Add WeatherNext2 worker and prevent empty scan cache

.gitignore

@@ -1,5 +1,8 @@
 # Secrets
 .env
+secrets/
+*.service-account.json
+gcp-sa.json
 
 # Scratch / temp scripts
 scratch/

deploy.sh

@@ -380,6 +380,9 @@ compose_up_retry "cache warmer" -d --no-deps polyweather_warmer
 echo "Updating training settlement worker..."
 compose_up_retry "training settlement" -d --no-deps polyweather_training_settlement
 
+echo "Updating WeatherNext2 worker..."
+compose_up_retry "WeatherNext2 worker" -d --no-deps polyweather_weathernext2_worker
+
 echo "Updating frontend..."
 compose_up_retry "frontend" -d --no-deps polyweather_frontend
 

docker-compose.yml

@@ -139,6 +139,9 @@ services:
       SUPABASE_SERVICE_ROLE_KEY: ${SUPABASE_SERVICE_ROLE_KEY}
       SUPABASE_URL: ${SUPABASE_URL}
       UVICORN_WORKERS: ${UVICORN_WORKERS:-2}
+      WEATHERNEXT2_CITY_HIGHS_PATH: ${WEATHERNEXT2_CITY_HIGHS_PATH:-/app/data/weathernext2_city_highs.json}
+      WEATHERNEXT2_ENABLED: ${WEATHERNEXT2_ENABLED:-1}
+      WEATHERNEXT2_MODEL_DIR: ${WEATHERNEXT2_MODEL_DIR:-/app/data/models/weathernext2_calibrator}
     healthcheck:
       interval: 30s
       retries: 3
@@ -296,6 +299,56 @@ services:
     volumes:
     - ${POLYWEATHER_RUNTIME_DATA_DIR:-/var/lib/polyweather}:/var/lib/polyweather
     - ${POLYWEATHER_RUNTIME_DATA_DIR:-/var/lib/polyweather}:/app/data
+  polyweather_weathernext2_worker:
+    command: python -m web.weathernext2_worker
+    container_name: polyweather_weathernext2_worker
+    depends_on:
+      polyweather_redis:
+        condition: service_healthy
+      polyweather_web:
+        condition: service_started
+    logging:
+      driver: "json-file"
+      options:
+        max-size: "50m"
+        max-file: "3"
+    cpus: ${POLYWEATHER_WEATHERNEXT2_CPUS:-1.50}
+    env_file: *id001
+    environment:
+      GOOGLE_APPLICATION_CREDENTIALS: ${GOOGLE_APPLICATION_CREDENTIALS:-/app/secrets/gcp-sa.json}
+      POLYWEATHER_EVENT_STORE: ${POLYWEATHER_EVENT_STORE:-redis}
+      POLYWEATHER_OBSERVATION_COLLECTOR_ENABLED: 'false'
+      POLYWEATHER_REDIS_URL: ${POLYWEATHER_REDIS_URL:-redis://polyweather_redis:6379/0}
+      POLYWEATHER_SCAN_TERMINAL_PREWARM_ENABLED: 'false'
+      POLYWEATHER_SERVICE_ROLE: weathernext2_worker
+      WEATHERNEXT2_BACKEND: ${WEATHERNEXT2_BACKEND:-gcs_zarr}
+      WEATHERNEXT2_CITY_HIGHS_PATH: ${WEATHERNEXT2_CITY_HIGHS_PATH:-/app/data/weathernext2_city_highs.json}
+      WEATHERNEXT2_DATA_ROOT: ${WEATHERNEXT2_DATA_ROOT:-/app/data/weathernext2}
+      WEATHERNEXT2_ENABLED: ${WEATHERNEXT2_ENABLED:-1}
+      WEATHERNEXT2_GCS_ZARR_URI: ${WEATHERNEXT2_GCS_ZARR_URI:-gs://weathernext/weathernext_2_0_0/zarr}
+      WEATHERNEXT2_MODEL_DIR: ${WEATHERNEXT2_MODEL_DIR:-/app/data/models/weathernext2_calibrator}
+      WEATHERNEXT2_WORKER_INITIAL_DELAY_SEC: ${WEATHERNEXT2_WORKER_INITIAL_DELAY_SEC:-90}
+      WEATHERNEXT2_WORKER_INTERVAL_SEC: ${WEATHERNEXT2_WORKER_INTERVAL_SEC:-21600}
+    healthcheck:
+      interval: 60s
+      retries: 3
+      test:
+      - CMD
+      - python
+      - -c
+      - import sqlite3; c=sqlite3.connect('/var/lib/polyweather/polyweather.db');
+        c.execute('SELECT 1'); c.close()
+      timeout: 10s
+    image: ghcr.io/yangyuan-zhen/polyweather-backend:${IMAGE_TAG:-latest}
+    mem_limit: ${POLYWEATHER_WEATHERNEXT2_MEM_LIMIT:-2g}
+    memswap_limit: ${POLYWEATHER_WEATHERNEXT2_MEMSWAP_LIMIT:-3g}
+    pids_limit: 512
+    restart: unless-stopped
+    user: ${UID:-1000}:${GID:-1000}
+    volumes:
+    - ${POLYWEATHER_RUNTIME_DATA_DIR:-/var/lib/polyweather}:/var/lib/polyweather
+    - ${POLYWEATHER_RUNTIME_DATA_DIR:-/var/lib/polyweather}:/app/data
+    - ./secrets:/app/secrets:ro
 x-polyweather-base:
   env_file: *id001
   image: ghcr.io/yangyuan-zhen/polyweather-backend:${IMAGE_TAG:-latest}

docs/weathernext2-integration.md

@@ -0,0 +1,97 @@
+# WeatherNext 2 接入说明
+
+PolyWeather 已新增 WeatherNext 2 的内部接入层，当前定位是“概率底座”和“DEB 参考模型”，不是公开 API。
+
+## 已落地
+
+- `src.data_collection.weathernext2_sources`
+  - 将 64 成员最高温预报聚合成 Polymarket 口径概率桶。
+  - 摄氏城市按单温度选项，例如 `33°C`。
+  - 华氏城市按两度市场选项，例如 `94-95°F`。
+  - 支持从 hourly 成员序列按城市当地日期切出今日最高温。
+- `WeatherDataCollector.fetch_weathernext2_probability`
+  - 默认关闭，不影响现有生产采集。
+  - `WEATHERNEXT2_ENABLED=1` 后启用。
+  - 优先读取 worker 生成的 `/app/data/weathernext2_city_highs.json`，再降级读取 fixture。
+  - 若存在 LightGBM 校准模型，会对 raw 成员分布做 quantile residual 校准后再生成市场桶概率。
+- `web.weathernext2_worker`
+  - 独立低频 worker，不在 API/Telegram 请求中训练或访问 GCS。
+  - 从 GCS Zarr 对城市经纬度最近格点取 2m temperature hourly ensemble。
+  - 按城市当地日期切今日最高温，写入离线 JSON 产物。
+  - 使用 `training_feature_records_store` + `truth_records_store` 训练 LightGBM q10/q50/q90 residual calibrator。
+- `trend_engine`
+  - 若 `weather_data["weathernext2"]` 存在，`distribution_full` 使用 WeatherNext 2 概率桶，`probability_engine=weathernext2`。
+  - WeatherNext 2 ensemble median/mean 会作为 `WeatherNext 2` 进入 `current_forecasts`，供 DEB 融合参考。
+
+## 环境变量
+
+```bash
+WEATHERNEXT2_ENABLED=1
+WEATHERNEXT2_BACKEND=gcs_zarr
+WEATHERNEXT2_DATA_ROOT=/app/data/weathernext2
+WEATHERNEXT2_CITY_HIGHS_PATH=/app/data/weathernext2_city_highs.json
+WEATHERNEXT2_MODEL_DIR=/app/data/models/weathernext2_calibrator
+WEATHERNEXT2_WORKER_INTERVAL_SEC=21600
+WEATHERNEXT2_CACHE_TTL_SEC=21600
+WEATHERNEXT2_GCS_ZARR_URI=gs://weathernext/weathernext_2_0_0/zarr
+WEATHERNEXT2_MEAN_GCS_ZARR_URI=gs://weathernext/weathernext_2_0_0_mean/zarr
+GOOGLE_APPLICATION_CREDENTIALS=/app/secrets/gcp-sa.json
+```
+
+Google 官方说明中 WeatherNext 2 GCS Zarr 路径为 `gs://weathernext/weathernext_2_0_0/zarr`，访问前需要 WeatherNext Data Request 权限。缺 GCP 凭据或无访问权限时，worker 会失败退出本轮，不覆盖旧产物；Web/API 仍按现有数据源运行。
+
+## Artifact / Fixture 格式
+
+```json
+{
+  "schema_version": 1,
+  "source": "weathernext2",
+  "backend": "gcs_zarr",
+  "generated_at": "2026-07-01T10:00:00+00:00",
+  "cities": {
+    "houston": {
+      "target_date": "2026-06-29",
+      "source_run": "2026-06-29T00:00:00Z",
+      "member_highs": {
+        "member_00": 94.1,
+        "member_01": 94.8,
+        "member_02": 95.2,
+        "member_03": 96.7
+      },
+      "summary": {
+        "members": 4,
+        "median": 95.0
+      },
+      "buckets": []
+    }
+  }
+}
+```
+
+也可以提供 hourly 成员序列：
+
+```json
+{
+  "shanghai": {
+    "target_date": "2026-06-29",
+    "timezone_offset_seconds": 28800,
+    "utc_times": ["2026-06-28T16:00:00Z", "2026-06-29T06:00:00Z"],
+    "member_hourly": {
+      "member_00": [31.2, 33.0],
+      "member_01": [30.8, 32.6]
+    }
+  }
+}
+```
+
+## 后续真实拉取
+
+Google 官方 WeatherNext 2 数据源包括 GCS Zarr、BigQuery、Earth Engine 和 Vertex AI。当前第一版已经按独立离线 job 落地：
+
+- 每 6 小时拉取最近一次 WeatherNext 2 run。
+- 对 50 个城市插值到机场/结算点。
+- 切城市当地日期今日最高温。
+- 写入 `/app/data/weathernext2_city_highs.json` 或数据库表。
+- Web/API 只读本地结果，避免实时请求 Google 数据集影响响应时间。
+
+后续可以评估 BigQuery 后端作为替代路径，但不应让 API 请求直接依赖外部 WeatherNext 数据集。

frontend/components/dashboard/scan-terminal/__tests__/modelSummaryDashboard.test.ts

@@ -30,6 +30,7 @@ export function runTests() {
       model_cluster_sources: {
         ECMWF: 28.8,
         GFS: 29.4,
+        "WeatherNext 2": 29.8,
       },
     },
     {
@@ -144,8 +145,9 @@ export function runTests() {
   assert(
     MODEL_SUMMARY_MODEL_COLUMNS.map((column) => column.key).includes("AROME HD") &&
       MODEL_SUMMARY_MODEL_COLUMNS.map((column) => column.key).includes("HRRR") &&
-      MODEL_SUMMARY_MODEL_COLUMNS.map((column) => column.key).includes("NAM"),
+      MODEL_SUMMARY_MODEL_COLUMNS.map((column) => column.key).includes("NAM") &&
-    "model summary must expose the fixed model columns including optional short-range models",
+      MODEL_SUMMARY_MODEL_COLUMNS.map((column) => column.key).includes("WeatherNext 2"),
+    "model summary must expose the fixed model columns including optional short-range and WeatherNext 2 models",
   );
   assert(summaryRows.length === 5, "model summary should keep one row per city");
   assert(summaryRows[0].cityName === "Beijing", "model summary should sort by resolved region then city name");
@@ -179,6 +181,7 @@ export function runTests() {
     "model summary should aggregate Fahrenheit probabilities into two-degree market option labels",
   );
   assert(parisRow.marketMatches.length === 3, "model summary should keep every Polymarket tradable bucket");
+  assert(beijingRow?.models["WeatherNext 2"] === 29.8, "model summary should preserve WeatherNext 2 model representative");
   assert(
     parisRow.marketMatches[0].label === "32°C" &&
       parisRow.marketMatches[0].modelProbability === null &&

frontend/components/dashboard/scan-terminal/__tests__/scanTerminalCache.test.ts

@@ -0,0 +1,58 @@
+import {
+  readScanCache,
+  writeScanCache,
+} from "@/components/dashboard/scan-terminal/use-scan-terminal-query";
+
+function assert(condition: unknown, message: string) {
+  if (!condition) throw new Error(message);
+}
+
+export function runTests() {
+  const originalLocalStorage = globalThis.localStorage;
+  const storage = new Map<string, string>();
+  const memoryStorage = {
+    getItem: (key: string) => storage.get(key) ?? null,
+    removeItem: (key: string) => { storage.delete(key); },
+    setItem: (key: string, value: string) => { storage.set(key, value); },
+  } as Storage;
+
+  Object.defineProperty(globalThis, "localStorage", {
+    configurable: true,
+    value: memoryStorage,
+  });
+
+  try {
+    writeScanCache(
+      {
+        generated_at: "2026-07-02T12:00:00Z",
+        rows: [{ city: "ankara", deb_prediction: 20.3 } as any],
+        status: "ready",
+      } as any,
+      "all",
+      "model-summary",
+    );
+    assert(
+      readScanCache("all", "model-summary", { allowStale: true })?.rows?.length === 1,
+      "non-empty scan cache should remain readable",
+    );
+
+    writeScanCache(
+      {
+        generated_at: "2026-07-02T12:05:00Z",
+        rows: [],
+        status: "ready",
+      } as any,
+      "all",
+      "model-summary",
+    );
+    assert(
+      readScanCache("all", "model-summary", { allowStale: true }) === null,
+      "empty scan cache responses must be ignored so model summary can revalidate rows",
+    );
+  } finally {
+    Object.defineProperty(globalThis, "localStorage", {
+      configurable: true,
+      value: originalLocalStorage,
+    });
+  }
+}

frontend/components/dashboard/scan-terminal/use-scan-terminal-query.ts

@@ -36,7 +36,7 @@ function scanCacheKey(tradingRegion: string, cacheScope: string): string {
   return `${SCAN_CACHE_PREFIX}:${cacheScope || "terminal"}:${tradingRegion || "all"}`;
 }
 
-function readScanCache(
+export function readScanCache(
   tradingRegion: string,
   cacheScope: string,
   options?: { allowStale?: boolean },
@@ -47,15 +47,27 @@ function readScanCache(
     const cached = JSON.parse(raw);
     const age = Date.now() - Number(cached.ts || 0);
     const maxAge = options?.allowStale ? MAX_STALE_SCAN_CACHE_MS : SCAN_CACHE_TTL_MS;
-    if (cached.ts && age >= 0 && age < maxAge && cached.data?.rows) {
+    if (
+      cached.ts &&
+      age >= 0 &&
+      age < maxAge &&
+      Array.isArray(cached.data?.rows) &&
+      cached.data.rows.length > 0
+    ) {
       return cached.data;
     }
   } catch { /* ignore */ }
   return null;
 }
 
-function writeScanCache(data: ScanTerminalResponse, tradingRegion: string, cacheScope: string) {
+export function writeScanCache(data: ScanTerminalResponse, tradingRegion: string, cacheScope: string) {
-  try { localStorage.setItem(scanCacheKey(tradingRegion, cacheScope), JSON.stringify({ ts: Date.now(), data })); } catch { /* ignore */ }
+  try {
+    if (!Array.isArray(data.rows) || data.rows.length <= 0) {
+      localStorage.removeItem(scanCacheKey(tradingRegion, cacheScope));
+      return;
+    }
+    localStorage.setItem(scanCacheKey(tradingRegion, cacheScope), JSON.stringify({ ts: Date.now(), data }));
+  } catch { /* ignore */ }
 }
 
 function normalizeCityKey(city: string | null | undefined) {

frontend/lib/model-summary.ts

@@ -16,6 +16,7 @@ export const MODEL_SUMMARY_MODEL_COLUMNS = [
   { key: "AROME HD", label: "AROME HD" },
   { key: "HRRR", label: "HRRR" },
   { key: "NAM", label: "NAM" },
+  { key: "WeatherNext 2", label: "WeatherNext 2" },
 ] as const;
 
 export type ModelSummaryColumnKey = (typeof MODEL_SUMMARY_MODEL_COLUMNS)[number]["key"];

requirements.lock

@@ -4,6 +4,7 @@ aiohappyeyeballs==2.6.2
     # via aiohttp
 aiohttp==3.14.1
     # via
+    #   gcsfs
     #   pytelegrambotapi
     #   web3
 aiosignal==1.4.0
@@ -28,6 +29,8 @@ certifi==2026.5.20
     #   httpx
     #   netcdf4
     #   requests
+cffi==2.0.0 ; platform_python_implementation != 'PyPy'
+    # via cryptography
 cftime==1.6.5
     # via netcdf4
 charset-normalizer==3.4.7
@@ -40,8 +43,14 @@ colorama==0.4.6 ; sys_platform == 'win32'
     # via
     #   click
     #   loguru
+cryptography==49.0.0
+    # via google-auth
 cytoolz==1.1.0 ; implementation_name == 'cpython'
     # via eth-utils
+decorator==5.3.1
+    # via gcsfs
+donfig==0.8.1.post1
+    # via zarr
 eth-abi==5.2.0
     # via
     #   eth-account
@@ -81,6 +90,58 @@ frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
+fsspec==2026.6.0
+    # via
+    #   -r requirements.txt
+    #   gcsfs
+gcsfs==2026.6.0
+    # via -r requirements.txt
+google-api-core==2.31.0
+    # via
+    #   google-cloud-core
+    #   google-cloud-storage
+    #   google-cloud-storage-control
+google-auth==2.55.1
+    # via
+    #   gcsfs
+    #   google-api-core
+    #   google-auth-oauthlib
+    #   google-cloud-core
+    #   google-cloud-storage
+    #   google-cloud-storage-control
+google-auth-oauthlib==1.4.0
+    # via gcsfs
+google-cloud-core==2.6.0
+    # via google-cloud-storage
+google-cloud-storage==3.12.0
+    # via
+    #   -r requirements.txt
+    #   gcsfs
+google-cloud-storage-control==1.12.0
+    # via gcsfs
+google-crc32c==1.8.0
+    # via
+    #   google-cloud-storage
+    #   google-resumable-media
+    #   zarr
+google-resumable-media==2.10.0
+    # via google-cloud-storage
+googleapis-common-protos==1.75.0
+    # via
+    #   google-api-core
+    #   grpc-google-iam-v1
+    #   grpcio-status
+grpc-google-iam-v1==0.14.4
+    # via google-cloud-storage-control
+grpcio==1.81.1
+    # via
+    #   google-api-core
+    #   google-cloud-storage-control
+    #   googleapis-common-protos
+    #   grpc-google-iam-v1
+    #   grpcio-status
+grpcio-status==1.81.1
+    # via google-api-core
 h11==0.16.0
     # via
     #   httpcore
@@ -100,33 +161,72 @@ idna==3.18
     #   httpx
     #   requests
     #   yarl
+joblib==1.5.3
+    # via
+    #   -r requirements.txt
+    #   scikit-learn
+lightgbm==4.6.0
+    # via -r requirements.txt
 loguru==0.7.3
     # via -r requirements.txt
 multidict==6.7.1
     # via
     #   aiohttp
     #   yarl
+narwhals==2.23.0
+    # via scikit-learn
 netcdf4==1.7.4
     # via -r requirements.txt
+numcodecs==0.16.5
+    # via zarr
 numpy==2.4.6
-
-torch
-typing_extensions==4.15.0
-    # via torch
     # via
     #   -r requirements.txt
     #   cftime
+    #   lightgbm
     #   netcdf4
+    #   numcodecs
+    #   pandas
+    #   scikit-learn
+    #   scipy
+    #   xarray
+    #   zarr
+oauthlib==3.3.1
+    # via requests-oauthlib
+packaging==26.2
+    # via
+    #   xarray
+    #   zarr
+pandas==3.0.3
+    # via xarray
 parsimonious==0.10.0
     # via eth-abi
 propcache==0.5.2
     # via
     #   aiohttp
     #   yarl
+proto-plus==1.28.0
+    # via
+    #   google-api-core
+    #   google-cloud-storage-control
+protobuf==7.35.1
+    # via
+    #   google-api-core
+    #   google-cloud-storage-control
+    #   googleapis-common-protos
+    #   grpc-google-iam-v1
+    #   grpcio-status
+    #   proto-plus
 pyaes==1.6.1
     # via telethon
 pyasn1==0.6.3
-    # via rsa
+    # via
+    #   pyasn1-modules
+    #   rsa
+pyasn1-modules==0.4.2
+    # via google-auth
+pycparser==3.0 ; implementation_name != 'PyPy' and platform_python_implementation != 'PyPy'
+    # via cffi
 pycryptodome==3.23.0
     # via
     #   eth-hash
@@ -141,6 +241,8 @@ pydantic-core==2.46.4
     # via pydantic
 pytelegrambotapi==4.34.0
     # via -r requirements.txt
+python-dateutil==2.9.0.post0
+    # via pandas
 python-dotenv==1.2.2
     # via -r requirements.txt
 pyunormalize==17.0.0
@@ -148,7 +250,9 @@ pyunormalize==17.0.0
 pywin32==312 ; sys_platform == 'win32'
     # via web3
 pyyaml==6.0.3
-    # via -r requirements.txt
+    # via
+    #   -r requirements.txt
+    #   donfig
 redis==6.4.0
     # via -r requirements.txt
 regex==2026.5.9
@@ -156,18 +260,40 @@ regex==2026.5.9
 requests==2.34.2
     # via
     #   -r requirements.txt
+    #   gcsfs
+    #   google-api-core
+    #   google-cloud-storage
     #   pytelegrambotapi
+    #   requests-oauthlib
     #   web3
+requests-oauthlib==2.0.0
+    # via google-auth-oauthlib
 rlp==4.1.0
     # via
     #   eth-account
     #   eth-rlp
 rsa==4.9.1
     # via telethon
+scikit-learn==1.9.0
+    # via -r requirements.txt
+scipy==1.17.1 ; python_full_version < '3.12'
+    # via
+    #   -r requirements.txt
+    #   lightgbm
+    #   scikit-learn
+scipy==1.18.0 ; python_full_version >= '3.12'
+    # via
+    #   -r requirements.txt
+    #   lightgbm
+    #   scikit-learn
+six==1.17.0
+    # via python-dateutil
 starlette==1.3.1
     # via fastapi
 telethon==1.43.2
     # via -r requirements.txt
+threadpoolctl==3.6.0
+    # via scikit-learn
 toolz==1.1.0 ; implementation_name == 'cpython' or implementation_name == 'pypy'
     # via
     #   cytoolz
@@ -181,17 +307,22 @@ typing-extensions==4.15.0
     #   anyio
     #   eth-typing
     #   fastapi
+    #   grpcio
+    #   numcodecs
     #   pydantic
     #   pydantic-core
     #   starlette
     #   typing-inspection
     #   web3
+    #   zarr
 typing-inspection==0.4.2
     # via
     #   fastapi
     #   pydantic
-tzdata==2026.2 ; sys_platform == 'win32'
+tzdata==2026.2 ; sys_platform == 'emscripten' or sys_platform == 'win32'
-    # via -r requirements.txt
+    # via
+    #   -r requirements.txt
+    #   pandas
 urllib3==2.7.0
     # via
     #   requests
@@ -206,5 +337,11 @@ websockets==15.0.1
     #   web3
 win32-setctime==1.2.0 ; sys_platform == 'win32'
     # via loguru
+xarray==2026.4.0
+    # via -r requirements.txt
 yarl==1.24.2
     # via aiohttp
+zarr==3.1.6 ; python_full_version < '3.12'
+    # via -r requirements.txt
+zarr==3.2.1 ; python_full_version >= '3.12'
+    # via -r requirements.txt

requirements.txt

@@ -7,6 +7,15 @@ python-dotenv>=1.1,<2
 PyYAML>=6.0,<7
 netCDF4>=1.7,<2
 numpy>=2.2,<3
+scipy>=1.10,<2
+scikit-learn>=1.4,<2
+lightgbm>=4.6,<5
+joblib>=1.3,<2
+xarray>=2024.7,<2027
+zarr>=2.18,<4
+fsspec>=2024.10,<2027
+gcsfs>=2024.10,<2027
+google-cloud-storage>=2.18,<4
 web3>=7.13,<8
 fastapi>=0.115,<1
 uvicorn>=0.34,<1

src/analysis/deb_algorithm.py

@@ -110,6 +110,7 @@ def _deb_model_priority(model_name: str) -> int:
         "hko": 45,
         "lgbm": 50,
         "openmeteo": 15,
+        "weathernext2": 30,
     }.get(normalized, 10)
 
 

src/analysis/trend_engine.py

@@ -50,6 +50,49 @@ def _sf(v):
         return None
 
 
+def _weathernext2_probability_payload(
+    weather_data: Dict[str, Any],
+) -> Optional[Dict[str, Any]]:
+    source = weather_data.get("weathernext2")
+    if not isinstance(source, dict):
+        return None
+    raw_buckets = source.get("buckets")
+    if not isinstance(raw_buckets, list) or not raw_buckets:
+        return None
+
+    buckets = []
+    for bucket in raw_buckets:
+        if not isinstance(bucket, dict):
+            continue
+        probability = _sf(bucket.get("probability"))
+        if probability is None or probability <= 0:
+            continue
+        copied = dict(bucket)
+        copied["probability"] = round(probability, 3)
+        buckets.append(copied)
+    if not buckets:
+        return None
+
+    summary = source.get("summary") if isinstance(source.get("summary"), dict) else {}
+    mu = _sf(summary.get("median"))
+    if mu is None:
+        mu = _sf(summary.get("mean"))
+    if mu is None:
+        top_bucket = max(buckets, key=lambda item: _sf(item.get("probability")) or 0)
+        mu = _sf(top_bucket.get("value"))
+
+    return {
+        "engine": "weathernext2",
+        "mu": mu,
+        "probabilities": sorted(
+            buckets,
+            key=lambda item: _sf(item.get("probability")) or 0,
+            reverse=True,
+        )[:4],
+        "probabilities_all": buckets,
+    }
+
+
 def _median(values: List[float]) -> Optional[float]:
     if not values:
         return None
@@ -548,6 +591,18 @@ def analyze_weather_trend(
     for m_name, m_val in mm_forecasts.items():
         if m_val is not None and not _is_excluded_model_name(m_name):
             current_forecasts[m_name] = _sf(m_val)
+    weathernext2 = weather_data.get("weathernext2")
+    if isinstance(weathernext2, dict):
+        weathernext2_summary = (
+            weathernext2.get("summary")
+            if isinstance(weathernext2.get("summary"), dict)
+            else {}
+        )
+        weathernext2_median = _sf(weathernext2_summary.get("median"))
+        if weathernext2_median is None:
+            weathernext2_median = _sf(weathernext2_summary.get("mean"))
+        if weathernext2_median is not None:
+            current_forecasts["WeatherNext 2"] = weathernext2_median
     forecast_highs = [h for h in current_forecasts.values() if h is not None]
     forecast_high = max(forecast_highs) if forecast_highs else None
     forecast_median = (
@@ -865,9 +920,12 @@ def analyze_weather_trend(
     # === Probability Engine ===
     probabilities: List[Dict[str, Any]] = []
     probabilities_all: List[Dict[str, Any]] = []
+    probability_engine = "legacy"
     forecast_miss_deg = 0.0
+    weathernext2_probs = _weathernext2_probability_payload(weather_data)
 
     if is_dead_market:
+        probability_engine = "dead_market"
         settled_wu = apply_city_settlement(city_name, max_so_far) if max_so_far is not None else 0
         dead_msg = (
             f"🎲 <b>结算预测</b>：已锁定 {settled_wu}{temp_symbol} "
@@ -881,6 +939,24 @@ def analyze_weather_trend(
                 {"value": settled_wu, "range": f"[{settled_wu-0.5}~{settled_wu+0.5})", "probability": 1.0}
             ]
             probabilities_all = probabilities
+    elif weathernext2_probs:
+        if max_so_far is not None and forecast_median is not None:
+            forecast_miss_deg = round(forecast_median - max_so_far, 1)
+        probability_engine = "weathernext2"
+        mu = weathernext2_probs.get("mu") or mu
+        probabilities = weathernext2_probs.get("probabilities", [])
+        probabilities_all = weathernext2_probs.get("probabilities_all", probabilities)
+        prob_parts = []
+        for bucket in probabilities[:4]:
+            label = str(bucket.get("label") or bucket.get("range") or bucket.get("value") or "").strip()
+            probability = _sf(bucket.get("probability"))
+            if label and probability is not None:
+                prob_parts.append(f"{label} {probability * 100:.0f}%")
+        if prob_parts:
+            mu_label = f"μ={mu:.1f}" if mu is not None else "μ=--"
+            prob_str = " | ".join(prob_parts)
+            insights.append(f"🎲 <b>WeatherNext 2 概率</b> ({mu_label})：{prob_str}")
+            ai_features.append(f"🎲 WeatherNext 2 概率分布：{prob_str}")
     elif (ens_p10 is not None and ens_p90 is not None) or fallback_sigma:
         # Forecast miss magnitude
         if max_so_far is not None and forecast_median is not None:
@@ -1117,7 +1193,7 @@ def analyze_weather_trend(
         "mu": mu,
         "probabilities": probabilities,
         "probabilities_all": probabilities_all or probabilities,
-        "probability_engine": "legacy",
+        "probability_engine": probability_engine,
         "trend_info": {
             "direction": trend_direction if 'trend_direction' in dir() else "unknown",
             "recent": recent_list,

src/analysis/weathernext2_calibration.py

@@ -0,0 +1,325 @@
+from __future__ import annotations
+
+import json
+import math
+import os
+import time
+from pathlib import Path
+from typing import Any, Dict, Iterable, Optional
+
+from src.data_collection.weathernext2_sources import build_weathernext2_city_probability
+
+
+QUANTILES = {"q10": 0.10, "q50": 0.50, "q90": 0.90}
+FEATURE_NAMES = [
+    "city_code",
+    "wn2_mean",
+    "wn2_median",
+    "wn2_p10",
+    "wn2_p25",
+    "wn2_p75",
+    "wn2_p90",
+    "wn2_spread",
+    "deb_prediction_c",
+    "model_median_c",
+    "model_spread",
+    "current_max_so_far_c",
+    "local_hour",
+    "month",
+    "day_of_year",
+    "observation_progress",
+]
+
+
+def _sf(value: Any) -> Optional[float]:
+    try:
+        if value is None or value == "":
+            return None
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if math.isfinite(parsed) else None
+
+
+def _date_parts(value: Any) -> tuple[float, float]:
+    text = str(value or "").strip()
+    try:
+        parsed = time.strptime(text[:10], "%Y-%m-%d")
+        return float(parsed.tm_mon), float(parsed.tm_yday)
+    except Exception:
+        return 0.0, 0.0
+
+
+def _summary_from_record(record: Dict[str, Any]) -> Dict[str, Any]:
+    wn2 = record.get("weathernext2") if isinstance(record.get("weathernext2"), dict) else {}
+    summary = wn2.get("summary") if isinstance(wn2.get("summary"), dict) else {}
+    return summary if isinstance(summary, dict) else {}
+
+
+def _city_key(value: Any) -> str:
+    return str(value or "").strip().lower()
+
+
+def _build_city_index(records: Iterable[Dict[str, Any]]) -> Dict[str, int]:
+    cities = sorted({_city_key(record.get("city")) for record in records if _city_key(record.get("city"))})
+    return {city: idx for idx, city in enumerate(cities)}
+
+
+def _feature_row(record: Dict[str, Any], city_index: Dict[str, int]) -> Optional[list[float]]:
+    summary = _summary_from_record(record)
+    median = _sf(summary.get("median"))
+    if median is None:
+        return None
+    target_date = record.get("target_date") or record.get("date")
+    month, day_of_year = _date_parts(target_date)
+    city = _city_key(record.get("city"))
+    current = _sf(record.get("current_max_so_far_c"))
+    progress = _sf(record.get("observation_progress"))
+    if progress is None:
+        local_hour = _sf(record.get("local_hour"))
+        progress = min(max((local_hour or 0.0) / 24.0, 0.0), 1.0)
+
+    def fallback(name: str, default: float) -> float:
+        parsed = _sf(summary.get(name))
+        return default if parsed is None else parsed
+
+    return [
+        float(city_index.get(city, -1)),
+        fallback("mean", median),
+        median,
+        fallback("p10", median),
+        fallback("p25", median),
+        fallback("p75", median),
+        fallback("p90", median),
+        fallback("spread", 0.0),
+        _sf(record.get("deb_prediction_c")) or median,
+        _sf(record.get("model_median_c")) or median,
+        _sf(record.get("model_spread")) or 0.0,
+        current if current is not None else median,
+        _sf(record.get("local_hour")) or 0.0,
+        month,
+        day_of_year,
+        progress,
+    ]
+
+
+def _training_xy(
+    records: Iterable[Dict[str, Any]],
+    city_index: Dict[str, int],
+) -> tuple[list[list[float]], list[float], list[Dict[str, Any]]]:
+    features: list[list[float]] = []
+    residuals: list[float] = []
+    kept: list[Dict[str, Any]] = []
+    for record in records:
+        summary = _summary_from_record(record)
+        median = _sf(summary.get("median"))
+        actual = _sf(record.get("actual_high_c", record.get("actual_high")))
+        row = _feature_row(record, city_index)
+        if median is None or actual is None or row is None:
+            continue
+        features.append(row)
+        residuals.append(actual - median)
+        kept.append(record)
+    return features, residuals, kept
+
+
+def train_lightgbm_quantile_calibrator(
+    records: Iterable[Dict[str, Any]],
+    *,
+    model_dir: os.PathLike[str] | str,
+    min_global_samples: int = 150,
+    min_city_samples: int = 5,
+) -> Dict[str, Any]:
+    rows = [record for record in records if isinstance(record, dict)]
+    city_index = _build_city_index(rows)
+    features, residuals, kept = _training_xy(rows, city_index)
+    if len(features) < int(min_global_samples):
+        return {
+            "trained": False,
+            "reason": "insufficient_global_samples",
+            "samples": len(features),
+        }
+
+    city_counts: Dict[str, int] = {}
+    for record in kept:
+        city_counts[_city_key(record.get("city"))] = city_counts.get(_city_key(record.get("city")), 0) + 1
+    if not any(count >= int(min_city_samples) for count in city_counts.values()):
+        return {
+            "trained": False,
+            "reason": "insufficient_city_samples",
+            "samples": len(features),
+        }
+
+    try:
+        import joblib  # type: ignore
+        from lightgbm import LGBMRegressor  # type: ignore
+    except Exception as exc:
+        return {
+            "trained": False,
+            "reason": "missing_lightgbm",
+            "samples": len(features),
+            "error": str(exc),
+        }
+
+    target_dir = Path(model_dir)
+    target_dir.mkdir(parents=True, exist_ok=True)
+    models = {}
+    for key, alpha in QUANTILES.items():
+        model = LGBMRegressor(
+            objective="quantile",
+            alpha=alpha,
+            n_estimators=45,
+            learning_rate=0.08,
+            num_leaves=15,
+            min_child_samples=5,
+            random_state=42,
+            n_jobs=2,
+            verbosity=-1,
+        )
+        model.fit(features, residuals)
+        models[key] = model
+        joblib.dump(model, target_dir / f"{key}.pkl")
+
+    metadata = {
+        "model_version": f"weathernext2_lightgbm_quantile_{int(time.time())}",
+        "engine": "lightgbm_quantile",
+        "samples": len(features),
+        "feature_names": FEATURE_NAMES,
+        "city_index": city_index,
+        "city_counts": city_counts,
+        "created_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
+    }
+    (target_dir / "metadata.json").write_text(
+        json.dumps(metadata, ensure_ascii=False, indent=2, sort_keys=True) + "\n",
+        encoding="utf-8",
+    )
+
+    ordered = True
+    for feature in features[: min(len(features), 50)]:
+        preds = sorted(float(models[key].predict([feature])[0]) for key in ("q10", "q50", "q90"))
+        if preds[0] > preds[1] or preds[1] > preds[2]:
+            ordered = False
+            break
+    return {
+        "trained": True,
+        "samples": len(features),
+        "model_dir": str(target_dir),
+        "model_version": metadata["model_version"],
+        "validation": {"ordered_quantiles": ordered},
+    }
+
+
+def _load_model_bundle(model_dir: os.PathLike[str] | str) -> Optional[Dict[str, Any]]:
+    target_dir = Path(model_dir)
+    metadata_path = target_dir / "metadata.json"
+    if not metadata_path.is_file():
+        return None
+    try:
+        import joblib  # type: ignore
+
+        metadata = json.loads(metadata_path.read_text(encoding="utf-8"))
+        return {
+            "metadata": metadata,
+            "models": {
+                key: joblib.load(target_dir / f"{key}.pkl")
+                for key in QUANTILES
+            },
+        }
+    except Exception:
+        return None
+
+
+def _predict_residual_quantiles(record: Dict[str, Any], bundle: Dict[str, Any]) -> Optional[Dict[str, float]]:
+    metadata = bundle.get("metadata") or {}
+    city_index = metadata.get("city_index") if isinstance(metadata.get("city_index"), dict) else {}
+    feature = _feature_row(record, city_index)
+    if feature is None:
+        return None
+    raw = {
+        key: float(model.predict([feature])[0])
+        for key, model in (bundle.get("models") or {}).items()
+    }
+    values = sorted([raw.get("q10", 0.0), raw.get("q50", 0.0), raw.get("q90", 0.0)])
+    return {"q10": values[0], "q50": values[1], "q90": values[2]}
+
+
+def _calibrate_members(
+    member_highs: Dict[str, Any],
+    raw_summary: Dict[str, Any],
+    residuals: Dict[str, float],
+) -> Dict[str, float]:
+    raw_median = _sf(raw_summary.get("median"))
+    raw_p10 = _sf(raw_summary.get("p10"))
+    raw_p90 = _sf(raw_summary.get("p90"))
+    if raw_median is None:
+        return {}
+    target_p10 = (raw_p10 if raw_p10 is not None else raw_median) + residuals["q10"]
+    target_median = raw_median + residuals["q50"]
+    target_p90 = (raw_p90 if raw_p90 is not None else raw_median) + residuals["q90"]
+
+    calibrated = {}
+    for member_id, value in (member_highs or {}).items():
+        parsed = _sf(value)
+        if parsed is None:
+            continue
+        if parsed <= raw_median:
+            raw_span = max(raw_median - (raw_p10 if raw_p10 is not None else raw_median), 0.1)
+            target_span = max(target_median - target_p10, 0.1)
+            adjusted = target_median - (raw_median - parsed) / raw_span * target_span
+        else:
+            raw_span = max((raw_p90 if raw_p90 is not None else raw_median) - raw_median, 0.1)
+            target_span = max(target_p90 - target_median, 0.1)
+            adjusted = target_median + (parsed - raw_median) / raw_span * target_span
+        calibrated[str(member_id)] = round(adjusted, 1)
+    return calibrated
+
+
+def apply_quantile_calibration_to_payload(
+    payload: Dict[str, Any],
+    *,
+    model_dir: os.PathLike[str] | str,
+) -> Dict[str, Any]:
+    bundle = _load_model_bundle(model_dir)
+    if not bundle:
+        return dict(payload)
+    raw_summary = payload.get("summary") if isinstance(payload.get("summary"), dict) else {}
+    member_highs = payload.get("member_highs") if isinstance(payload.get("member_highs"), dict) else {}
+    if not raw_summary or not member_highs:
+        return dict(payload)
+
+    record = {
+        "city": payload.get("city"),
+        "target_date": payload.get("target_date"),
+        "weathernext2": {"summary": raw_summary},
+    }
+    residuals = _predict_residual_quantiles(record, bundle)
+    if residuals is None:
+        return dict(payload)
+
+    calibrated_members = _calibrate_members(member_highs, raw_summary, residuals)
+    if not calibrated_members:
+        return dict(payload)
+
+    calibrated = build_weathernext2_city_probability(
+        city=str(payload.get("city") or ""),
+        member_highs=calibrated_members,
+        temp_symbol=str(payload.get("temp_symbol") or "°C"),
+        target_date=payload.get("target_date"),
+        source_run=payload.get("source_run"),
+        generated_at=payload.get("generated_at"),
+    )
+    metadata = bundle.get("metadata") or {}
+    calibrated["calibration"] = {
+        "engine": "lightgbm_quantile",
+        "model_version": metadata.get("model_version"),
+        "samples": metadata.get("samples"),
+        "residual_quantiles": {key: round(value, 3) for key, value in residuals.items()},
+        "raw_summary": raw_summary,
+        "calibrated_summary": calibrated.get("summary"),
+    }
+    calibrated["raw_weathernext2"] = {
+        "summary": raw_summary,
+        "buckets": payload.get("buckets") or [],
+        "top_bucket": payload.get("top_bucket"),
+    }
+    return calibrated

src/data_collection/weather_sources.py

@@ -18,6 +18,7 @@ from src.data_collection.metar_sources import MetarSourceMixin
 from src.data_collection.mgm_sources import MgmSourceMixin
 from src.data_collection.jma_amedas_sources import JmaAmedasSourceMixin
 from src.data_collection.nws_open_meteo_sources import NwsOpenMeteoSourceMixin
+from src.data_collection.weathernext2_sources import WeatherNext2SourceMixin
 from src.data_collection.amos_station_sources import AmosStationSourceMixin
 from src.data_collection.amsc_awos_sources import AmscAwosSourceMixin
 from src.data_collection.fmi_sources import FmiSourceMixin
@@ -35,7 +36,7 @@ from src.data_collection.forecast_source_bundle import fetch_open_meteo_forecast
 from src.database.db_manager import DBManager
 
 
-class WeatherDataCollector(OpenMeteoCacheMixin, SettlementSourceMixin, MetarSourceMixin, MgmSourceMixin, JmaAmedasSourceMixin, NwsOpenMeteoSourceMixin, AmosStationSourceMixin, AmscAwosSourceMixin, FmiSourceMixin, KnmiSourceMixin, HkoObsSourceMixin, CowinSourceMixin, MadisSourceMixin, SingaporeMssSourceMixin, ImsSourceMixin, NcmSourceMixin, AerowebSourceMixin, WundergroundHistoricalMixin):
+class WeatherDataCollector(OpenMeteoCacheMixin, SettlementSourceMixin, MetarSourceMixin, MgmSourceMixin, JmaAmedasSourceMixin, NwsOpenMeteoSourceMixin, WeatherNext2SourceMixin, AmosStationSourceMixin, AmscAwosSourceMixin, FmiSourceMixin, KnmiSourceMixin, HkoObsSourceMixin, CowinSourceMixin, MadisSourceMixin, SingaporeMssSourceMixin, ImsSourceMixin, NcmSourceMixin, AerowebSourceMixin, WundergroundHistoricalMixin):
     """
     Multi-source weather data collector
 
@@ -189,9 +190,11 @@ class WeatherDataCollector(OpenMeteoCacheMixin, SettlementSourceMixin, MetarSour
         self._open_meteo_cache: Dict[str, Dict] = {}
         self._ensemble_cache: Dict[str, Dict] = {}
         self._multi_model_cache: Dict[str, Dict] = {}
+        self._weathernext2_cache: Dict[str, Dict] = {}
         self._open_meteo_cache_lock = threading.Lock()
         self._ensemble_cache_lock = threading.Lock()
         self._multi_model_cache_lock = threading.Lock()
+        self._weathernext2_cache_lock = threading.Lock()
         # Open-Meteo 共享 429 冷却计时器：触发限流后所有 OM 端点暂停请求
         self._open_meteo_rate_limit_until: float = 0.0
         self._open_meteo_rl_cooldown: int = int(
@@ -1740,6 +1743,26 @@ class WeatherDataCollector(OpenMeteoCacheMixin, SettlementSourceMixin, MetarSour
             if multi_model_data:
                 results["multi_model"] = multi_model_data
 
+    def _attach_weathernext2_model(
+        self,
+        results: Dict,
+        city: str,
+        lat: float,
+        lon: float,
+        use_fahrenheit: bool,
+        *,
+        timezone_offset_seconds: Optional[int] = None,
+    ) -> None:
+        payload = self.fetch_weathernext2_probability(
+            city,
+            lat,
+            lon,
+            use_fahrenheit=use_fahrenheit,
+            timezone_offset_seconds=timezone_offset_seconds,
+        )
+        if payload:
+            results["weathernext2"] = payload
+
     def fetch_all_sources(
         self,
         city: str,
@@ -1794,6 +1817,14 @@ class WeatherDataCollector(OpenMeteoCacheMixin, SettlementSourceMixin, MetarSour
                 results["open-meteo"] = open_meteo
                 # 获取时区偏移以过滤 METAR
                 utc_offset = open_meteo.get("utc_offset", 0)
+                self._attach_weathernext2_model(
+                    results,
+                    city_lower,
+                    lat,
+                    lon,
+                    use_fahrenheit,
+                    timezone_offset_seconds=utc_offset,
+                )
                 if supports_aviationweather:
                     metar_data = self.fetch_metar(
                         city, use_fahrenheit=use_fahrenheit, utc_offset=utc_offset
@@ -1844,6 +1875,14 @@ class WeatherDataCollector(OpenMeteoCacheMixin, SettlementSourceMixin, MetarSour
                 fallback_utc_offset = int(
                     self.CITY_REGISTRY.get(city_lower, {}).get("tz_offset", 0)
                 )
+                self._attach_weathernext2_model(
+                    results,
+                    city_lower,
+                    lat,
+                    lon,
+                    use_fahrenheit,
+                    timezone_offset_seconds=fallback_utc_offset,
+                )
                 if supports_aviationweather:
                     metar_data = self.fetch_metar(
                         city,

src/data_collection/weathernext2_fetcher.py

@@ -0,0 +1,243 @@
+from __future__ import annotations
+
+import math
+import os
+from datetime import datetime, timezone
+from typing import Any, Dict, Mapping, Optional, Sequence
+
+
+TEMPERATURE_VARIABLE_CANDIDATES = (
+    "temperature_2m",
+    "2m_temperature",
+    "t2m",
+    "air_temperature_2m",
+    "temperature",
+)
+
+
+def _mapping_keys(dataset: Any) -> list[str]:
+    if isinstance(dataset, Mapping):
+        return [str(key) for key in dataset.keys()]
+    data_vars = getattr(dataset, "data_vars", None)
+    if data_vars is not None:
+        try:
+            return [str(key) for key in data_vars.keys()]
+        except Exception:
+            return []
+    return []
+
+
+def select_temperature_variable(dataset: Any, preferred: Optional[str] = None) -> str:
+    if preferred:
+        keys = set(_mapping_keys(dataset))
+        if preferred in keys:
+            return preferred
+        raise KeyError(f"WeatherNext2 temperature variable not found: {preferred}")
+
+    keys = _mapping_keys(dataset)
+    lowered = {key.lower(): key for key in keys}
+    for candidate in TEMPERATURE_VARIABLE_CANDIDATES:
+        if candidate.lower() in lowered:
+            return lowered[candidate.lower()]
+
+    for key in keys:
+        normalized = key.lower().replace("-", "_")
+        if "temperature" in normalized and ("2m" in normalized or "two_meter" in normalized):
+            return key
+    raise KeyError("WeatherNext2 2m temperature variable was not found")
+
+
+def normalize_temperature_value(
+    value: Any,
+    units: str = "",
+    *,
+    use_fahrenheit: bool = False,
+) -> Optional[float]:
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return None
+    if not math.isfinite(parsed):
+        return None
+
+    unit = str(units or "").strip().lower()
+    if unit in {"k", "kelvin"} or parsed > 150:
+        celsius = parsed - 273.15
+    elif unit in {"f", "fahrenheit", "degf", "degree_fahrenheit"}:
+        celsius = (parsed - 32.0) * 5.0 / 9.0
+    else:
+        celsius = parsed
+
+    if use_fahrenheit:
+        return round(celsius * 9.0 / 5.0 + 32.0, 1)
+    return round(celsius, 1)
+
+
+def _get_dataset_value(dataset: Any, key: str) -> Any:
+    if isinstance(dataset, Mapping):
+        return dataset[key]
+    return dataset[key]
+
+
+def _as_list(values: Any) -> list[Any]:
+    if hasattr(values, "values"):
+        try:
+            values = values.values
+        except Exception:
+            pass
+    if hasattr(values, "tolist"):
+        try:
+            return values.tolist()
+        except Exception:
+            pass
+    return list(values or [])
+
+
+def _nearest_index(values: Sequence[Any], target: float) -> int:
+    numeric_values = [float(value) for value in values]
+    return min(range(len(numeric_values)), key=lambda idx: abs(numeric_values[idx] - target))
+
+
+def _variable_units(dataset: Any, temp_var: str) -> str:
+    if isinstance(dataset, Mapping):
+        units = dataset.get("units")
+        if isinstance(units, Mapping):
+            return str(units.get(temp_var) or "")
+        return ""
+    variable = dataset[temp_var]
+    attrs = getattr(variable, "attrs", {}) or {}
+    return str(attrs.get("units") or attrs.get("unit") or "")
+
+
+def _variable_values(dataset: Any, temp_var: str) -> Any:
+    variable = _get_dataset_value(dataset, temp_var)
+    if hasattr(variable, "values"):
+        return variable.values
+    return variable
+
+
+def _index_nested(values: Any, indexes: Sequence[int]) -> Any:
+    current = values
+    for idx in indexes:
+        current = current[idx]
+    return current
+
+
+def _dataset_dimension_values(dataset: Any, candidates: Sequence[str]) -> tuple[str, list[Any]]:
+    keys = _mapping_keys(dataset)
+    lowered = {key.lower(): key for key in keys}
+    for candidate in candidates:
+        key = lowered.get(candidate.lower())
+        if key is not None:
+            return key, _as_list(_get_dataset_value(dataset, key))
+
+    coords = getattr(dataset, "coords", None)
+    if coords is not None:
+        for candidate in candidates:
+            if candidate in coords:
+                return candidate, _as_list(coords[candidate])
+    raise KeyError(f"WeatherNext2 dimension not found: {candidates}")
+
+
+def _parse_time(value: Any) -> str:
+    if isinstance(value, datetime):
+        dt = value
+    else:
+        text = str(value or "").strip()
+        if text.endswith("Z"):
+            text = f"{text[:-1]}+00:00"
+        try:
+            dt = datetime.fromisoformat(text)
+        except ValueError:
+            return str(value)
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo=timezone.utc)
+    return dt.astimezone(timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def extract_member_hourly_from_grid_dataset(
+    dataset: Any,
+    *,
+    lat: float,
+    lon: float,
+    temp_var: Optional[str] = None,
+    use_fahrenheit: bool = False,
+) -> Dict[str, Any]:
+    """Extract all member hourly temperatures for the nearest grid point."""
+    selected_var = select_temperature_variable(dataset, temp_var or os.getenv("WEATHERNEXT2_TEMP_VAR") or None)
+    lat_name, lat_values = _dataset_dimension_values(dataset, ("lat", "latitude"))
+    lon_name, lon_values = _dataset_dimension_values(dataset, ("lon", "longitude"))
+    member_name, member_values = _dataset_dimension_values(dataset, ("member", "realization", "ensemble_member"))
+    time_name, time_values = _dataset_dimension_values(dataset, ("time", "valid_time"))
+
+    lat_idx = _nearest_index(lat_values, lat)
+    lon_idx = _nearest_index(lon_values, lon)
+    units = _variable_units(dataset, selected_var)
+
+    variable = _get_dataset_value(dataset, selected_var)
+    dims = list(getattr(variable, "dims", []) or [member_name, time_name, lat_name, lon_name])
+    values = _variable_values(dataset, selected_var)
+    dim_indexes = {
+        member_name: None,
+        time_name: None,
+        lat_name: lat_idx,
+        lon_name: lon_idx,
+    }
+
+    member_hourly: Dict[str, list[Optional[float]]] = {}
+    for member_idx, member in enumerate(member_values):
+        hourly = []
+        for time_idx, _time_value in enumerate(time_values):
+            dim_indexes[member_name] = member_idx
+            dim_indexes[time_name] = time_idx
+            indexes = [int(dim_indexes[dim]) for dim in dims]
+            hourly.append(
+                normalize_temperature_value(
+                    _index_nested(values, indexes),
+                    units,
+                    use_fahrenheit=use_fahrenheit,
+                )
+            )
+        try:
+            member_label = f"member_{int(member):02d}"
+        except Exception:
+            member_label = f"member_{member_idx:02d}"
+        member_hourly[member_label] = hourly
+
+    return {
+        "temp_var": selected_var,
+        "units": units,
+        "lat_dim": lat_name,
+        "lon_dim": lon_name,
+        "member_dim": member_name,
+        "time_dim": time_name,
+        "nearest_lat": float(lat_values[lat_idx]),
+        "nearest_lon": float(lon_values[lon_idx]),
+        "utc_times": [_parse_time(value) for value in time_values],
+        "member_hourly": member_hourly,
+    }
+
+
+def open_weathernext2_zarr_dataset(uri: str):
+    """Open a WeatherNext 2 Zarr dataset lazily with optional runtime dependencies."""
+    try:
+        import xarray as xr  # type: ignore
+    except Exception as exc:  # pragma: no cover - depends on deployment extras
+        raise RuntimeError("xarray is required for WeatherNext2 GCS/Zarr access") from exc
+
+    storage_options = {}
+    credentials = str(os.getenv("GOOGLE_APPLICATION_CREDENTIALS", "") or "").strip()
+    if credentials:
+        storage_options["token"] = "google_default"
+    try:
+        return xr.open_zarr(
+            uri,
+            storage_options=storage_options or None,
+            consolidated=True,
+        )
+    except Exception:
+        return xr.open_zarr(
+            uri,
+            storage_options=storage_options or None,
+            consolidated=False,
+        )

src/data_collection/weathernext2_sources.py

@@ -0,0 +1,581 @@
+from __future__ import annotations
+
+import math
+import os
+import json
+import threading
+import time
+from datetime import date, datetime, timedelta, timezone
+from typing import Any, Dict, Iterable, Mapping, Optional, Sequence, Union
+
+from loguru import logger
+
+
+WEATHERNEXT2_SOURCE = "weathernext2"
+WEATHERNEXT2_PROVIDER = "google_deepmind"
+WEATHERNEXT2_GCS_ZARR_URI = "gs://weathernext/weathernext_2_0_0/zarr"
+WEATHERNEXT2_MEAN_GCS_ZARR_URI = "gs://weathernext/weathernext_2_0_0_mean/zarr"
+
+Number = Union[int, float]
+
+
+def _numeric(value: Any) -> Optional[float]:
+    if value is None:
+        return None
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if math.isfinite(parsed) else None
+
+
+def _round1(value: Number) -> float:
+    return round(float(value), 1)
+
+
+def _round3(value: Number) -> float:
+    return round(float(value), 3)
+
+
+def _settle_integer(value: Number) -> int:
+    parsed = float(value)
+    if parsed >= 0:
+        return int(math.floor(parsed + 0.5))
+    return int(math.ceil(parsed - 0.5))
+
+
+def _is_fahrenheit(temp_symbol: str) -> bool:
+    return "F" in str(temp_symbol or "").upper()
+
+
+def _unit(temp_symbol: str) -> str:
+    return "°F" if _is_fahrenheit(temp_symbol) else "°C"
+
+
+def market_bucket_for_temperature(value: Number, temp_symbol: str = "°C") -> Dict[str, Any]:
+    """Return the tradable market option bucket for a single member temperature."""
+    unit = _unit(temp_symbol)
+    settled = _settle_integer(value)
+    if _is_fahrenheit(unit):
+        lower_value = settled if settled % 2 == 0 else settled - 1
+        upper_value = lower_value + 1
+        return {
+            "key": f"{lower_value}-{upper_value}{unit}",
+            "label": f"{lower_value}-{upper_value}{unit}",
+            "lower": float(lower_value) - 0.5,
+            "upper": float(upper_value) + 0.5,
+            "sort_value": float(lower_value),
+        }
+    return {
+        "key": f"{settled}{unit}",
+        "label": f"{settled}{unit}",
+        "lower": float(settled) - 0.5,
+        "upper": float(settled) + 0.5,
+        "sort_value": float(settled),
+    }
+
+
+def _member_items(member_highs: Union[Mapping[str, Any], Sequence[Any]]) -> Iterable[tuple[str, float]]:
+    if isinstance(member_highs, Mapping):
+        iterable = member_highs.items()
+    else:
+        iterable = ((f"member_{idx:02d}", value) for idx, value in enumerate(member_highs))
+    for member_id, value in iterable:
+        parsed = _numeric(value)
+        if parsed is not None:
+            yield str(member_id), parsed
+
+
+def summarize_member_highs(member_highs: Union[Mapping[str, Any], Sequence[Any]]) -> Dict[str, Any]:
+    values = sorted(value for _member_id, value in _member_items(member_highs))
+    if not values:
+        return {
+            "members": 0,
+            "mean": None,
+            "median": None,
+            "p10": None,
+            "p25": None,
+            "p75": None,
+            "p90": None,
+            "min": None,
+            "max": None,
+            "spread": None,
+        }
+
+    def percentile(q: float) -> float:
+        if len(values) == 1:
+            return values[0]
+        position = (len(values) - 1) * q
+        lower_idx = int(math.floor(position))
+        upper_idx = int(math.ceil(position))
+        if lower_idx == upper_idx:
+            return values[lower_idx]
+        weight = position - lower_idx
+        return values[lower_idx] * (1 - weight) + values[upper_idx] * weight
+
+    return {
+        "members": len(values),
+        "mean": _round1(sum(values) / len(values)),
+        "median": _round1(percentile(0.5)),
+        "p10": _round1(percentile(0.1)),
+        "p25": _round1(percentile(0.25)),
+        "p75": _round1(percentile(0.75)),
+        "p90": _round1(percentile(0.9)),
+        "min": _round1(values[0]),
+        "max": _round1(values[-1]),
+        "spread": _round1(values[-1] - values[0]),
+    }
+
+
+def build_market_bucket_probabilities(
+    member_highs: Union[Mapping[str, Any], Sequence[Any]],
+    temp_symbol: str = "°C",
+) -> list[Dict[str, Any]]:
+    grouped: Dict[str, Dict[str, Any]] = {}
+    total_members = 0
+
+    for _member_id, value in _member_items(member_highs):
+        total_members += 1
+        option = market_bucket_for_temperature(value, temp_symbol)
+        bucket = grouped.setdefault(
+            option["key"],
+            {
+                "key": option["key"],
+                "label": option["label"],
+                "lower": option["lower"],
+                "upper": option["upper"],
+                "sort_value": option["sort_value"],
+                "weighted_sum": 0.0,
+                "member_count": 0,
+            },
+        )
+        bucket["weighted_sum"] += value
+        bucket["member_count"] += 1
+
+    if total_members <= 0:
+        return []
+
+    buckets = []
+    for bucket in grouped.values():
+        member_count = int(bucket["member_count"])
+        buckets.append(
+            {
+                "key": bucket["key"],
+                "label": bucket["label"],
+                "lower": bucket["lower"],
+                "upper": bucket["upper"],
+                "value": _round1(bucket["weighted_sum"] / member_count),
+                "probability": _round3(member_count / total_members),
+                "member_count": member_count,
+                "total_members": total_members,
+            }
+        )
+
+    return sorted(buckets, key=lambda item: (float(item["lower"]), float(item["upper"])))
+
+
+def _parse_utc_time(value: Any) -> Optional[datetime]:
+    if isinstance(value, datetime):
+        parsed = value
+    else:
+        text = str(value or "").strip()
+        if not text:
+            return None
+        if text.endswith("Z"):
+            text = f"{text[:-1]}+00:00"
+        try:
+            parsed = datetime.fromisoformat(text)
+        except ValueError:
+            return None
+    if parsed.tzinfo is None:
+        parsed = parsed.replace(tzinfo=timezone.utc)
+    return parsed.astimezone(timezone.utc)
+
+
+def _parse_date(value: Any) -> Optional[date]:
+    if isinstance(value, date) and not isinstance(value, datetime):
+        return value
+    text = str(value or "").strip()
+    if not text:
+        return None
+    try:
+        return datetime.fromisoformat(text[:10]).date()
+    except ValueError:
+        return None
+
+
+def build_city_local_daily_highs_from_hourly(
+    member_hourly: Mapping[str, Sequence[Any]],
+    utc_times: Sequence[Any],
+    timezone_offset_seconds: int,
+    target_local_date: Any,
+) -> Dict[str, float]:
+    """Reduce hourly ensemble member temperatures to local-date daily highs."""
+    target_date = _parse_date(target_local_date)
+    if target_date is None:
+        return {}
+
+    parsed_times = [_parse_utc_time(value) for value in utc_times]
+    highs: Dict[str, float] = {}
+    offset = timedelta(seconds=int(timezone_offset_seconds or 0))
+
+    for member_id, values in member_hourly.items():
+        member_values = []
+        for idx, utc_time in enumerate(parsed_times):
+            if utc_time is None or idx >= len(values):
+                continue
+            local_date = (utc_time + offset).date()
+            if local_date != target_date:
+                continue
+            parsed = _numeric(values[idx])
+            if parsed is not None:
+                member_values.append(parsed)
+        if member_values:
+            highs[str(member_id)] = _round1(max(member_values))
+
+    return highs
+
+
+def build_weathernext2_city_probability(
+    *,
+    city: str,
+    member_highs: Union[Mapping[str, Any], Sequence[Any]],
+    temp_symbol: str = "°C",
+    target_date: Optional[str] = None,
+    source_run: Optional[str] = None,
+    generated_at: Optional[str] = None,
+) -> Dict[str, Any]:
+    """Build a WeatherNext 2 probability payload aligned with tradable market options."""
+    buckets = build_market_bucket_probabilities(member_highs, temp_symbol=temp_symbol)
+    summary = summarize_member_highs(member_highs)
+    normalized_member_highs = {
+        member_id: _round1(value)
+        for member_id, value in _member_items(member_highs)
+    }
+    top_bucket = (
+        max(buckets, key=lambda item: (float(item["probability"]), float(item["lower"])))
+        if buckets
+        else None
+    )
+
+    return {
+        "source": WEATHERNEXT2_SOURCE,
+        "provider": WEATHERNEXT2_PROVIDER,
+        "city": str(city or "").strip(),
+        "target_date": target_date,
+        "source_run": source_run,
+        "generated_at": generated_at or datetime.now(timezone.utc).isoformat(),
+        "temp_symbol": _unit(temp_symbol),
+        "members": int(summary["members"] or 0),
+        "member_highs": normalized_member_highs,
+        "summary": summary,
+        "buckets": buckets,
+        "top_bucket": top_bucket,
+        "bucket_policy": "celsius_single_fahrenheit_two_degree_market_options",
+        "gcs_zarr_uri": os.getenv("WEATHERNEXT2_GCS_ZARR_URI", WEATHERNEXT2_GCS_ZARR_URI),
+        "mean_gcs_zarr_uri": os.getenv("WEATHERNEXT2_MEAN_GCS_ZARR_URI", WEATHERNEXT2_MEAN_GCS_ZARR_URI),
+    }
+
+
+def _env_enabled(name: str, default: str = "0") -> bool:
+    return str(os.getenv(name, default) or "").strip().lower() in {
+        "1",
+        "true",
+        "yes",
+        "on",
+    }
+
+
+def _city_key(city: str) -> str:
+    return str(city or "").strip().lower()
+
+
+def _load_fixture_payload(path: str, city: str) -> Optional[Dict[str, Any]]:
+    if not path:
+        return None
+    try:
+        with open(path, "r", encoding="utf-8") as fh:
+            payload = json.load(fh)
+    except Exception as exc:
+        logger.warning("WeatherNext2 fixture load failed path={}: {}", path, exc)
+        return None
+    if not isinstance(payload, dict):
+        return None
+
+    key = _city_key(city)
+    if isinstance(payload.get(key), dict):
+        return dict(payload[key])
+    for candidate, value in payload.items():
+        if _city_key(candidate) == key and isinstance(value, dict):
+            return dict(value)
+    if "member_highs" in payload or "member_hourly" in payload:
+        return dict(payload)
+    return None
+
+
+def _weathernext2_data_root() -> str:
+    return str(os.getenv("WEATHERNEXT2_DATA_ROOT", "/app/data/weathernext2") or "").strip()
+
+
+def _weathernext2_artifact_path() -> str:
+    configured = str(os.getenv("WEATHERNEXT2_CITY_HIGHS_PATH", "") or "").strip()
+    if configured:
+        return configured
+    return os.path.join(_weathernext2_data_root(), "weathernext2_city_highs.json")
+
+
+def _weathernext2_model_dir() -> str:
+    return str(
+        os.getenv(
+            "WEATHERNEXT2_MODEL_DIR",
+            "/app/data/models/weathernext2_calibrator",
+        )
+        or ""
+    ).strip()
+
+
+def _load_artifact_payload(path: str, city: str) -> Optional[Dict[str, Any]]:
+    if not path or not os.path.isfile(path):
+        return None
+    try:
+        with open(path, "r", encoding="utf-8") as fh:
+            payload = json.load(fh)
+    except Exception as exc:
+        logger.warning("WeatherNext2 artifact load failed path={}: {}", path, exc)
+        return None
+    if not isinstance(payload, dict):
+        return None
+
+    key = _city_key(city)
+    city_payloads: Any = payload.get("cities")
+    if isinstance(city_payloads, dict):
+        if isinstance(city_payloads.get(key), dict):
+            return dict(city_payloads[key])
+        for candidate, value in city_payloads.items():
+            if _city_key(candidate) == key and isinstance(value, dict):
+                return dict(value)
+    if isinstance(city_payloads, list):
+        for value in city_payloads:
+            if not isinstance(value, dict):
+                continue
+            if _city_key(value.get("city")) == key:
+                return dict(value)
+
+    if isinstance(payload.get(key), dict):
+        return dict(payload[key])
+    for candidate, value in payload.items():
+        if _city_key(candidate) == key and isinstance(value, dict):
+            return dict(value)
+    return None
+
+
+class WeatherNext2SourceMixin:
+    """Optional WeatherNext 2 probability source.
+
+    The live Google datasets require project access and optional heavy client
+    dependencies. This mixin keeps production safe by supporting a fixture or
+    prepared payload first, while reserving backend configuration for the real
+    GCS/BigQuery fetcher.
+    """
+
+    def _weathernext2_enabled(self) -> bool:
+        return _env_enabled("WEATHERNEXT2_ENABLED")
+
+    def _weathernext2_cache_key(
+        self,
+        city: str,
+        lat: float,
+        lon: float,
+        use_fahrenheit: bool,
+        target_date: Optional[str],
+    ) -> str:
+        return (
+            f"{_city_key(city)}:{round(float(lat), 4)}:{round(float(lon), 4)}:"
+            f"{'f' if use_fahrenheit else 'c'}:{target_date or ''}"
+        )
+
+    def _weathernext2_cache_state(self) -> tuple[Dict[str, Dict[str, Any]], threading.Lock, int]:
+        if not hasattr(self, "_weathernext2_cache"):
+            self._weathernext2_cache = {}
+        if not hasattr(self, "_weathernext2_cache_lock"):
+            self._weathernext2_cache_lock = threading.Lock()
+        ttl_sec = int(os.getenv("WEATHERNEXT2_CACHE_TTL_SEC", "21600"))
+        return self._weathernext2_cache, self._weathernext2_cache_lock, ttl_sec
+
+    def _weathernext2_target_date(
+        self,
+        timezone_offset_seconds: Optional[int],
+        target_date: Optional[str],
+    ) -> str:
+        if target_date:
+            return str(target_date)
+        offset = timedelta(seconds=int(timezone_offset_seconds or 0))
+        return (datetime.now(timezone.utc) + offset).date().isoformat()
+
+    def _weathernext2_from_fixture(
+        self,
+        *,
+        city: str,
+        use_fahrenheit: bool,
+        target_date: str,
+        timezone_offset_seconds: Optional[int],
+    ) -> Optional[Dict[str, Any]]:
+        fixture_path = str(os.getenv("WEATHERNEXT2_FIXTURE_PATH", "") or "").strip()
+        fixture = _load_fixture_payload(fixture_path, city)
+        if not fixture:
+            return None
+
+        temp_symbol = "°F" if use_fahrenheit else "°C"
+        fixture_target_date = str(fixture.get("target_date") or target_date)
+        member_highs = fixture.get("member_highs")
+        if member_highs is None and isinstance(fixture.get("member_hourly"), dict):
+            member_highs = build_city_local_daily_highs_from_hourly(
+                fixture["member_hourly"],
+                fixture.get("utc_times") or fixture.get("times") or [],
+                int(timezone_offset_seconds or fixture.get("timezone_offset_seconds") or 0),
+                fixture_target_date,
+            )
+        if member_highs is None:
+            return None
+
+        return build_weathernext2_city_probability(
+            city=city,
+            member_highs=member_highs,
+            temp_symbol=temp_symbol,
+            target_date=fixture_target_date,
+            source_run=fixture.get("source_run"),
+            generated_at=fixture.get("generated_at"),
+        )
+
+    def _weathernext2_normalize_prepared_payload(
+        self,
+        *,
+        payload: Mapping[str, Any],
+        city: str,
+        use_fahrenheit: bool,
+        target_date: str,
+        timezone_offset_seconds: Optional[int],
+    ) -> Optional[Dict[str, Any]]:
+        temp_symbol = "°F" if use_fahrenheit else "°C"
+        payload_target_date = str(payload.get("target_date") or target_date)
+        member_highs = payload.get("member_highs")
+        if member_highs is None and isinstance(payload.get("member_hourly"), dict):
+            member_highs = build_city_local_daily_highs_from_hourly(
+                payload["member_hourly"],
+                payload.get("utc_times") or payload.get("times") or [],
+                int(timezone_offset_seconds or payload.get("timezone_offset_seconds") or 0),
+                payload_target_date,
+            )
+        if member_highs is None and isinstance(payload.get("buckets"), list):
+            normalized = dict(payload)
+            normalized.setdefault("source", WEATHERNEXT2_SOURCE)
+            normalized.setdefault("provider", WEATHERNEXT2_PROVIDER)
+            normalized.setdefault("city", city)
+            normalized.setdefault("target_date", payload_target_date)
+            normalized.setdefault("temp_symbol", temp_symbol)
+            normalized.setdefault(
+                "gcs_zarr_uri",
+                os.getenv("WEATHERNEXT2_GCS_ZARR_URI", WEATHERNEXT2_GCS_ZARR_URI),
+            )
+            return normalized
+        if member_highs is None:
+            return None
+        return build_weathernext2_city_probability(
+            city=city,
+            member_highs=member_highs,
+            temp_symbol=temp_symbol,
+            target_date=payload_target_date,
+            source_run=payload.get("source_run"),
+            generated_at=payload.get("generated_at"),
+        )
+
+    def _weathernext2_from_artifact(
+        self,
+        *,
+        city: str,
+        use_fahrenheit: bool,
+        target_date: str,
+        timezone_offset_seconds: Optional[int],
+    ) -> Optional[Dict[str, Any]]:
+        artifact = _load_artifact_payload(_weathernext2_artifact_path(), city)
+        if not artifact:
+            return None
+        payload = self._weathernext2_normalize_prepared_payload(
+            payload=artifact,
+            city=city,
+            use_fahrenheit=use_fahrenheit,
+            target_date=target_date,
+            timezone_offset_seconds=timezone_offset_seconds,
+        )
+        if payload is None:
+            return None
+        try:
+            from src.analysis.weathernext2_calibration import apply_quantile_calibration_to_payload
+
+            return apply_quantile_calibration_to_payload(
+                payload,
+                model_dir=_weathernext2_model_dir(),
+            )
+        except Exception as exc:
+            logger.warning("WeatherNext2 calibration apply failed city={}: {}", city, exc)
+            return payload
+
+    def fetch_weathernext2_probability(
+        self,
+        city: str,
+        lat: float,
+        lon: float,
+        *,
+        use_fahrenheit: bool = False,
+        target_date: Optional[str] = None,
+        timezone_offset_seconds: Optional[int] = None,
+    ) -> Optional[Dict[str, Any]]:
+        if not self._weathernext2_enabled():
+            return None
+        if lat is None or lon is None:
+            return None
+
+        resolved_target_date = self._weathernext2_target_date(
+            timezone_offset_seconds,
+            target_date,
+        )
+        cache_key = self._weathernext2_cache_key(
+            city,
+            lat,
+            lon,
+            use_fahrenheit,
+            resolved_target_date,
+        )
+        cache, lock, ttl_sec = self._weathernext2_cache_state()
+        now_ts = time.time()
+        with lock:
+            cached = cache.get(cache_key)
+            cached_data = cached.get("data") if isinstance(cached, dict) else None
+            if isinstance(cached_data, dict) and now_ts - float(cached.get("t", 0)) < ttl_sec:
+                return dict(cached_data)
+
+        payload = self._weathernext2_from_artifact(
+            city=city,
+            use_fahrenheit=use_fahrenheit,
+            target_date=resolved_target_date,
+            timezone_offset_seconds=timezone_offset_seconds,
+        )
+        if payload is None:
+            payload = self._weathernext2_from_fixture(
+                city=city,
+                use_fahrenheit=use_fahrenheit,
+                target_date=resolved_target_date,
+                timezone_offset_seconds=timezone_offset_seconds,
+            )
+        if payload is None:
+            backend = str(os.getenv("WEATHERNEXT2_BACKEND", "") or "").strip().lower()
+            if backend:
+                logger.warning(
+                    "WeatherNext2 backend={} configured but live fetcher is not installed; use fixture/prepared payload first",
+                    backend,
+                )
+            return None
+
+        with lock:
+            cache[cache_key] = {"t": now_ts, "data": dict(payload)}
+        return payload

tests/test_deployment_runtime_config.py

@@ -100,6 +100,13 @@ def test_docker_compose_isolates_collector_from_web_and_bot_services():
     training_settlement_block = compose.split(
         "  polyweather_training_settlement:",
         1,
+    )[1].split(
+        "\n  polyweather_weathernext2_worker:",
+        1,
+    )[0]
+    weathernext2_block = compose.split(
+        "  polyweather_weathernext2_worker:",
+        1,
     )[1].split(
         "\nx-polyweather-base:",
         1,
@@ -110,6 +117,7 @@ def test_docker_compose_isolates_collector_from_web_and_bot_services():
     assert "POLYWEATHER_SERVICE_ROLE: collector" in collector_block
     assert "POLYWEATHER_SERVICE_ROLE: warmer" in warmer_block
     assert "POLYWEATHER_SERVICE_ROLE: training_settlement" in training_settlement_block
+    assert "POLYWEATHER_SERVICE_ROLE: weathernext2_worker" in weathernext2_block
     assert "redis-server --appendonly yes --maxmemory ${POLYWEATHER_REDIS_MAXMEMORY:-512mb} --maxmemory-policy noeviction" in compose
     assert "POLYWEATHER_SCAN_TERMINAL_PREWARM_ENABLED: 'false'" in bot_block
     assert "POLYWEATHER_EVENT_STORE: ${POLYWEATHER_EVENT_STORE:-redis}" in web_block
@@ -126,7 +134,9 @@ def test_docker_compose_isolates_collector_from_web_and_bot_services():
     assert "POLYWEATHER_OBSERVATION_COLLECTOR_ENABLED: 'true'" in collector_block
     assert "POLYWEATHER_OBSERVATION_COLLECTOR_ENABLED: 'false'" in warmer_block
     assert "POLYWEATHER_OBSERVATION_COLLECTOR_ENABLED: 'false'" in training_settlement_block
+    assert "POLYWEATHER_OBSERVATION_COLLECTOR_ENABLED: 'false'" in weathernext2_block
     assert "command: python -m web.training_settlement_worker" in training_settlement_block
+    assert "command: python -m web.weathernext2_worker" in weathernext2_block
     assert (
         "POLYWEATHER_TRAINING_SETTLEMENT_INTERVAL_SEC: "
         "${POLYWEATHER_TRAINING_SETTLEMENT_INTERVAL_SEC:-21600}"
@@ -142,6 +152,25 @@ def test_docker_compose_isolates_collector_from_web_and_bot_services():
     assert "POLYWEATHER_CITY_DETAIL_BATCH_QUEUE_WAIT_MS: ${POLYWEATHER_CITY_DETAIL_BATCH_QUEUE_WAIT_MS:-3000}" in web_block
     assert "POLYWEATHER_CITY_DETAIL_BATCH_PARTIAL_TIMEOUT_MS: ${POLYWEATHER_CITY_DETAIL_BATCH_PARTIAL_TIMEOUT_MS:-8000}" in web_block
     assert "UVICORN_WORKERS: ${UVICORN_WORKERS:-2}" in web_block
+    assert "WEATHERNEXT2_ENABLED: ${WEATHERNEXT2_ENABLED:-1}" in web_block
+    assert "WEATHERNEXT2_ENABLED: ${WEATHERNEXT2_ENABLED:-1}" in weathernext2_block
+    assert "WEATHERNEXT2_BACKEND: ${WEATHERNEXT2_BACKEND:-gcs_zarr}" in weathernext2_block
+    assert (
+        "WEATHERNEXT2_GCS_ZARR_URI: "
+        "${WEATHERNEXT2_GCS_ZARR_URI:-gs://weathernext/weathernext_2_0_0/zarr}"
+        in weathernext2_block
+    )
+    assert (
+        "WEATHERNEXT2_MODEL_DIR: "
+        "${WEATHERNEXT2_MODEL_DIR:-/app/data/models/weathernext2_calibrator}"
+        in weathernext2_block
+    )
+    assert (
+        "GOOGLE_APPLICATION_CREDENTIALS: "
+        "${GOOGLE_APPLICATION_CREDENTIALS:-/app/secrets/gcp-sa.json}"
+        in weathernext2_block
+    )
+    assert "./secrets:/app/secrets:ro" in weathernext2_block
     assert "POLYWEATHER_COLLECTOR_PATCH_ENDPOINT: ''" in bot_block
     assert "POLYWEATHER_COLLECTOR_PATCH_ENDPOINT: ''" in web_block
     assert (
@@ -312,6 +341,7 @@ def test_deploy_script_retries_compose_recreate_races():
     assert 'compose_up_retry "observation collector" -d --no-deps polyweather_collector' in script
     assert 'compose_up_retry "cache warmer" -d --no-deps polyweather_warmer' in script
     assert 'compose_up_retry "training settlement" -d --no-deps polyweather_training_settlement' in script
+    assert 'compose_up_retry "WeatherNext2 worker" -d --no-deps polyweather_weathernext2_worker' in script
     assert 'compose_up_retry "frontend" -d --no-deps polyweather_frontend' in script
 
 

tests/test_trend_engine.py

@@ -241,6 +241,64 @@ class TestMuCalculation:
         assert sd["peak_hours"] == ["15:00", "16:00"]
         assert sd["peak_status"] == "before"
 
+    @patch("src.analysis.trend_engine.calculate_dynamic_weights", return_value=(None, ""))
+    @patch("src.analysis.trend_engine.get_deb_accuracy", return_value=None)
+    @patch("src.analysis.trend_engine.update_daily_record")
+    def test_weathernext2_probability_replaces_legacy_distribution(
+        self, _udr, _deb_acc, _dw
+    ):
+        data = _make_weather_data(
+            cur_temp=30.0,
+            max_so_far=31.0,
+            om_today_high=33.0,
+            ens_median=32.0,
+            local_time="2026-03-04 11:00",
+        )
+        data["weathernext2"] = {
+            "summary": {"median": 33.1},
+            "buckets": [
+                {
+                    "label": "32°C",
+                    "value": 32.1,
+                    "lower": 31.5,
+                    "upper": 32.5,
+                    "probability": 0.41,
+                },
+                {
+                    "label": "33°C",
+                    "value": 33.0,
+                    "lower": 32.5,
+                    "upper": 33.5,
+                    "probability": 0.56,
+                },
+            ],
+        }
+
+        _, ai_context, sd = analyze_weather_trend(data, "°C", "test_city")
+
+        assert sd["probability_engine"] == "weathernext2"
+        assert sd["probabilities_all"] == data["weathernext2"]["buckets"]
+        assert sd["mu"] == 33.1
+        assert "WeatherNext 2 概率分布" in ai_context
+
+    @patch("src.analysis.trend_engine.calculate_dynamic_weights", return_value=(None, ""))
+    @patch("src.analysis.trend_engine.get_deb_accuracy", return_value=None)
+    @patch("src.analysis.trend_engine.update_daily_record")
+    def test_weathernext2_median_enters_current_forecasts(
+        self, _udr, _deb_acc, _dw
+    ):
+        data = _make_weather_data(local_time="2026-03-04 11:00")
+        data["weathernext2"] = {
+            "summary": {"median": 33.1},
+            "buckets": [
+                {"label": "33°C", "value": 33.0, "probability": 0.56},
+            ],
+        }
+
+        _, _, sd = analyze_weather_trend(data, "°C", "test_city")
+
+        assert sd["current_forecasts"]["WeatherNext 2"] == 33.1
+
 
 class TestDebEnsembleSignal:
     @patch(

tests/test_weathernext2_calibration.py

@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+from src.analysis.weathernext2_calibration import (
+    apply_quantile_calibration_to_payload,
+    train_lightgbm_quantile_calibrator,
+)
+from src.data_collection.weathernext2_sources import build_weathernext2_city_probability
+
+
+def _synthetic_training_rows(count: int = 170):
+    rows = []
+    for idx in range(count):
+        city = "houston" if idx % 2 == 0 else "shanghai"
+        median = 30.0 + (idx % 7) * 0.2
+        residual = 1.0 if city == "houston" else -0.5
+        rows.append(
+            {
+                "city": city,
+                "target_date": f"2026-05-{idx % 28 + 1:02d}",
+                "actual_high_c": median + residual,
+                "weathernext2": {
+                    "summary": {
+                        "mean": median,
+                        "median": median,
+                        "p10": median - 1.0,
+                        "p25": median - 0.5,
+                        "p75": median + 0.5,
+                        "p90": median + 1.0,
+                        "spread": 2.0,
+                    }
+                },
+                "deb_prediction_c": median + 0.2,
+                "model_median_c": median + 0.1,
+                "model_spread": 1.4,
+                "current_max_so_far_c": median - 2.0,
+                "local_hour": 12,
+            }
+        )
+    return rows
+
+
+def test_lightgbm_quantile_calibrator_trains_and_saves_ordered_quantiles(tmp_path):
+    result = train_lightgbm_quantile_calibrator(
+        _synthetic_training_rows(),
+        model_dir=tmp_path,
+        min_global_samples=150,
+        min_city_samples=5,
+    )
+
+    assert result["trained"] is True
+    assert result["samples"] == 170
+    assert (tmp_path / "metadata.json").is_file()
+    assert (tmp_path / "q10.pkl").is_file()
+    assert (tmp_path / "q50.pkl").is_file()
+    assert (tmp_path / "q90.pkl").is_file()
+    assert result["validation"]["ordered_quantiles"] is True
+
+
+def test_lightgbm_quantile_calibrator_skips_when_samples_are_insufficient(tmp_path):
+    result = train_lightgbm_quantile_calibrator(
+        _synthetic_training_rows(20),
+        model_dir=tmp_path,
+        min_global_samples=150,
+        min_city_samples=5,
+    )
+
+    assert result["trained"] is False
+    assert result["reason"] == "insufficient_global_samples"
+
+
+def test_calibrated_distribution_rebuilds_market_buckets_from_shifted_members(tmp_path):
+    train_lightgbm_quantile_calibrator(
+        _synthetic_training_rows(),
+        model_dir=tmp_path,
+        min_global_samples=150,
+        min_city_samples=5,
+    )
+    raw = build_weathernext2_city_probability(
+        city="houston",
+        member_highs=[30.0, 30.2, 30.4, 30.6],
+        temp_symbol="°C",
+        target_date="2026-06-29",
+    )
+
+    calibrated = apply_quantile_calibration_to_payload(raw, model_dir=tmp_path)
+
+    assert calibrated["calibration"]["engine"] == "lightgbm_quantile"
+    assert calibrated["calibration"]["samples"] == 170
+    assert calibrated["calibration"]["raw_summary"]["median"] == raw["summary"]["median"]
+    assert calibrated["calibration"]["calibrated_summary"]["median"] > raw["summary"]["median"]
+    assert calibrated["buckets"]
+    assert calibrated["top_bucket"]["label"].endswith("°C")

tests/test_weathernext2_sources.py

@@ -0,0 +1,259 @@
+from datetime import datetime, timezone
+import json
+
+from src.data_collection.weather_sources import WeatherDataCollector
+from src.data_collection.weathernext2_fetcher import (
+    extract_member_hourly_from_grid_dataset,
+    normalize_temperature_value,
+    open_weathernext2_zarr_dataset,
+    select_temperature_variable,
+)
+from src.data_collection.weathernext2_sources import (
+    build_city_local_daily_highs_from_hourly,
+    build_weathernext2_city_probability,
+    market_bucket_for_temperature,
+    summarize_member_highs,
+)
+
+
+def test_market_bucket_for_temperature_uses_single_celsius_options():
+    bucket = market_bucket_for_temperature(32.6, "°C")
+
+    assert bucket["label"] == "33°C"
+    assert bucket["lower"] == 32.5
+    assert bucket["upper"] == 33.5
+
+
+def test_market_bucket_for_temperature_groups_fahrenheit_by_two_degree_market_option():
+    assert market_bucket_for_temperature(94.2, "°F")["label"] == "94-95°F"
+    assert market_bucket_for_temperature(95.4, "°F")["label"] == "94-95°F"
+    assert market_bucket_for_temperature(96.1, "°F")["label"] == "96-97°F"
+
+
+def test_weathernext2_probability_aggregates_member_highs_to_market_buckets():
+    probability = build_weathernext2_city_probability(
+        city="Houston",
+        member_highs={
+            "member_00": 94.1,
+            "member_01": 94.8,
+            "member_02": 95.2,
+            "member_03": 96.7,
+        },
+        temp_symbol="°F",
+        target_date="2026-06-29",
+        source_run="2026-06-29T00:00:00Z",
+    )
+
+    assert probability["source"] == "weathernext2"
+    assert probability["members"] == 4
+    assert probability["top_bucket"]["label"] == "94-95°F"
+    assert probability["top_bucket"]["probability"] == 0.75
+    assert probability["buckets"] == [
+        {
+            "key": "94-95°F",
+            "label": "94-95°F",
+            "lower": 93.5,
+            "upper": 95.5,
+            "value": 94.7,
+            "probability": 0.75,
+            "member_count": 3,
+            "total_members": 4,
+        },
+        {
+            "key": "96-97°F",
+            "label": "96-97°F",
+            "lower": 95.5,
+            "upper": 97.5,
+            "value": 96.7,
+            "probability": 0.25,
+            "member_count": 1,
+            "total_members": 4,
+        },
+    ]
+
+
+def test_weathernext2_probability_keeps_celsius_market_option_labels():
+    probability = build_weathernext2_city_probability(
+        city="Shanghai",
+        member_highs=[31.6, 32.2, 32.8, 33.1],
+        temp_symbol="°C",
+        target_date="2026-06-29",
+    )
+
+    labels = [bucket["label"] for bucket in probability["buckets"]]
+
+    assert labels == ["32°C", "33°C"]
+    assert probability["top_bucket"]["label"] == "33°C"
+    assert probability["top_bucket"]["probability"] == 0.5
+
+
+def test_city_local_daily_highs_from_hourly_uses_target_local_date():
+    highs = build_city_local_daily_highs_from_hourly(
+        member_hourly={
+            "member_00": [21.0, 26.0, 28.0, 24.0],
+            "member_01": [20.0, 25.0, None, 27.0],
+        },
+        utc_times=[
+            datetime(2026, 6, 28, 15, tzinfo=timezone.utc),  # local 2026-06-28 23:00
+            datetime(2026, 6, 28, 16, tzinfo=timezone.utc),  # local 2026-06-29 00:00
+            datetime(2026, 6, 29, 6, tzinfo=timezone.utc),  # local 2026-06-29 14:00
+            datetime(2026, 6, 29, 16, tzinfo=timezone.utc),  # local 2026-06-30 00:00
+        ],
+        timezone_offset_seconds=8 * 3600,
+        target_local_date="2026-06-29",
+    )
+
+    assert highs == {"member_00": 28.0, "member_01": 25.0}
+
+
+def test_summarize_member_highs_reports_ensemble_shape():
+    summary = summarize_member_highs([30, 32, 34, 36, 38])
+
+    assert summary == {
+        "members": 5,
+        "mean": 34.0,
+        "median": 34.0,
+        "p10": 30.8,
+        "p25": 32.0,
+        "p75": 36.0,
+        "p90": 37.2,
+        "min": 30.0,
+        "max": 38.0,
+        "spread": 8.0,
+    }
+
+
+def test_collector_reads_weathernext2_fixture_when_enabled(monkeypatch, tmp_path):
+    fixture_path = tmp_path / "weathernext2.json"
+    fixture_path.write_text(
+        json.dumps(
+            {
+                "houston": {
+                    "target_date": "2026-06-29",
+                    "source_run": "2026-06-29T00:00:00Z",
+                    "member_highs": [94.1, 94.8, 95.2, 96.7],
+                }
+            }
+        ),
+        encoding="utf-8",
+    )
+    monkeypatch.setenv("WEATHERNEXT2_ENABLED", "1")
+    monkeypatch.setenv("WEATHERNEXT2_FIXTURE_PATH", str(fixture_path))
+
+    collector = WeatherDataCollector({})
+
+    payload = collector.fetch_weathernext2_probability(
+        "Houston",
+        lat=29.7604,
+        lon=-95.3698,
+        use_fahrenheit=True,
+        target_date="2026-06-29",
+        timezone_offset_seconds=-5 * 3600,
+    )
+
+    assert payload["source"] == "weathernext2"
+    assert payload["top_bucket"]["label"] == "94-95°F"
+    assert payload["buckets"][0]["probability"] == 0.75
+
+
+def test_collector_reads_weathernext2_worker_artifact_before_fixture(monkeypatch, tmp_path):
+    artifact_path = tmp_path / "weathernext2_city_highs.json"
+    fixture_path = tmp_path / "fixture.json"
+    artifact_path.write_text(
+        json.dumps(
+            {
+                "schema_version": 1,
+                "cities": {
+                    "houston": {
+                        "target_date": "2026-06-29",
+                        "source_run": "2026-06-29T00:00:00Z",
+                        "member_highs": [94.1, 94.8, 95.2, 96.7],
+                    }
+                },
+            }
+        ),
+        encoding="utf-8",
+    )
+    fixture_path.write_text(
+        json.dumps({"houston": {"member_highs": [80.0, 80.2]}}),
+        encoding="utf-8",
+    )
+    monkeypatch.setenv("WEATHERNEXT2_ENABLED", "1")
+    monkeypatch.setenv("WEATHERNEXT2_CITY_HIGHS_PATH", str(artifact_path))
+    monkeypatch.setenv("WEATHERNEXT2_FIXTURE_PATH", str(fixture_path))
+    monkeypatch.setenv("WEATHERNEXT2_MODEL_DIR", str(tmp_path / "missing-model"))
+
+    collector = WeatherDataCollector({})
+
+    payload = collector.fetch_weathernext2_probability(
+        "Houston",
+        lat=29.7604,
+        lon=-95.3698,
+        use_fahrenheit=True,
+        target_date="2026-06-29",
+        timezone_offset_seconds=-5 * 3600,
+    )
+
+    assert payload["source"] == "weathernext2"
+    assert payload["top_bucket"]["label"] == "94-95°F"
+    assert payload["buckets"][0]["probability"] == 0.75
+
+
+def test_weathernext2_grid_extractor_selects_temp_var_nearest_point_and_converts_kelvin():
+    dataset = {
+        "lat": [10.0, 20.0],
+        "lon": [30.0, 40.0],
+        "member": [0, 1],
+        "time": [
+            "2026-06-29T00:00:00Z",
+            "2026-06-29T06:00:00Z",
+            "2026-06-29T12:00:00Z",
+        ],
+        "temperature_2m": [
+            [
+                [[290.0, 291.0], [292.0, 293.0]],
+                [[294.0, 295.0], [296.0, 297.0]],
+                [[298.0, 299.0], [300.0, 301.0]],
+            ],
+            [
+                [[289.0, 290.0], [291.0, 292.0]],
+                [[293.0, 294.0], [295.0, 296.0]],
+                [[297.0, 298.0], [299.0, 300.0]],
+            ],
+        ],
+        "units": {"temperature_2m": "K"},
+    }
+
+    assert select_temperature_variable(dataset) == "temperature_2m"
+    assert normalize_temperature_value(300.15, "K") == 27.0
+
+    extracted = extract_member_hourly_from_grid_dataset(
+        dataset,
+        lat=18.0,
+        lon=38.0,
+    )
+
+    assert extracted["temp_var"] == "temperature_2m"
+    assert extracted["nearest_lat"] == 20.0
+    assert extracted["nearest_lon"] == 40.0
+    assert extracted["member_hourly"]["member_00"] == [19.9, 23.9, 27.9]
+    assert extracted["member_hourly"]["member_01"] == [18.9, 22.9, 26.9]
+
+
+def test_weathernext2_zarr_uses_google_default_credentials(monkeypatch):
+    calls = []
+
+    class FakeXarray:
+        @staticmethod
+        def open_zarr(uri, **kwargs):
+            calls.append((uri, kwargs))
+            return {"ok": True}
+
+    import src.data_collection.weathernext2_fetcher as fetcher
+
+    monkeypatch.setattr(fetcher, "xr", FakeXarray, raising=False)
+    monkeypatch.setitem(__import__("sys").modules, "xarray", FakeXarray)
+    monkeypatch.setenv("GOOGLE_APPLICATION_CREDENTIALS", "/app/secrets/gcp-sa.json")
+
+    assert open_weathernext2_zarr_dataset("gs://weathernext/example") == {"ok": True}
+    assert calls[0][1]["storage_options"]["token"] == "google_default"

tests/test_weathernext2_worker_service.py

@@ -0,0 +1,108 @@
+from __future__ import annotations
+
+import json
+from datetime import date, timedelta
+
+from src.database.runtime_state import (
+    RuntimeStateDB,
+    TrainingFeatureRecordRepository,
+    TruthRecordRepository,
+)
+from web.weathernext2_worker_service import run_weathernext2_cycle
+
+
+def _seed_training_rows(feature_repo, truth_repo, count: int = 170):
+    for idx in range(count):
+        city = "houston" if idx % 2 == 0 else "shanghai"
+        target_date = (date(2026, 1, 1) + timedelta(days=idx)).isoformat()
+        median = 30.0 + (idx % 7) * 0.2
+        residual = 1.0 if city == "houston" else -0.5
+        feature_repo.upsert_record(
+            city,
+            target_date,
+            {
+                "weathernext2": {
+                    "summary": {
+                        "mean": median,
+                        "median": median,
+                        "p10": median - 1.0,
+                        "p25": median - 0.5,
+                        "p75": median + 0.5,
+                        "p90": median + 1.0,
+                        "spread": 2.0,
+                    }
+                },
+                "deb_prediction": median + 0.2,
+                "model_median_c": median + 0.1,
+                "model_spread": 1.4,
+                "current_max_so_far_c": median - 2.0,
+                "local_hour": 12,
+            },
+        )
+        truth_repo.upsert_truth(
+            city=city,
+            target_date=target_date,
+            actual_high=median + residual,
+            settlement_source="metar",
+            settlement_station_code="TEST",
+            settlement_station_label="TEST",
+            truth_version="test",
+            updated_by="test",
+            is_final=True,
+        )
+
+
+def test_weathernext2_worker_writes_city_artifact_and_trains_calibrator(tmp_path):
+    db = RuntimeStateDB(str(tmp_path / "polyweather.db"))
+    feature_repo = TrainingFeatureRecordRepository(db)
+    truth_repo = TruthRecordRepository(db)
+    _seed_training_rows(feature_repo, truth_repo)
+
+    def fake_fetcher(city, meta, target_date):
+        return {
+            "city": city,
+            "target_date": target_date,
+            "source_run": "2026-06-29T00:00:00Z",
+            "member_highs": [30.0, 30.2, 30.4, 30.6],
+        }
+
+    output_path = tmp_path / "weathernext2_city_highs.json"
+    model_dir = tmp_path / "model"
+    result = run_weathernext2_cycle(
+        city_registry={
+            "houston": {"name": "Houston", "lat": 29.7, "lon": -95.3, "use_fahrenheit": False, "tz_offset": -5 * 3600},
+        },
+        output_path=str(output_path),
+        model_dir=str(model_dir),
+        fetcher=fake_fetcher,
+        feature_repo=feature_repo,
+        truth_repo=truth_repo,
+        min_global_samples=150,
+        min_city_samples=5,
+    )
+
+    assert result["status"] == "written"
+    assert result["city_count"] == 1
+    assert result["calibration"]["trained"] is True
+    assert (model_dir / "metadata.json").is_file()
+    payload = json.loads(output_path.read_text(encoding="utf-8"))
+    assert payload["cities"]["houston"]["member_highs"] == [30.0, 30.2, 30.4, 30.6]
+    assert payload["calibration_training"]["samples"] == 170
+
+
+def test_weathernext2_worker_does_not_overwrite_artifact_when_no_city_payloads(tmp_path):
+    db = RuntimeStateDB(str(tmp_path / "polyweather.db"))
+    output_path = tmp_path / "weathernext2_city_highs.json"
+    output_path.write_text('{"cities":{"old":{"member_highs":[1]}}}', encoding="utf-8")
+
+    result = run_weathernext2_cycle(
+        city_registry={"houston": {"name": "Houston", "lat": 29.7, "lon": -95.3}},
+        output_path=str(output_path),
+        model_dir=str(tmp_path / "model"),
+        fetcher=lambda _city, _meta, _target_date: None,
+        feature_repo=TrainingFeatureRecordRepository(db),
+        truth_repo=TruthRecordRepository(db),
+    )
+
+    assert result["status"] == "no_city_payloads"
+    assert json.loads(output_path.read_text(encoding="utf-8"))["cities"]["old"]["member_highs"] == [1]

web/scan_terminal_city_row.py

@@ -78,6 +78,34 @@ def _city_local_time(city: str, utc_offset_seconds: Optional[int] = None) -> str
     return _city_local_now(city, utc_offset_seconds).strftime("%H:%M")
 
 
+def _model_sources_with_weathernext2(
+    base_models: Any,
+    data: Dict[str, Any],
+) -> Dict[str, Any]:
+    models = (
+        {
+            str(k): v
+            for k, v in (base_models or {}).items()
+            if v is not None
+        }
+        if isinstance(base_models, dict)
+        else {}
+    )
+    weathernext2 = data.get("weathernext2")
+    if isinstance(weathernext2, dict):
+        summary = (
+            weathernext2.get("summary")
+            if isinstance(weathernext2.get("summary"), dict)
+            else {}
+        )
+        representative = _safe_float(summary.get("median"))
+        if representative is None:
+            representative = _safe_float(summary.get("mean"))
+        if representative is not None:
+            models["WeatherNext 2"] = round(float(representative), 1)
+    return models
+
+
 def _panel_cache_stale_reason(city: str, cached_entry: Dict[str, Any], payload: Dict[str, Any]) -> Optional[str]:
     updated_at_ts = _safe_float(cached_entry.get("updated_at_ts"))
     if updated_at_ts is None or time.time() - updated_at_ts > SCAN_PANEL_CACHE_MAX_AGE_SEC:
@@ -611,7 +639,12 @@ def _build_terminal_row(
         "distribution_full": scan.get("distribution_full") or scan.get("distribution_preview") or row.get("distribution_preview") or [],
         "probability_engine": scan.get("probability_engine") or (data.get("probabilities") or {}).get("engine"),
         "probability_calibration_mode": scan.get("probability_calibration_mode") or (data.get("probabilities") or {}).get("calibration_mode"),
-        "model_cluster_sources": daily_entry.get("models") if isinstance(daily_entry.get("models"), dict) else data.get("multi_model", {}).get("forecasts"),
+        "model_cluster_sources": _model_sources_with_weathernext2(
+            daily_entry.get("models")
+            if isinstance(daily_entry.get("models"), dict)
+            else data.get("multi_model", {}).get("forecasts"),
+            data,
+        ),
         "window_phase": row.get("window_phase") or scan.get("window_phase"),
         "window_score": row.get("window_score") if row.get("window_score") is not None else scan.get("window_score"),
         "signal_status": scan.get("signal_status"),
@@ -744,13 +777,11 @@ def _build_quick_row(
         "network_provider": data.get("official_network_source") or official_status.get("provider_code"),
         "network_provider_label": official_status.get("provider_label"),
         "deb_prediction": deb.get("prediction"),
-        "model_cluster_sources": (
+        "model_cluster_sources": _model_sources_with_weathernext2(
             daily_entry.get("models")
             if isinstance(daily_entry.get("models"), dict)
-            else {
+            else multi.get("forecasts", {}),
-                str(k): v for k, v in multi.get("forecasts", {}).items()
+            data,
-                if v is not None
-            }
         ),
         "distribution_preview": distribution[:6] if distribution else [],
         "distribution_full": probs.get("distribution_all") or distribution,

web/weathernext2_worker.py

@@ -0,0 +1,91 @@
+"""Standalone WeatherNext 2 offline worker."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import signal
+import threading
+import time
+from types import FrameType
+from typing import Optional
+
+from loguru import logger
+
+from web.weathernext2_worker_service import run_weathernext2_cycle
+
+
+_STOP_EVENT = threading.Event()
+
+
+def _env_int(name: str, default: int) -> int:
+    raw = os.getenv(name)
+    if raw is None:
+        return default
+    try:
+        return int(raw)
+    except Exception:
+        return default
+
+
+def _handle_stop_signal(signum: int, _frame: Optional[FrameType]) -> None:
+    logger.info("WeatherNext2 worker stopping signal={}", signum)
+    _STOP_EVENT.set()
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Fetch WeatherNext 2 city highs and train LightGBM calibration."
+    )
+    parser.add_argument("--once", action="store_true", help="Run one cycle and exit.")
+    parser.add_argument(
+        "--interval-sec",
+        type=int,
+        default=_env_int("WEATHERNEXT2_WORKER_INTERVAL_SEC", 21600),
+    )
+    parser.add_argument(
+        "--initial-delay-sec",
+        type=int,
+        default=_env_int("WEATHERNEXT2_WORKER_INITIAL_DELAY_SEC", 90),
+    )
+    parser.add_argument("--cities", nargs="*", default=None)
+    return parser.parse_args()
+
+
+def _run_once(*, cities: Optional[list[str]]) -> dict:
+    result = run_weathernext2_cycle(cities=cities)
+    logger.info("WeatherNext2 worker result={}", json.dumps(result, ensure_ascii=False))
+    return result
+
+
+def main() -> None:
+    signal.signal(signal.SIGINT, _handle_stop_signal)
+    signal.signal(signal.SIGTERM, _handle_stop_signal)
+    args = _parse_args()
+
+    if args.once:
+        result = _run_once(cities=args.cities)
+        print(json.dumps(result, ensure_ascii=False))
+        return
+
+    interval_sec = max(1800, int(args.interval_sec or 21600))
+    initial_delay_sec = max(0, int(args.initial_delay_sec or 0))
+    logger.info("WeatherNext2 worker started interval={}s", interval_sec)
+    if initial_delay_sec and _STOP_EVENT.wait(initial_delay_sec):
+        return
+
+    while not _STOP_EVENT.is_set():
+        started = time.time()
+        try:
+            _run_once(cities=args.cities)
+        except Exception as exc:
+            logger.exception("WeatherNext2 cycle failed: {}", exc)
+        elapsed = time.time() - started
+        wait_for = max(5.0, interval_sec - elapsed)
+        if _STOP_EVENT.wait(wait_for):
+            break
+
+
+if __name__ == "__main__":
+    main()

web/weathernext2_worker_service.py

@@ -0,0 +1,288 @@
+"""Offline WeatherNext 2 fetch and calibration worker service."""
+
+from __future__ import annotations
+
+import json
+import os
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Callable, Dict, Iterable, Mapping, Optional, Sequence
+
+from loguru import logger
+
+from src.analysis.weathernext2_calibration import train_lightgbm_quantile_calibrator
+from src.data_collection.city_registry import CITY_REGISTRY
+from src.data_collection.weathernext2_fetcher import (
+    extract_member_hourly_from_grid_dataset,
+    open_weathernext2_zarr_dataset,
+)
+from src.data_collection.weathernext2_sources import (
+    WEATHERNEXT2_GCS_ZARR_URI,
+    build_city_local_daily_highs_from_hourly,
+    build_weathernext2_city_probability,
+)
+from src.database.runtime_state import (
+    TrainingFeatureRecordRepository,
+    TruthRecordRepository,
+)
+
+
+CityFetcher = Callable[[str, Mapping[str, Any], str], Optional[Mapping[str, Any]]]
+
+
+def _env_int(name: str, default: int) -> int:
+    raw = os.getenv(name)
+    if raw is None:
+        return default
+    try:
+        return int(raw)
+    except Exception:
+        return default
+
+
+def _normalize_city(city: str) -> str:
+    return str(city or "").strip().lower().replace("-", " ")
+
+
+def _selected_city_names(
+    city_registry: Mapping[str, Mapping[str, Any]],
+    cities: Optional[Iterable[str]],
+) -> Sequence[str]:
+    selected = {_normalize_city(city) for city in (cities or []) if _normalize_city(city)}
+    names = []
+    for city in sorted(city_registry.keys()):
+        normalized = _normalize_city(city)
+        if selected and normalized not in selected:
+            continue
+        names.append(normalized)
+    return tuple(names)
+
+
+def _city_target_date(city_meta: Mapping[str, Any]) -> str:
+    offset_seconds = int(city_meta.get("tz_offset") or 0)
+    now_utc = datetime.now(timezone.utc).timestamp()
+    return datetime.fromtimestamp(now_utc + offset_seconds, tz=timezone.utc).date().isoformat()
+
+
+def _safe_float(value: Any) -> Optional[float]:
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if parsed == parsed else None
+
+
+def _dataset_source_run(dataset: Any) -> Optional[str]:
+    attrs = getattr(dataset, "attrs", None)
+    if isinstance(attrs, Mapping):
+        for key in ("source_run", "forecast_reference_time", "init_time", "run_time"):
+            value = attrs.get(key)
+            if value:
+                return str(value)
+    for key in ("init_time", "forecast_reference_time", "time"):
+        try:
+            values = dataset[key].values
+        except Exception:
+            try:
+                values = dataset[key]
+            except Exception:
+                continue
+        try:
+            first = values[-1]
+        except Exception:
+            first = values
+        if first is not None:
+            return str(first)
+    return None
+
+
+def _default_city_fetcher(dataset: Any, temp_var: Optional[str]) -> CityFetcher:
+    source_run = _dataset_source_run(dataset)
+
+    def fetch(city: str, meta: Mapping[str, Any], target_date: str) -> Optional[Mapping[str, Any]]:
+        lat = _safe_float(meta.get("lat"))
+        lon = _safe_float(meta.get("lon"))
+        if lat is None or lon is None:
+            return None
+        use_fahrenheit = bool(meta.get("use_fahrenheit") or meta.get("f"))
+        extracted = extract_member_hourly_from_grid_dataset(
+            dataset,
+            lat=lat,
+            lon=lon,
+            temp_var=temp_var,
+            use_fahrenheit=use_fahrenheit,
+        )
+        member_highs = build_city_local_daily_highs_from_hourly(
+            extracted["member_hourly"],
+            extracted["utc_times"],
+            int(meta.get("tz_offset") or 0),
+            target_date,
+        )
+        if not member_highs:
+            return None
+        payload = build_weathernext2_city_probability(
+            city=city,
+            member_highs=member_highs,
+            temp_symbol="°F" if use_fahrenheit else "°C",
+            target_date=target_date,
+            source_run=source_run,
+        )
+        payload["fetch_metadata"] = {
+            "backend": "gcs_zarr",
+            "temp_var": extracted.get("temp_var"),
+            "units": extracted.get("units"),
+            "nearest_lat": extracted.get("nearest_lat"),
+            "nearest_lon": extracted.get("nearest_lon"),
+            "requested_lat": lat,
+            "requested_lon": lon,
+        }
+        return payload
+
+    return fetch
+
+
+def _atomic_write_json(path: Path, payload: Mapping[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp_path = path.with_suffix(f"{path.suffix}.tmp")
+    tmp_path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+    os.replace(str(tmp_path), str(path))
+
+
+def _flatten_training_records_for_weathernext2(
+    *,
+    feature_repo: Optional[TrainingFeatureRecordRepository] = None,
+    truth_repo: Optional[TruthRecordRepository] = None,
+) -> list[Dict[str, Any]]:
+    features = (feature_repo or TrainingFeatureRecordRepository()).load_all()
+    truth = (truth_repo or TruthRecordRepository()).load_all()
+    rows: list[Dict[str, Any]] = []
+    for city, by_date in features.items():
+        city_truth = truth.get(city, {})
+        for target_date, payload in by_date.items():
+            if not isinstance(payload, dict):
+                continue
+            actual = payload.get("actual_high")
+            if actual is None and isinstance(city_truth.get(target_date), dict):
+                actual = city_truth[target_date].get("actual_high")
+            if actual is None:
+                continue
+            row = dict(payload)
+            row["city"] = city
+            row["target_date"] = target_date
+            row["actual_high_c"] = actual
+            if "deb_prediction_c" not in row and row.get("deb_prediction") is not None:
+                row["deb_prediction_c"] = row.get("deb_prediction")
+            if "weathernext2" not in row and isinstance(row.get("weather_data"), dict):
+                wn2 = row["weather_data"].get("weathernext2")
+                if isinstance(wn2, dict):
+                    row["weathernext2"] = wn2
+            rows.append(row)
+    return rows
+
+
+def run_weathernext2_cycle(
+    *,
+    city_registry: Optional[Mapping[str, Mapping[str, Any]]] = None,
+    cities: Optional[Iterable[str]] = None,
+    output_path: Optional[str] = None,
+    model_dir: Optional[str] = None,
+    fetcher: Optional[CityFetcher] = None,
+    dataset: Any = None,
+    feature_repo: Optional[TrainingFeatureRecordRepository] = None,
+    truth_repo: Optional[TruthRecordRepository] = None,
+    min_global_samples: Optional[int] = None,
+    min_city_samples: Optional[int] = None,
+) -> Dict[str, Any]:
+    registry = city_registry or CITY_REGISTRY
+    data_root = Path(os.getenv("WEATHERNEXT2_DATA_ROOT", "/app/data/weathernext2"))
+    out_path = Path(output_path or os.getenv("WEATHERNEXT2_CITY_HIGHS_PATH", "") or data_root / "weathernext2_city_highs.json")
+    calibrator_dir = Path(model_dir or os.getenv("WEATHERNEXT2_MODEL_DIR", "/app/data/models/weathernext2_calibrator"))
+    selected = _selected_city_names(registry, cities)
+    backend = str(os.getenv("WEATHERNEXT2_BACKEND", "gcs_zarr") or "gcs_zarr").strip().lower()
+
+    generated_at = datetime.now(timezone.utc).isoformat()
+    opened_dataset = dataset
+    if fetcher is None:
+        if opened_dataset is None:
+            if backend != "gcs_zarr":
+                return {"ok": False, "status": "skipped", "reason": "unsupported_backend", "backend": backend}
+            uri = os.getenv("WEATHERNEXT2_GCS_ZARR_URI", WEATHERNEXT2_GCS_ZARR_URI)
+            opened_dataset = open_weathernext2_zarr_dataset(uri)
+        fetcher = _default_city_fetcher(
+            opened_dataset,
+            str(os.getenv("WEATHERNEXT2_TEMP_VAR", "") or "").strip() or None,
+        )
+
+    city_payloads: Dict[str, Dict[str, Any]] = {}
+    failed = 0
+    items = []
+    for city in selected:
+        meta = registry.get(city) or {}
+        target_date = _city_target_date(meta)
+        try:
+            payload = fetcher(city, meta, target_date)
+            if not isinstance(payload, Mapping):
+                items.append({"city": city, "ok": True, "status": "empty", "target_date": target_date})
+                continue
+            city_payloads[city] = dict(payload)
+            items.append({"city": city, "ok": True, "status": "written", "target_date": target_date})
+        except Exception as exc:
+            failed += 1
+            logger.warning("WeatherNext2 city fetch failed city={}: {}", city, exc)
+            items.append({"city": city, "ok": False, "status": "failed", "target_date": target_date, "error": str(exc)})
+
+    training_rows = _flatten_training_records_for_weathernext2(
+        feature_repo=feature_repo,
+        truth_repo=truth_repo,
+    )
+    calibration = train_lightgbm_quantile_calibrator(
+        training_rows,
+        model_dir=calibrator_dir,
+        min_global_samples=min_global_samples
+        if min_global_samples is not None
+        else _env_int("WEATHERNEXT2_MIN_GLOBAL_SAMPLES", 150),
+        min_city_samples=min_city_samples
+        if min_city_samples is not None
+        else _env_int("WEATHERNEXT2_MIN_CITY_SAMPLES", 5),
+    )
+
+    if not city_payloads:
+        return {
+            "ok": failed == 0,
+            "status": "no_city_payloads",
+            "backend": backend,
+            "generated_at": generated_at,
+            "city_count": 0,
+            "failed": failed,
+            "items": items,
+            "calibration": calibration,
+            "output_path": str(out_path),
+        }
+
+    artifact = {
+        "schema_version": 1,
+        "source": "weathernext2",
+        "backend": backend,
+        "gcs_zarr_uri": os.getenv("WEATHERNEXT2_GCS_ZARR_URI", WEATHERNEXT2_GCS_ZARR_URI),
+        "generated_at": generated_at,
+        "city_count": len(city_payloads),
+        "cities": city_payloads,
+        "fetch_metadata": {
+            "selected_city_count": len(selected),
+            "failed": failed,
+        },
+        "calibration_training": calibration,
+    }
+    _atomic_write_json(out_path, artifact)
+
+    return {
+        "ok": failed == 0,
+        "status": "written",
+        "backend": backend,
+        "generated_at": generated_at,
+        "city_count": len(city_payloads),
+        "failed": failed,
+        "items": items,
+        "calibration": calibration,
+        "output_path": str(out_path),
+    }