknown>
     r.price = close
     r.change_pct = pre > 0 ? (close / pre - 1) * 100 : 0
-    r.market = Number(row.market) === 1 ? 'SH' : 'SZ'
+    // 与 fetchBoardMembers 同口径：MAC 协议 market 1=SH / 2=BJ / 其余 SZ
+    const m = Number(row.market)
+    r.market = m === 1 ? 'SH' : m === 2 ? 'BJ' : 'SZ'
     return r as RankRow
   })
 }
@@ -799,6 +801,9 @@ export async function fetchLlmChatTask(taskId: string): Promise<TaskState> {
  *
  * 大报告解读 1-3 分钟属正常：轮询间隔放宽到 1.5s（回测是 0.3s），
  * 前端上限 20 分钟兜底（后端 LLM 读超时最大 600s，正常应先于此前返回）。
+ *
+ * @param signal 可选中止信号：弹窗等调用方卸载时 abort，循环立即以
+ *               name='AbortError' 的错误退出，不再继续轮询。
  */
 export async function runLlmChatWithPolling(
   prompt: string,
@@ -806,11 +811,13 @@ export async function runLlmChatWithPolling(
   onPoll?: (state: TaskState) => void,
   intervalMs = 1_500,
   timeoutMs = 20 * 60_000,
+  signal?: AbortSignal,
 ): Promise<TaskState> {
   const { task_id } = await submitLlmChatTask(prompt, context)
   const start = Date.now()
   // eslint-disable-next-line no-constant-condition
   while (true) {
+    if (signal?.aborted) throw abortError()
     const state = await fetchLlmChatTask(task_id)
     onPoll?.(state)
     if (state.status === 'done' || state.status === 'failed') return state
@@ -818,9 +825,15 @@ export async function runLlmChatWithPolling(
       throw new Error(`AI 解读任务超时（${timeoutMs / 1000}s），任务仍在后台运行，可稍后重试`)
     }
     await new Promise((r) => setTimeout(r, intervalMs))
+    if (signal?.aborted) throw abortError()
   }
 }
 
+/** 构造与 fetch 中止一致的 AbortError（便于调用方按 name 识别并静默）。 */
+function abortError(): Error {
+  return new DOMException('AI 解读已取消', 'AbortError')
+}
+
 // ── AI 解读历史 ──────────────────────────────────────────────────────────────
 
 /** 列出 AI 解读历史（时间倒序，含 Prompt/正文/策略上下文）。 */

web-ui/src/components/AiInterpretModal.vue

@@ -2,7 +2,7 @@
 // AI 解读弹窗（单标的/组合回测通用）：Prompt 预览 + 复制/下载 + 一键直接解读。
 // Prompt 由父组件实时组装传入（附加分析跑完内容自动变全），本组件只管交互；
 // 直接解读走后端 LLM 后台任务（配置见「AI 设置」页），解读记录旁路落历史库。
-import { onMounted, ref, watch } from 'vue'
+import { onBeforeUnmount, onMounted, ref, watch } from 'vue'
 
 import { formatError, fetchLlmConfig, runLlmChatWithPolling } from '../api'
 import type { LlmChatContext, LlmChatResult } from '../types'
@@ -28,6 +28,8 @@ const aiRunning = ref(false)
 const aiElapsed = ref(0)
 const aiReply = ref('')
 let aiTimer = 0
+/** 卸载时中止后台轮询（runLlmChatWithPolling 检查 signal 立即退出）。 */
+let abortCtrl: AbortController | null = null
 
 onMounted(() => {
   // 打开时探测 LLM 是否已配置（失败静默——导出 Prompt 的老路径不依赖后端）
@@ -42,6 +44,15 @@ onMounted(() => {
     })
 })
 
+onBeforeUnmount(() => {
+  if (aiTimer) {
+    window.clearInterval(aiTimer)
+    aiTimer = 0
+  }
+  abortCtrl?.abort()
+  abortCtrl = null
+})
+
 watch(
   () => props.prompt,
   () => {
@@ -61,8 +72,16 @@ async function runAiInterpret() {
   aiTimer = window.setInterval(() => {
     aiElapsed.value += 1
   }, 1000)
+  abortCtrl = new AbortController()
   try {
-    const state = await runLlmChatWithPolling(props.prompt, props.context)
+    const state = await runLlmChatWithPolling(
+      props.prompt,
+      props.context,
+      undefined,
+      undefined,
+      undefined,
+      abortCtrl.signal,
+    )
     // TaskState.result 是多任务类型联合，按 LLM 任务结构收窄
     const r = state.result as LlmChatResult | null
     // 后端已保证非空正文（空白正文会以 failed 上浮），前端再拦一道纯空白
@@ -75,9 +94,13 @@ async function runAiInterpret() {
       aiMsg.value = `解读失败：${state.error ?? '未知错误'}（可在「AI 设置」检查配置，或复制 Prompt 手动使用）`
     }
   } catch (e) {
+    // 卸载触发的取消不打扰（组件已不在 DOM）
+    if ((e as Error)?.name === 'AbortError') return
     aiMsg.value = `解读失败：${formatError(e)}（可在「AI 设置」检查配置，或复制 Prompt 手动使用）`
   } finally {
     window.clearInterval(aiTimer)
+    aiTimer = 0
+    abortCtrl = null
     aiRunning.value = false
   }
 }

web-ui/src/components/BoardDialog.vue

@@ -139,16 +139,24 @@ const membersLoading = ref(false)
 // 后端单页 80 自动翻页，1000 覆盖最大概念板块，拉满后按成员清单自然终止。
 const MEMBER_FETCH_COUNT = 1000
 
+/** 请求序号守卫：升/降序快速连点、或 props.code 切换时，旧响应直接丢弃。 */
+let memberSeq = 0
+
 async function loadMembers() {
+  const my = ++memberSeq
   membersLoading.value = true
   membersError.value = ''
   try {
-    members.value = await fetchBoardMembers(props.code, MEMBER_FETCH_COUNT, memberOrder.value)
+    const rows = await fetchBoardMembers(props.code, MEMBER_FETCH_COUNT, memberOrder.value)
+    // 响应已过期（顺序已再切换 / 板块已切换）：不得覆盖新请求的结果
+    if (my !== memberSeq) return
+    members.value = rows
   } catch (e) {
+    if (my !== memberSeq) return
     members.value = []
     membersError.value = formatError(e)
   } finally {
-    membersLoading.value = false
+    if (my === memberSeq) membersLoading.value = false
   }
 }
 

web-ui/src/components/HotspotCorrelation.vue

@@ -18,6 +18,8 @@ const error = ref('')
 const loading = ref(false)
 
 let poll: number | null = null
+/** 请求序号守卫：await 后比对，过期响应（旧类型/旧窗口）直接丢弃。 */
+let loadSeq = 0
 
 function stopPoll() {
   if (poll !== null) {
@@ -27,9 +29,12 @@ function stopPoll() {
 }
 
 async function load() {
+  const my = ++loadSeq
   error.value = ''
   try {
     const r = await fetchHotspotCorrelation(props.boardType, props.days, props.perDay)
+    // 响应已过期：不覆盖新参数的 resp、不重启构建轮询
+    if (my !== loadSeq) return
     if (r.status === 'building') {
       resp.value = null
       loading.value = true
@@ -45,6 +50,7 @@ async function load() {
     resp.value = r
     render()
   } catch (e) {
+    if (my !== loadSeq) return
     stopPoll()
     loading.value = false
     error.value = formatError(e)

web-ui/src/components/Sparkline.vue

@@ -35,7 +35,9 @@ const geom = computed(() => {
 </script>
 
 <template>
-  <svg :width="W" :height="H" class="sparkline" viewBox="0 0 92 28" preserveAspectRatio="none">
+  <!-- viewBox 必须跟随 width/height：几何坐标按 W/H 计算，写死 92×28 会让
+       传大尺寸的调用方（Dashboard 150×30）右侧被 viewport 裁剪 -->
+  <svg :width="W" :height="H" class="sparkline" :viewBox="`0 0 ${W} ${H}`" preserveAspectRatio="none">
     <template v-if="geom">
       <line
         v-if="geom.baseValid"
@@ -56,7 +58,7 @@ const geom = computed(() => {
         vector-effect="non-scaling-stroke"
       />
     </template>
-    <text v-else x="46" y="18" text-anchor="middle" fill="#5c6370" font-size="9">加载中…</text>
+    <text v-else :x="W / 2" :y="H - 10" text-anchor="middle" fill="#5c6370" font-size="9">加载中…</text>
   </svg>
 </template>
 

web-ui/src/grading/combinedMetrics.ts

@@ -12,7 +12,7 @@
  *           夏普/波动率按 √252 年化。如果是周线/月线，年化因子需要调整。
  */
 
-import type { EquityPoint } from '../types'
+import type { EquityPoint } from '../types.ts'
 
 /** 年化因子（按交易日）。 */
 const TRADING_DAYS_PER_YEAR = 252

web-ui/src/grading/engine.ts

@@ -4,8 +4,8 @@
  * 这一层是纯函数 + 零业务依赖，所有场景（单标的/组合/寻优）共用。
  */
 
-import { GRADE_THRESHOLDS, type DimensionScore, type Grade, type GradeResult, type VetoHit } from './types'
-import { THRESHOLDS, type DimensionKey } from './thresholds'
+import { GRADE_THRESHOLDS, type DimensionScore, type Grade, type GradeResult, type VetoHit } from './types.ts'
+import { THRESHOLDS, type DimensionKey } from './thresholds.ts'
 
 /**
  * 按锚点列表做线性插值，返回 0–100 的分数。

web-ui/src/grading/index.ts

@@ -12,10 +12,10 @@
  * @see docs/superpowers/plans 评级系统设计文档
  */
 
-import type { BacktestResult, EquityPoint, GridPointResult, Performance, PortfolioResult } from '../types'
-import { buildResult, scoreDimension } from './engine'
-import { computeCombinedMetrics } from './combinedMetrics'
-import type { DimensionScore, GradeResult, VetoHit } from './types'
+import type { BacktestResult, EquityPoint, GridPointResult, Performance, PortfolioResult } from '../types.ts'
+import { buildResult, scoreDimension } from './engine.ts'
+import { computeCombinedMetrics } from './combinedMetrics.ts'
+import type { DimensionScore, GradeResult, VetoHit } from './types.ts'
 
 // ════════════════════════════════════════════════════════════════════════════
 // 一票否决规则（所有场景共用）
@@ -315,8 +315,10 @@ export function gradeBacktestResult(result: BacktestResult): GradeResult {
 }
 
 // ── 重新导出常用类型和工具，便于调用方一处 import ───────────────────────────
-export { GRADE_META, GRADE_THRESHOLDS } from './types'
-export type { Grade, GradeResult, DimensionScore, VetoHit, GradeMeta } from './types'
-export { worseGrade, scoreToGrade } from './engine'
-export { computeCombinedMetrics } from './combinedMetrics'
-export type { CombinedMetrics } from './combinedMetrics'
+// 注：相对路径统一带 .ts 扩展——Node --test 直跑（type-stripping 不改写
+// import 说明符）与 Vite/vue-tsc（allowImportingTsExtensions）两侧都可用。
+export { GRADE_META, GRADE_THRESHOLDS } from './types.ts'
+export type { Grade, GradeResult, DimensionScore, VetoHit, GradeMeta } from './types.ts'
+export { worseGrade, scoreToGrade } from './engine.ts'
+export { computeCombinedMetrics } from './combinedMetrics.ts'
+export type { CombinedMetrics } from './combinedMetrics.ts'

web-ui/src/views/BoardOverviewView.vue

@@ -37,18 +37,26 @@ const error = ref('')
 const lastRefresh = ref('')
 const stat = ref<MarketStat | null>(null)
 
+/** 请求序号守卫：await 后比对，过期响应直接丢弃。 */
+let overviewSeq = 0
+
 async function loadOverview() {
+  const my = ++overviewSeq
   loading.value = rows.value.length === 0
   error.value = ''
   try {
     const resp = await fetchBoardOverview(activeType.value)
+    // 响应已过期（期间切换了一级/二级）：不覆盖新类型的 rows，
+    // 也不进入 diffFlips 产生幽灵翻红/翻绿事件
+    if (my !== overviewSeq) return
     rows.value = resp.rows
     lastRefresh.value = new Date().toLocaleTimeString('zh-CN', { hour12: false })
     diffFlips(resp.rows)
   } catch (e) {
+    if (my !== overviewSeq) return
     error.value = formatError(e)
   } finally {
-    loading.value = false
+    if (my === overviewSeq) loading.value = false
   }
 }
 

web-ui/src/views/DashboardView.vue

@@ -265,10 +265,11 @@ function distHeight(count: number): string {
 }
 
 function distColor(i: number): string {
-  // 桶 1..20 对应 -10..+9：前 10 绿（跌），后 10 红（涨）；两端按方向
+  // 桶 1..20 对应 -10..+9：1..10 绿（跌），11（0 轴）中性灰，12..20 红（涨）；两端按方向
   if (i === 0) return 'var(--down)'
   if (i === BUCKETS.length - 1) return 'var(--up)'
-  return i <= 10 ? 'var(--down)' : 'var(--up)'
+  if (i === 11) return 'var(--text-dim)'
+  return i < 11 ? 'var(--down)' : 'var(--up)'
 }
 
 // ── 板块热度 + 冰冷（一次拉 120 个，前端切热/冷两端） ───────────────────────
@@ -454,7 +455,8 @@ const boardDialog = ref<{ code: string; name: string } | null>(null)
 
 function openDialog(code: string, name: string, marketHint?: string) {
   if (!code) return
-  const mkt = marketHint ?? (/^(6|9|5)/.test(code) ? 'SH' : /^(4|8|92|43)/.test(code) ? 'BJ' : 'SZ')
+  // 先判北交所再判沪市：920xxx（北交所新段）以 9 开头，若先匹配 9 会被误判 SH
+  const mkt = marketHint ?? (/^(4|8|92|43)/.test(code) ? 'BJ' : /^[659]/.test(code) ? 'SH' : 'SZ')
   dialog.value = { market: mkt, code, name }
 }
 

web-ui/src/views/HotspotView.vue

@@ -26,7 +26,10 @@ const viewMode = ref<'matrix' | 'corr'>('matrix')
 watch(
   () => props.boardType,
   (t) => {
-    if (t && t !== boardType.value) setType(t)
+    // /styles 路由注入 FG；/hotspots 无 props（undefined）——复用组件实例时
+    // 需回退默认 HY，否则 FG 数据滞留在热点滚动路由下
+    const target = t ?? 'HY'
+    if (target !== boardType.value) setType(target)
   },
 )
 
@@ -48,6 +51,8 @@ const loading = ref(false)
 const lastRefresh = ref('')
 
 let buildTimer = 0
+/** 请求序号守卫：await 后比对，过期响应直接丢弃（不覆盖新类型/新参数的状态）。 */
+let loadSeq = 0
 
 function stopBuildPoll() {
   if (buildTimer) {
@@ -57,8 +62,12 @@ function stopBuildPoll() {
 }
 
 async function load(retry = false) {
+  const my = ++loadSeq
   try {
     const r = await fetchBoardHotspot(boardType.value, days.value, mode.value, PER_DAY, retry)
+    // 响应已过期（期间用户切换了类型/天数/模式）：不得覆盖 resp、不得
+    // 杀死或重启新参数的构建轮询，整体丢弃。
+    if (my !== loadSeq) return
     if (r.status === 'building') {
       buildError.value = ''
       resp.value = null
@@ -78,6 +87,7 @@ async function load(retry = false) {
     loading.value = false
     lastRefresh.value = new Date().toLocaleTimeString('zh-CN', { hour12: false })
   } catch (e) {
+    if (my !== loadSeq) return
     stopBuildPoll()
     buildingProgress.value = null
     buildError.value = formatError(e)

web-ui/src/views/IndexCalendarView.vue

@@ -24,8 +24,11 @@ const bars = ref<Bar[]>([])
 const lastUpdate = ref('')
 
 async function loadIndex(idx: number) {
-  if (barsByIndex.has(idx)) {
-    bars.value = barsByIndex.get(idx)!
+  const cached = barsByIndex.get(idx)
+  if (cached) {
+    bars.value = cached
+    error.value = '' // 其他指数的失败/空数据提示不带到已缓存的指数上
+    loading.value = false
     return
   }
   loading.value = true
@@ -33,9 +36,13 @@ async function loadIndex(idx: number) {
   try {
     const meta = INDICES[idx]
     const data = await fetchIndexBars(meta.market, meta.code, 550) // ≈2.2 年
+    if (data.length === 0) {
+      // 空数据不入缓存：否则"重试"命中缓存直接 return，永远无法重新请求
+      error.value = `${meta.name} 日K返回空`
+      return
+    }
     barsByIndex.set(idx, data)
     bars.value = data
-    if (data.length === 0) error.value = `${meta.name} 日K返回空`
     lastUpdate.value = new Date().toLocaleTimeString('zh-CN', { hour12: false })
   } catch (e) {
     error.value = formatError(e)

web-ui/src/views/SentimentView.vue

@@ -230,7 +230,8 @@ async function loadVolume() {
       fetchBars('SH', '000001', 'MIN_5', start),
       fetchBars('SZ', '399001', 'MIN_5', start),
     ])
-    // 按日期聚合两市场 5 分钟 amount（元）
+    // 按日期聚合上证指数+深证成指的 5 分钟 amount（元；成指成分口径，
+    // 非深市全市场——仅用于与自身近 5 日同期均值做相对比较）
     const byDate = new Map<string, Map<string, number>>()
     for (const b of [...sh, ...sz]) {
       const d = b.datetime.slice(0, 10)
@@ -343,7 +344,7 @@ async function buildDigest(): Promise<string> {
     )
   }
   if (volRatio.value !== null) {
-    lines.push(`量能：当日两市累计成交较近 5 日同期均值 ${fmtPctSigned(volRatio.value)}。`)
+    lines.push(`量能：沪深指数（上证指数+深证成指）当日成交额较近 5 日同期均值 ${fmtPctSigned(volRatio.value)}。`)
   }
   try {
     const eco = await fetchLimitUpEcology()
@@ -471,8 +472,9 @@ onBeforeUnmount(() => {
       <!-- 量能仪表盘 -->
       <div class="section">
         <div class="sec-title">
-          量能 · 两市累计成交额（最近交易日{{ volDate ? ` ${volDate.slice(5)}` : '' }} vs 近 5 日同期均值
+          量能 · 沪深指数成交额（上证指数 + 深证成指 MIN_5 合计，最近交易日{{ volDate ? ` ${volDate.slice(5)}` : '' }} vs 近 5 日同期均值
           <span v-if="volRatio !== null" :class="volRatio > 0 ? 'up' : 'down'">{{ fmtPctSigned(volRatio) }}</span>）
+          <span class="dim">· 非全市场口径，全市场总成交见顶部「今日总成交」</span>
         </div>
         <div class="card chart-card">
           <div ref="volEl" class="chart"></div>
@@ -486,7 +488,10 @@ onBeforeUnmount(() => {
           <div v-for="d in fundDays" :key="d.date" class="fund-row">
             <span class="mono dim fund-date">{{ String(d.date).slice(4, 6) }}-{{ String(d.date).slice(6, 8) }}</span>
             <span v-for="b in d.boards" :key="b.code" class="fund-chip mono">
-              {{ b.name }} <span class="up">+{{ (b.main_net / 1e8).toFixed(1) }}亿</span>
+              {{ b.name }}
+              <span :class="b.main_net >= 0 ? 'up' : 'down'">
+                {{ b.main_net >= 0 ? '+' : '-' }}{{ (Math.abs(b.main_net) / 1e8).toFixed(1) }}亿
+              </span>
             </span>
           </div>
           <div v-if="fundDays.length === 0" class="empty-hint dim">

web-ui/src/views/WatchlistView.vue

@@ -68,15 +68,20 @@ async function fillMissingNames() {
 
 // ── 行情（SSE 快照 + REST 首次兜底） ────────────────────────────────────────
 
-/** SSE 未覆盖时（自选刚加、服务重启间隙）用 REST 主动拉一次。 */
+/** SSE 未覆盖时（自选刚加、服务重启间隙）用 REST 主动拉一次。
+ *  后端 /quotes 单次最多 80 只（通达信协议上限），超量需分批。 */
+const QUOTE_BATCH = 80
+
 async function restFallback() {
   if (items.value.length === 0) return
   const missing = items.value.filter((i) => !quoteStore.getQuote(i.symbol))
   if (missing.length === 0) return
-  try {
-    await fetchQuotes(missing.map((i) => ({ market: i.market, code: i.code })))
-  } catch {
-    // SSE 会补上，静默
+  for (let i = 0; i < missing.length; i += QUOTE_BATCH) {
+    try {
+      await fetchQuotes(missing.slice(i, i + QUOTE_BATCH).map((it) => ({ market: it.market, code: it.code })))
+    } catch {
+      // SSE 会补上，静默（单批失败不阻断后续批次）
+    }
   }
 }