v1.2.0：报错页面修复、数据刷到最新、更稳的构建

本次发布修复了多处页面报错与空图表，并把所有宏观指标刷新到当前最新。

🐛 问题修复

• 「财政与外需」页不再整页报错：此前该页 5 张图全部显示错误卡，现在正常展示财政收支、货物进出口、ISM 制造业景气等数据。
• 「债务周期」页「居民真实杠杆空间」不再空白：补齐派生数据后该图正常出线。
• CPI 再也不会停更：此前 CPI 因一处内部校验设置过严，长期停在旧值刷不动；现已修复，正常更新到最新月份。
• 社会融资规模补到最新：此前停在 4 月，现补齐到 7 月，与新增贷款、货币供应等同口径月份对齐。
• 缺数据的表不再拖垮整页：某张表暂时无数据时，页面显示「暂无数据」而非整页错误卡。

✨ 体验改进

• 全部 18 类宏观数据刷新到各自最新：月度指标到 2026 年 7 月，LPR 与 10 年期国债到 8 月，宏观杠杆率补到 2026 年二季度，季度与年度指标均为最新一期。
• 更稳的发布构建：修复了持续集成长期失败的问题，让每次发布都完整跑过自动化测试后再交付。

---

This release fixes several page errors and empty charts, and brings every macro indicator up to the latest available period.

🐛 Bug fixes

• The "Fiscal & External Demand" page no longer errors out: all five charts (fiscal revenue/expenditure, goods trade, ISM manufacturing, …) now render instead of showing an error card.
• The "Household deleveraging headroom" chart on the Debt Cycle page is no longer blank: it plots correctly after the derived data was rebuilt.
• CPI no longer gets stuck: an over-strict internal validation had frozen CPI at an old value; it now updates to the latest month.
• Social financing is current again: it was stuck at April and is now through July, aligned with new loans and money supply.
• A table with no data no longer breaks the whole page: it shows "No data" instead of a full-page error card.

✨ Improvements

• All 18 macro series refreshed to their latest period: monthlies through July 2026, LPR and the 10-year bond yield through August, macro leverage through 2026 Q2, quarterly and annual series at their newest release.
• More reliable release builds: fixed a long-standing CI failure so every release now ships only after passing the full automated test suite.

v1.1.0：相位纠错、本机安全、数据可信

本次更新集中修正经济周期判断的准确性、堵住本机安全隐患，并让数据失效不再被悄悄掩盖。

⚠️ 需要注意

1. 打开地址改为 http://localhost:8000

• 为什么改：过去前端和接口跑在两个进程、两个端口，经常出现「页面是旧的、接口是新的」，刷新也不生效。
• 新规则：双击「启动面板」后，界面和接口都在 8000 一个地址上，不再有 5173。
• 你要做什么：更新收藏夹。旧链接不会自动跳转。

旧：http://localhost:5173
新：http://localhost:8000

2. 手动触发「刷新数据」「重新生成评述」的方式变了

• 为什么改：过去只要你浏览的任意一个网页里藏了一张图片链接，就能在你不知情的情况下触发全量数据采集，甚至触发按次计费的模型调用。
• 新规则：这些操作现在需要本机凭据，页面内点击按钮一切照常，凭据由本机自动下发。
• 你要做什么：网页端无需任何操作。如果你写过脚本或收藏夹直接访问采集地址，需要改为先提交任务、再订阅进度。

🐛 问题修复

• 经济相位不再把 5% 的增长判成「衰退」：过去趋势基准被 2021 年的异常低基数长期污染，导致时钟长期停在衰退象限、综合信号被压到 -2。现在改用抗异常值的趋势基准并加入防抖，同样的数据会正确落在「复苏」。
• 数据源失效不再显示绿灯：过去某个数据源整体抓取失败时，健康指示灯依然是绿的，采集脚本也照样以「成功」退出。现在会真实报警并以非零状态退出。
• 图表缩放不再被自动刷新重置：过去后台刷新一次，你手动拉的时间区间、隐藏的图例都会被清空。
• 打开页面偶发 500 或全空白：修复了指标出现空值时接口直接报错、以及后端未启动时首页显示「暂无数据」而非提示连接失败的问题。
• 利率与采购经理指数表的重复与陈旧数据：修复同一天存在多条互相矛盾记录（利率表由 1536 行去重为 154 行），以及库存信号被一个已停更的数据源拖在 10 个月前的问题。
• 缺失数据不再被当作「强烈看空」：过去某个分项没有数据时会被当成最悲观的取值计入综合信号，现在会被剔除并重新归一，界面上也能看到各分项的数据截止时间。
• 本次还修复了并发刷新可能损坏数据库、换库后页面长期显示旧数值等约 40 项问题。

✨ 体验改进

• 所有页面的加载中、出错、无数据三种状态统一呈现，不再出现「看起来正常但其实是空的」页面。
• 重新生成评述期间，旧评述会继续展示，不再空一大块。
• 长序列图表（三千余个点）滚动与悬浮更顺滑。
• 数据采集新增超时与整轮预算，单个数据源卡死不会再拖住整次刷新。

---

This release focuses on the accuracy of the economic-cycle read-out, closes a local security hole, and stops data failures from being silently hidden.

⚠️ Heads-up

1. The app now opens at http://localhost:8000

• Why: the interface and the API used to run as two processes on two ports, which regularly served a stale page against a fresh API — and refreshing did not help.
• What is new: after launching, both the interface and the API live at 8000. There is no 5173 any more.
• What to do: update your bookmark. The old link does not redirect.

Old: http://localhost:5173
New: http://localhost:8000

2. Triggering "refresh data" and "regenerate commentary" works differently

• Why: previously any web page you happened to visit could hide an image link that silently started a full data collection on your machine — and even a metered model call.
• What is new: these actions now require a local credential. Clicking the buttons in the interface behaves exactly as before; the credential is issued locally and automatically.
• What to do: nothing, if you use the interface. If you have a script or bookmark that hits the collection address directly, it must now submit the job first and then subscribe to progress.

🐛 Bug fixes

• A 5% growth reading is no longer classified as "recession": the trend baseline had been polluted for years by the unusual 2021 base, pinning the clock in the recession quadrant and dragging the composite signal to -2. With an outlier-resistant baseline plus anti-flapping, the same data now correctly reads "recovery".
• A failed data source no longer shows a green light: when a source failed to fetch entirely, the health indicator stayed green and the collection script still exited as "successful". It now reports the failure and exits non-zero.
• Chart zoom is no longer reset by background refreshes: a single refresh used to wipe the time range you had dragged and the series you had hidden.
• Occasional 500 errors or a blank page on open: fixed the API failing outright when an indicator contained empty values, and the home page claiming "no data yet" when the backend was simply unreachable.
• Duplicate and stale rows in the rate and PMI tables: fixed contradictory records sharing one date (the rate table went from 1536 rows to 154 after de-duplication), and an inventory signal held 10 months in the past by a source that had stopped updating.
• Missing data is no longer read as "strongly bearish": a sub-signal with no data used to enter the composite at its most pessimistic value; it is now excluded and the composite renormalised, and the interface shows each sub-signal's data cut-off.
• Around 40 further issues are fixed in this release, including concurrent refreshes that could corrupt the database and pages that kept showing old values after the data was replaced.

✨ Improvements

• Loading, error and empty states are presented consistently across every page, so a page can no longer look fine while actually being empty.
• The previous commentary stays on screen while a new one is generated, instead of leaving a gap.
• Long series (three thousand-plus points) scroll and hover more smoothly.
• Data collection gained per-source timeouts and an overall run budget, so one stuck source no longer holds up the whole refresh.