Finalize SQLite defaults and refresh calibration artifacts

.env.example

@@ -17,7 +17,7 @@ OPEN_METEO_DISK_CACHE_PATH=/var/lib/polyweather/open_meteo_cache.json
 # Windows / macOS can usually keep the defaults.
 UID=1000
 GID=1000
-POLYWEATHER_STATE_STORAGE_MODE=dual
+POLYWEATHER_STATE_STORAGE_MODE=sqlite
 POLYWEATHER_PROMETHEUS_PORT=9090
 POLYWEATHER_ALERTMANAGER_PORT=9093
 POLYWEATHER_ALERT_RELAY_PORT=9099

README.md

@@ -25,7 +25,7 @@ Public docs center: `/docs/intro` on the main site (bilingual product documentat
 - Auto-reconciliation live: event listener + periodic confirm loop.
 - Ops dashboard live: `/ops` for memberships, leaderboard, manual point grants, and payment incident triage.
 - Lightweight observability live: `/healthz`, `/api/system/status`, `/metrics`.
-- Runtime state supports gradual SQLite migration (`file / dual / sqlite`).
+- Runtime state, cache, and core offline training/backfill flows now use SQLite as the primary path; legacy JSON/JSONL files remain only for migration, export, and explicit fallback input.
 - EMOS/CRPS pipeline is integrated in `shadow` mode with rollout gating.
 - Intraday structural signal is now peak-window aware and bilingual (`zh-CN` / `en-US`).
 - Non-Hong Kong airport cities now ingest `TAF` and parse `FM / TEMPO / BECMG / PROB30/40`.
@@ -101,7 +101,8 @@ npm run dev
 
 ## Recent Highlights
 
-- Taipei settlement is aligned to `NOAA RCTP` and rounded whole-degree Celsius logic.
+- Taipei settlement is aligned to `Wunderground RCSS` with whole-degree Celsius resolution logic.
+- Shenzhen settlement is aligned to `Wunderground ZGSZ`.
 - Hong Kong keeps `HKO` official readings in dashboard and history, without falling back to airport METAR lines.
 - Intraday analysis now separates:
   - `Surface Structure`
@@ -117,6 +118,7 @@ Use external runtime storage to avoid SQLite/git conflicts:
 ```env
 POLYWEATHER_RUNTIME_DATA_DIR=/var/lib/polyweather
 POLYWEATHER_DB_PATH=/var/lib/polyweather/polyweather.db
+POLYWEATHER_STATE_STORAGE_MODE=sqlite
 ```
 
 ## Ops Verification

README_ZH.md

@@ -24,7 +24,7 @@
 - 已上线轻量运营后台：`/ops`（会员、周榜、补分、支付异常单）。
 - 已上线轻量可观测性：`/healthz`、`/api/system/status`、`/metrics`。
 - 已补最小外部监控栈：Prometheus + Alertmanager + Grafana + Telegram 告警 relay。
-- 运行态状态与缓存已支持 SQLite 渐进迁移：`file / dual / sqlite`。
+- 运行态状态、缓存与核心离线训练/回填链路已完成 SQLite 主路径收口；legacy JSON/JSONL 仅保留给迁移、导出与显式回退输入。
 - 已接入 EMOS/CRPS 校准链路，但当前仍保持 `emos_shadow`。
 
 ## 许可证与商用边界（重要）
@@ -64,7 +64,7 @@ flowchart LR
     ANA --> PAY["支付状态（Intent + Event + Confirm Loop）"]
     ANA --> PM["Polymarket 只读层"]
     API --> OBS["healthz / system status / metrics"]
-    ANA --> STATE["SQLite runtime state + dual fallback"]
+    ANA --> STATE["SQLite runtime state<br/>legacy files only for migration/export fallback"]
 ```
 
 ## 监控城市（30）
@@ -98,7 +98,7 @@ npm run dev
 ```env
 POLYWEATHER_RUNTIME_DATA_DIR=/var/lib/polyweather
 POLYWEATHER_DB_PATH=/var/lib/polyweather/polyweather.db
-POLYWEATHER_STATE_STORAGE_MODE=dual
+POLYWEATHER_STATE_STORAGE_MODE=sqlite
 ```
 
 ## 运维验收

artifacts/probability_calibration/default.json

@@ -1,20 +1,20 @@
 {
-  "version": "emos-20260320132525",
+  "version": "emos-20260402151301",
-  "trained_at": "2026-03-20T13:25:25.836021+00:00",
+  "trained_at": "2026-04-02T15:13:01.073253+00:00",
   "global": {
     "mu": {
-      "intercept": -1.57406048,
+      "intercept": 1.52725527,
-      "raw_mu_coef": 2.80583627,
+      "raw_mu_coef": 2.96318139,
-      "deb_coef": -0.06819634,
+      "deb_coef": -1.54490032,
-      "ens_median_coef": -1.81560215,
+      "ens_median_coef": -0.71820793,
-      "max_so_far_gap_coef": 0.0
+      "max_so_far_gap_coef": 10.22802132
     },
     "sigma": {
-      "intercept": 0.67509915,
+      "intercept": 0.67227661,
-      "raw_sigma_coef": 0.14431833,
+      "raw_sigma_coef": 0.69444674,
-      "spread_coef": 0.14431833,
+      "spread_coef": 0.09134583,
-      "peak_flag_coef": 0.0,
+      "peak_flag_coef": -0.58201761,
-      "max_so_far_gap_coef": 0.0
+      "max_so_far_gap_coef": -0.81482466
     }
   },
   "sigma_constraints": {
@@ -30,143 +30,59 @@
   },
   "blending": {
     "alpha_mu": 0.0,
-    "alpha_sigma": 0.0
+    "alpha_sigma": 0.45
   },
   "cities": {
     "ankara": {
-      "samples": 7,
+      "samples": 3,
-      "mu_bias": 0.566273,
+      "mu_bias": 1.283918,
-      "sigma_scale": 2.0,
+      "sigma_scale": 1.484056,
-      "confidence": 0.875
+      "confidence": 0.375
-    },
-    "london": {
-      "samples": 6,
-      "mu_bias": 0.489961,
-      "sigma_scale": 2.0,
-      "confidence": 0.75
-    },
-    "new york": {
-      "samples": 4,
-      "mu_bias": 1.852451,
-      "sigma_scale": 0.830444,
-      "confidence": 0.5
-    },
-    "paris": {
-      "samples": 7,
-      "mu_bias": 0.308599,
-      "sigma_scale": 2.0,
-      "confidence": 0.875
-    },
-    "seoul": {
-      "samples": 6,
-      "mu_bias": -1.486024,
-      "sigma_scale": 1.569585,
-      "confidence": 0.75
-    },
-    "toronto": {
-      "samples": 5,
-      "mu_bias": -0.734395,
-      "sigma_scale": 1.656751,
-      "confidence": 0.625
-    },
-    "buenos aires": {
-      "samples": 5,
-      "mu_bias": -1.753334,
-      "sigma_scale": 1.831739,
-      "confidence": 0.625
-    },
-    "wellington": {
-      "samples": 6,
-      "mu_bias": 0.350757,
-      "sigma_scale": 1.377974,
-      "confidence": 0.75
-    },
-    "chicago": {
-      "samples": 4,
-      "mu_bias": 3.01062,
-      "sigma_scale": 0.825575,
-      "confidence": 0.5
-    },
-    "sao paulo": {
-      "samples": 5,
-      "mu_bias": 1.632457,
-      "sigma_scale": 2.0,
-      "confidence": 0.625
-    },
-    "dallas": {
-      "samples": 4,
-      "mu_bias": 3.77714,
-      "sigma_scale": 0.796874,
-      "confidence": 0.5
-    },
-    "miami": {
-      "samples": 5,
-      "mu_bias": -4.868741,
-      "sigma_scale": 2.0,
-      "confidence": 0.625
-    },
-    "atlanta": {
-      "samples": 5,
-      "mu_bias": -7.648823,
-      "sigma_scale": 2.0,
-      "confidence": 0.625
-    },
-    "seattle": {
-      "samples": 4,
-      "mu_bias": 4.058619,
-      "sigma_scale": 2.0,
-      "confidence": 0.5
-    },
-    "lucknow": {
-      "samples": 4,
-      "mu_bias": 3.257609,
-      "sigma_scale": 2.0,
-      "confidence": 0.5
-    },
-    "munich": {
-      "samples": 6,
-      "mu_bias": -0.780811,
-      "sigma_scale": 2.0,
-      "confidence": 0.75
     },
     "hong kong": {
+      "samples": 4,
+      "mu_bias": 1.329042,
+      "sigma_scale": 1.045379,
+      "confidence": 0.5
+    },
+    "milan": {
       "samples": 3,
-      "mu_bias": -0.492675,
+      "mu_bias": -3.932299,
+      "sigma_scale": 2.0,
+      "confidence": 0.375
+    },
+    "shanghai": {
+      "samples": 3,
+      "mu_bias": 1.827618,
       "sigma_scale": 2.0,
       "confidence": 0.375
     },
     "taipei": {
       "samples": 3,
-      "mu_bias": 1.462462,
+      "mu_bias": 3.387896,
-      "sigma_scale": 1.132834,
-      "confidence": 0.375
-    },
-    "milan": {
-      "samples": 3,
-      "mu_bias": -2.729203,
       "sigma_scale": 2.0,
       "confidence": 0.375
     },
     "warsaw": {
       "samples": 3,
-      "mu_bias": 0.349319,
+      "mu_bias": -0.614031,
-      "sigma_scale": 1.337949,
+      "sigma_scale": 1.252706,
       "confidence": 0.375
     }
   },
   "metrics": {
-    "sample_count": 105,
+    "sample_count": 54,
-    "mean_crps": 2.923823,
+    "mean_crps": 3.78558,
-    "legacy_mean_crps": 2.793938,
+    "legacy_mean_crps": 4.326547,
-    "legacy_mean_mae": 2.721143,
+    "legacy_mean_mae": 4.528889,
-    "legacy_bucket_hit_rate": 0.695238,
+    "legacy_bucket_hit_rate": 0.518519,
-    "legacy_bucket_brier": 0.775463,
+    "legacy_bucket_brier": 0.83496,
-    "selected_mean_crps": 2.700275,
+    "selected_mean_crps": 4.239815,
-    "selected_mean_mae": 2.721143,
+    "selected_mean_mae": 4.528889,
-    "selected_bucket_hit_rate": 0.695238,
+    "selected_bucket_hit_rate": 0.537037,
-    "selected_bucket_brier": 0.765459,
+    "selected_bucket_brier": 0.853485,
-    "selected_score": 4.003626,
+    "selected_score": 6.027262,
-    "legacy_score": 4.104792,
+    "legacy_score": 6.127879,
     "filled_actual_from_history": 2,
     "settlement_history_city_count": 30
   },

artifacts/probability_calibration/evaluation_report.json

@@ -1,190 +1,199 @@
 {
   "summary": {
-    "sample_count": 105,
+    "sample_count": 54,
     "filled_actual_from_history": 2,
     "legacy": {
-      "mean_crps": 2.793938,
+      "mean_crps": 4.287017,
-      "mean_mae": 2.721143,
+      "mean_mae": 4.488148,
-      "bucket_hit_rate": 0.695238
+      "bucket_hit_rate": 0.537037
     },
     "emos": {
-      "mean_crps": 2.700275,
+      "mean_crps": 4.201116,
-      "mean_mae": 2.721143,
+      "mean_mae": 4.488148,
-      "bucket_hit_rate": 0.695238
+      "bucket_hit_rate": 0.537037
     },
     "delta": {
-      "crps": -0.093663,
+      "crps": -0.085901,
       "mae": 0.0,
       "bucket_hit_rate": 0.0
     }
   },
   "by_city": {
     "ankara": {
-      "samples": 7,
+      "samples": 3,
-      "legacy_mean_crps": 2.023242,
+      "legacy_mean_crps": 0.327701,
-      "emos_mean_crps": 2.023242,
+      "emos_mean_crps": 0.439705,
-      "legacy_mean_mae": 1.984286,
+      "legacy_mean_mae": 0.066667,
-      "emos_mean_mae": 1.984286,
+      "emos_mean_mae": 0.066667,
-      "legacy_bucket_hit_rate": 0.714286,
+      "legacy_bucket_hit_rate": 1.0,
-      "emos_bucket_hit_rate": 0.714286
+      "emos_bucket_hit_rate": 1.0
     },
     "atlanta": {
-      "samples": 5,
+      "samples": 2,
-      "legacy_mean_crps": 12.792034,
+      "legacy_mean_crps": 30.449382,
-      "emos_mean_crps": 12.694543,
+      "emos_mean_crps": 30.578432,
-      "legacy_mean_mae": 12.806,
+      "legacy_mean_mae": 32.015,
-      "emos_mean_mae": 12.806,
+      "emos_mean_mae": 32.015,
-      "legacy_bucket_hit_rate": 0.6,
+      "legacy_bucket_hit_rate": 0.0,
-      "emos_bucket_hit_rate": 0.6
+      "emos_bucket_hit_rate": 0.0
     },
     "buenos aires": {
-      "samples": 5,
+      "samples": 2,
-      "legacy_mean_crps": 3.846144,
+      "legacy_mean_crps": 9.113412,
-      "emos_mean_crps": 3.846144,
+      "emos_mean_crps": 8.759954,
-      "legacy_mean_mae": 4.168,
+      "legacy_mean_mae": 10.27,
-      "emos_mean_mae": 4.168,
+      "emos_mean_mae": 10.27,
-      "legacy_bucket_hit_rate": 0.6,
+      "legacy_bucket_hit_rate": 0.0,
-      "emos_bucket_hit_rate": 0.6
+      "emos_bucket_hit_rate": 0.0
     },
     "chicago": {
-      "samples": 4,
+      "samples": 1,
-      "legacy_mean_crps": 1.346667,
+      "legacy_mean_crps": 1.250268,
-      "emos_mean_crps": 0.601765,
+      "emos_mean_crps": 0.701085,
       "legacy_mean_mae": 0.0,
       "emos_mean_mae": 0.0,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "dallas": {
-      "samples": 4,
+      "samples": 1,
-      "legacy_mean_crps": 1.256111,
+      "legacy_mean_crps": 2.173363,
-      "emos_mean_crps": 0.651425,
+      "emos_mean_crps": 0.701085,
       "legacy_mean_mae": 0.0,
       "emos_mean_mae": 0.0,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "hong kong": {
-      "samples": 3,
+      "samples": 4,
-      "legacy_mean_crps": 0.261027,
+      "legacy_mean_crps": 0.29509,
-      "emos_mean_crps": 0.261027,
+      "emos_mean_crps": 0.387946,
-      "legacy_mean_mae": 0.1,
+      "legacy_mean_mae": 0.075,
-      "emos_mean_mae": 0.1,
+      "emos_mean_mae": 0.075,
       "legacy_bucket_hit_rate": 1.0,
-      "emos_bucket_hit_rate": 0.666667
+      "emos_bucket_hit_rate": 0.75
     },
     "london": {
-      "samples": 6,
+      "samples": 2,
-      "legacy_mean_crps": 2.079624,
+      "legacy_mean_crps": 3.885033,
-      "emos_mean_crps": 2.079624,
+      "emos_mean_crps": 3.866915,
-      "legacy_mean_mae": 2.451667,
+      "legacy_mean_mae": 4.135,
-      "emos_mean_mae": 2.451667,
+      "emos_mean_mae": 4.135,
-      "legacy_bucket_hit_rate": 0.166667,
+      "legacy_bucket_hit_rate": 0.0,
-      "emos_bucket_hit_rate": 0.166667
+      "emos_bucket_hit_rate": 0.0
     },
     "lucknow": {
-      "samples": 4,
+      "samples": 2,
-      "legacy_mean_crps": 1.468528,
+      "legacy_mean_crps": 2.487193,
-      "emos_mean_crps": 1.468528,
+      "emos_mean_crps": 2.342342,
-      "legacy_mean_mae": 1.6025,
+      "legacy_mean_mae": 3.205,
-      "emos_mean_mae": 1.6025,
+      "emos_mean_mae": 3.205,
-      "legacy_bucket_hit_rate": 0.5,
+      "legacy_bucket_hit_rate": 0.0,
-      "emos_bucket_hit_rate": 0.5
+      "emos_bucket_hit_rate": 0.0
     },
     "madrid": {
       "samples": 2,
       "legacy_mean_crps": 6.27726,
-      "emos_mean_crps": 6.27726,
+      "emos_mean_crps": 5.967277,
       "legacy_mean_mae": 7.33,
       "emos_mean_mae": 7.33,
       "legacy_bucket_hit_rate": 0.0,
       "emos_bucket_hit_rate": 0.0
     },
     "miami": {
-      "samples": 5,
+      "samples": 2,
-      "legacy_mean_crps": 11.665378,
+      "legacy_mean_crps": 28.637631,
-      "emos_mean_crps": 11.665378,
+      "emos_mean_crps": 28.482516,
-      "legacy_mean_mae": 12.07,
+      "legacy_mean_mae": 30.175,
-      "emos_mean_mae": 12.07,
+      "emos_mean_mae": 30.175,
-      "legacy_bucket_hit_rate": 0.6,
+      "legacy_bucket_hit_rate": 0.0,
-      "emos_bucket_hit_rate": 0.6
+      "emos_bucket_hit_rate": 0.0
     },
     "milan": {
       "samples": 3,
       "legacy_mean_crps": 4.401392,
-      "emos_mean_crps": 3.928883,
+      "emos_mean_crps": 3.858031,
       "legacy_mean_mae": 4.06,
       "emos_mean_mae": 4.06,
       "legacy_bucket_hit_rate": 0.666667,
       "emos_bucket_hit_rate": 0.666667
     },
     "munich": {
-      "samples": 6,
+      "samples": 2,
-      "legacy_mean_crps": 2.988583,
+      "legacy_mean_crps": 3.145192,
-      "emos_mean_crps": 2.988583,
+      "emos_mean_crps": 3.011312,
-      "legacy_mean_mae": 3.143333,
+      "legacy_mean_mae": 3.64,
-      "emos_mean_mae": 3.143333,
+      "emos_mean_mae": 3.64,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "new york": {
+      "samples": 1,
+      "legacy_mean_crps": 3.692845,
+      "emos_mean_crps": 3.407357,
+      "legacy_mean_mae": 4.94,
+      "emos_mean_mae": 4.94,
+      "legacy_bucket_hit_rate": 0.0,
+      "emos_bucket_hit_rate": 0.0
+    },
+    "paris": {
+      "samples": 2,
+      "legacy_mean_crps": 4.013782,
+      "emos_mean_crps": 3.979293,
+      "legacy_mean_mae": 4.265,
+      "emos_mean_mae": 4.265,
       "legacy_bucket_hit_rate": 0.5,
       "emos_bucket_hit_rate": 0.5
     },
-    "new york": {
-      "samples": 4,
-      "legacy_mean_crps": 1.861101,
-      "emos_mean_crps": 1.409393,
-      "legacy_mean_mae": 1.3725,
-      "emos_mean_mae": 1.3725,
-      "legacy_bucket_hit_rate": 0.75,
-      "emos_bucket_hit_rate": 0.75
-    },
-    "paris": {
-      "samples": 7,
-      "legacy_mean_crps": 2.430082,
-      "emos_mean_crps": 2.430082,
-      "legacy_mean_mae": 2.518571,
-      "emos_mean_mae": 2.518571,
-      "legacy_bucket_hit_rate": 0.571429,
-      "emos_bucket_hit_rate": 0.571429
-    },
     "sao paulo": {
-      "samples": 5,
+      "samples": 2,
-      "legacy_mean_crps": 2.454756,
+      "legacy_mean_crps": 5.540967,
-      "emos_mean_crps": 2.454756,
+      "emos_mean_crps": 5.272063,
-      "legacy_mean_mae": 2.628,
+      "legacy_mean_mae": 6.57,
-      "emos_mean_mae": 2.628,
+      "emos_mean_mae": 6.57,
-      "legacy_bucket_hit_rate": 0.6,
+      "legacy_bucket_hit_rate": 0.0,
-      "emos_bucket_hit_rate": 0.6
+      "emos_bucket_hit_rate": 0.0
     },
     "seattle": {
-      "samples": 4,
+      "samples": 1,
-      "legacy_mean_crps": 0.531656,
+      "legacy_mean_crps": 0.315488,
-      "emos_mean_crps": 0.452784,
+      "emos_mean_crps": 0.425909,
       "legacy_mean_mae": 0.0,
       "emos_mean_mae": 0.0,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "seoul": {
-      "samples": 6,
+      "samples": 2,
-      "legacy_mean_crps": 0.328088,
+      "legacy_mean_crps": 0.313754,
-      "emos_mean_crps": 0.328088,
+      "emos_mean_crps": 0.412831,
-      "legacy_mean_mae": 0.2,
+      "legacy_mean_mae": 0.15,
-      "emos_mean_mae": 0.2,
+      "emos_mean_mae": 0.15,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "shanghai": {
-      "samples": 2,
+      "samples": 3,
-      "legacy_mean_crps": 0.250034,
+      "legacy_mean_crps": 0.299116,
-      "emos_mean_crps": 0.250034,
+      "emos_mean_crps": 0.394855,
-      "legacy_mean_mae": 0.15,
+      "legacy_mean_mae": 0.1,
-      "emos_mean_mae": 0.15,
+      "emos_mean_mae": 0.1,
+      "legacy_bucket_hit_rate": 1.0,
+      "emos_bucket_hit_rate": 1.0
+    },
+    "shenzhen": {
+      "samples": 1,
+      "legacy_mean_crps": 0.063696,
+      "emos_mean_crps": 0.073016,
+      "legacy_mean_mae": 0.1,
+      "emos_mean_mae": 0.1,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "singapore": {
       "samples": 2,
       "legacy_mean_crps": 0.281993,
-      "emos_mean_crps": 0.281993,
+      "emos_mean_crps": 0.37264,
       "legacy_mean_mae": 0.15,
       "emos_mean_mae": 0.15,
       "legacy_bucket_hit_rate": 1.0,
@@ -193,16 +202,16 @@
     "taipei": {
       "samples": 3,
       "legacy_mean_crps": 0.356996,
-      "emos_mean_crps": 0.356996,
+      "emos_mean_crps": 0.472738,
       "legacy_mean_mae": 0.1,
       "emos_mean_mae": 0.1,
-      "legacy_bucket_hit_rate": 1.0,
+      "legacy_bucket_hit_rate": 0.666667,
-      "emos_bucket_hit_rate": 1.0
+      "emos_bucket_hit_rate": 0.666667
     },
     "tel aviv": {
       "samples": 2,
       "legacy_mean_crps": 0.446758,
-      "emos_mean_crps": 0.446758,
+      "emos_mean_crps": 0.578006,
       "legacy_mean_mae": 0.3,
       "emos_mean_mae": 0.3,
       "legacy_bucket_hit_rate": 1.0,
@@ -211,36 +220,36 @@
     "tokyo": {
       "samples": 2,
       "legacy_mean_crps": 0.450128,
-      "emos_mean_crps": 0.450128,
+      "emos_mean_crps": 0.582151,
       "legacy_mean_mae": 0.25,
       "emos_mean_mae": 0.25,
       "legacy_bucket_hit_rate": 0.5,
       "emos_bucket_hit_rate": 1.0
     },
     "toronto": {
-      "samples": 5,
+      "samples": 2,
-      "legacy_mean_crps": 2.647861,
+      "legacy_mean_crps": 5.497916,
-      "emos_mean_crps": 2.566068,
+      "emos_mean_crps": 5.240552,
-      "legacy_mean_mae": 2.532,
+      "legacy_mean_mae": 6.33,
-      "emos_mean_mae": 2.532,
+      "emos_mean_mae": 6.33,
-      "legacy_bucket_hit_rate": 0.6,
+      "legacy_bucket_hit_rate": 0.0,
-      "emos_bucket_hit_rate": 0.6
+      "emos_bucket_hit_rate": 0.0
     },
     "warsaw": {
       "samples": 3,
       "legacy_mean_crps": 1.618875,
-      "emos_mean_crps": 1.618875,
+      "emos_mean_crps": 1.553232,
       "legacy_mean_mae": 2.056667,
       "emos_mean_mae": 2.056667,
       "legacy_bucket_hit_rate": 0.333333,
       "emos_bucket_hit_rate": 0.333333
     },
     "wellington": {
-      "samples": 6,
+      "samples": 2,
-      "legacy_mean_crps": 0.266349,
+      "legacy_mean_crps": 0.364919,
-      "emos_mean_crps": 0.266349,
+      "emos_mean_crps": 0.475875,
-      "legacy_mean_mae": 0.2,
+      "legacy_mean_mae": 0.15,
-      "emos_mean_mae": 0.2,
+      "emos_mean_mae": 0.15,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     }

docs/TECH_DEBT.md

@@ -29,8 +29,7 @@ flowchart TD
     end
 
     subgraph S["状态与概率"]
-        S1["SQLite 第二阶段收口（离线脚本统一数据源）"]
+        S1["EMOS shadow -> primary 门禁稳定化"]
-        S2["EMOS shadow -> primary 门禁稳定化"]
     end
 
     A --> P
@@ -48,7 +47,7 @@ flowchart TD
 - 钱包绑定支持浏览器钱包 + WalletConnect。
 - 账户中心与 Pro 权限展示链路打通。
 - 钱包异动支持独立频道路由。
-- 运行态状态/缓存已完成 SQLite 主读切换。
+- 运行态状态/缓存与核心离线训练、评估、回填链路已完成 SQLite 主路径收口。
 - 轻量可观测性已上线（`/healthz`、`/api/system/status`、`/metrics`）。
 - EMOS/CRPS 校准链路已上线 shadow 模式。
 
@@ -56,7 +55,6 @@ flowchart TD
 
 | 项目 | 影响 | 建议动作 |
 | :-- | :-- | :-- |
-| SQLite 第二阶段收口 | 线上主链路已切到 `sqlite`，但离线脚本仍有文件默认输入 | 训练、回填、报表脚本统一改成 SQLite 优先，文件参数仅保留显式回退 |
 | EMOS 上线门禁 | 当前 `hold`，不能切 primary | 继续积累样本，重点压 `bucket_brier` |
 | 外部监控与告警 | 只有轻量指标，无外部抓取 | 接 Prometheus/Grafana 或最小巡检 |
 | 退款与售后链路 | 商业闭环不完整 | 增加退款状态机与工单系统 |
@@ -78,7 +76,6 @@ flowchart TD
 
 ## 6. 下阶段里程碑
 
-1. 完成 SQLite 第二阶段收口，统一离线脚本与训练输入到 SQLite 优先。
+1. 稳定 EMOS shadow，达到 rollout `observe/promote` 条件。
-2. 稳定 EMOS shadow，达到 rollout `observe/promote` 条件。
+2. 补外部监控抓取与告警阈值。
-3. 补外部监控抓取与告警阈值。
+3. 评估并推进支付合约 V2 升级。
-4. 评估并推进支付合约 V2 升级。

docs/TECH_DEBT_ZH.md

@@ -29,8 +29,7 @@ flowchart TD
     end
 
     subgraph S["状态与概率"]
-        S1["SQLite 第二阶段收口（离线脚本统一数据源）"]
+        S1["EMOS shadow -> primary 门禁稳定化"]
-        S2["EMOS shadow -> primary 门禁稳定化"]
     end
 
     A --> P
@@ -48,7 +47,7 @@ flowchart TD
 - 钱包绑定支持浏览器钱包 + WalletConnect。
 - 账户中心与 Pro 权限展示链路打通。
 - 钱包异动支持独立频道路由。
-- 运行态状态/缓存已完成 SQLite 主读切换。
+- 运行态状态/缓存与核心离线训练、评估、回填链路已完成 SQLite 主路径收口。
 - 轻量可观测性已上线（`/healthz`、`/api/system/status`、`/metrics`）。
 - EMOS/CRPS 校准链路已上线 shadow 模式。
 
@@ -56,7 +55,6 @@ flowchart TD
 
 | 项目 | 影响 | 建议动作 |
 | :-- | :-- | :-- |
-| SQLite 第二阶段收口 | 线上主链路已切到 `sqlite`，但离线脚本仍有文件默认输入 | 训练、回填、报表脚本统一改成 SQLite 优先，文件参数仅保留显式回退 |
 | EMOS 上线门禁 | 当前 `hold`，不能切 primary | 继续积累样本，重点压 `bucket_brier` |
 | 外部监控与告警 | 只有轻量指标，无外部抓取 | 接 Prometheus/Grafana 或最小巡检 |
 | 退款与售后链路 | 商业闭环不完整 | 增加退款状态机与工单系统 |
@@ -78,7 +76,6 @@ flowchart TD
 
 ## 6. 下阶段里程碑
 
-1. 完成 SQLite 第二阶段收口，统一离线脚本与训练输入到 SQLite 优先。
+1. 稳定 EMOS shadow，达到 rollout `observe/promote` 条件。
-2. 稳定 EMOS shadow，达到 rollout `observe/promote` 条件。
+2. 补外部监控抓取与告警阈值。
-3. 补外部监控抓取与告警阈值。
+3. 评估并推进支付合约 V2 升级。
-4. 评估并推进支付合约 V2 升级。

docs/deep-research-report.md

@@ -3,9 +3,9 @@
 ## 执行摘要
 
 PolyWeather（仓库：`yangyuan-zhen/PolyWeather`）定位为**面向温度类结算预测市场（如 Polymarket 的温度结算合约）**的“生产级气象情报系统”，核心在于把多源天气观测/预报转化为**结算导向的概率桶（μ + bucket distribution）**，并进一步映射到市场报价完成**错价扫描**；同时提供 Web 仪表盘与 Telegram Bot 两套交互入口，并包含 Polygon 链上 USDC/USDC.e 支付、自动补单与订阅/积分体系。项目 README 现明确仓库代码采用 `AGPL-3.0-only`，同时将品牌、商标、生产私有数据与运营阈值保留在代码许可证之外。
-从工程实现看，截至 `2026-03-21`，项目已经完成一轮明确的工程化收口：多源天气采集仍保持现有业务能力，同时已完成采集层与 Web API 大文件拆分、CI 质量门禁、配置分级（`.env.example` / `.env.secrets.example` / 中文部署文档）、EMOS/CRPS 校准链路、运行态状态与缓存向 SQLite 的渐进迁移，以及基础可观测性接口（`/healthz`、`/api/system/status`、`/metrics`）。
+从工程实现看，截至 `2026-04-02`，项目已经完成一轮明确的工程化收口：多源天气采集仍保持现有业务能力，同时已完成采集层与 Web API 大文件拆分、CI 质量门禁、配置分级（`.env.example` / `.env.secrets.example` / 中文部署文档）、EMOS/CRPS 校准链路、运行态状态与缓存迁移到 SQLite 主路径，以及基础可观测性接口（`/healthz`、`/api/system/status`、`/metrics`）。
 这意味着报告里最初最突出的“工程地基缺失”问题，已经有一部分被关闭：`src/data_collection/weather_sources.py` 与 `web/app.py` 不再是原来的超大单文件；GitHub Actions 已覆盖 Python、前端和 Docker build；配置与密钥治理已成体系；运行态状态不再只能依赖 JSON/JSONL 文件；EMOS 也不再只是概念，而是进入了可训练、可评估、可 shadow、可门禁判断的阶段。
-但项目仍处在“从可用走向稳态”的中段，而不是终局。当前真正的高优先级问题已收敛为三类：第一，**运行态 SQLite 已经完成主读切换，但离线训练/回填脚本仍存在 legacy JSON/JSONL 默认输入**，需要继续把数据科学与运维脚本统一到单一数据源；第二，**可观测性只完成了轻量级指标层**，还没有形成完整的外部监控、阈值告警与趋势面板；第三，**EMOS 仍未达到生产切换标准**，当前门禁结论明确为 `hold`，阻塞原因是 shadow bucket brier 明显退化。支付链路方面，链下审计与容灾已明显增强：事件重放、SQLite 审计事件、RPC 多节点容灾、合约静态检查、`/ops` 支付异常单、按邮箱恢复脚本都已补齐；当前剩余风险主要集中在**链上合约本身仍是最小实现**，尚未升级到 SafeERC20、Pausable、链上套餐绑定等更强防护版本。
+但项目仍处在“从可用走向稳态”的中段，而不是终局。当前真正的高优先级问题已收敛为两类：第一，**可观测性只完成了轻量级指标层**，还没有形成完整的外部监控、阈值告警与趋势面板；第二，**EMOS 仍未达到生产切换标准**，当前门禁结论明确为 `hold`，阻塞原因是 shadow bucket brier 明显退化。SQLite 迁移方面，运行态主读切换和核心离线训练/回填链路已经完成验收：在移除 `data/*.json` / `data/*.jsonl` 后，训练、评估、shadow report 与关键 backfill 脚本仍可仅依赖运行时数据库正常执行；当前保留的 legacy 文件路径主要用于迁移、导出、校验和显式回退输入。支付链路方面，链下审计与容灾已明显增强：事件重放、SQLite 审计事件、RPC 多节点容灾、合约静态检查、`/ops` 支付异常单、按邮箱恢复脚本都已补齐；当前剩余风险主要集中在**链上合约本身仍是最小实现**，尚未升级到 SafeERC20、Pausable、链上套餐绑定等更强防护版本。
 因此，当前阶段最正确的策略已经不是继续做“大范围基础重构”，而是围绕**迁移验收、可观测性补全、EMOS 上线门禁稳定化**这三条线持续收口。短中期内更高 ROI 的方向依然不是引入新的大模型，而是把现有“采集→后处理→市场映射→支付/订阅”的链路做成**状态一致、指标可见、发布可控、回退明确**的生产平台。
 ## 项目概览
 
@@ -58,7 +58,7 @@ flowchart TB
  WX[WeatherDataCollector]
  CITY[CITY_REGISTRY]
  HIST[(SQLite runtime state<br/>daily_records / cache / snapshots)]
- JSON[Legacy JSON files<br/>dual-mode fallback]
+JSON[Legacy JSON files<br/>migration/export/explicit fallback only]
  end
 
  subgraph ExternalAPIs
@@ -122,7 +122,7 @@ flowchart TB
 **推理流水线（在线）**：
 Web/Telegram 请求 → FastAPI 调用采集器抓取/复用缓存 → 分析引擎输出结构化结果（μ、概率桶、趋势、死盘/窗口判定、DEB 预测、市场扫描）→ 前端渲染或 bot 消息格式化。
 **检查点（checkpoints）**：传统 ML checkpoint 不适用；但项目现已形成两类“业务状态 checkpoint”：
-（a）SQLite 运行态存储（当前线上主路径）；（b）legacy JSON/JSONL 文件（主要保留给迁移回滚与部分离线脚本默认输入）。当前设计仍支持 `POLYWEATHER_STATE_STORAGE_MODE=file|dual|sqlite`，但对线上部署而言，推荐目标状态已经是 `sqlite`。
+（a）SQLite 运行态存储（当前线上与核心离线链路主路径）；（b）legacy JSON/JSONL 文件（主要保留给迁移回滚、导出比对与显式回退输入）。当前设计仍支持 `POLYWEATHER_STATE_STORAGE_MODE=file|dual|sqlite`，但对线上部署与离线训练/回填而言，推荐目标状态都已经是 `sqlite`。
 ### 测试、CI/CD 与运维验证
 
 **测试**：仓库存在 `tests/test_trend_engine.py`，覆盖 μ 计算、死盘判定、预报崩盘提示、趋势方向等核心逻辑（通过 patch 隔离外部依赖）。
@@ -142,7 +142,7 @@ Web/Telegram 请求 → FastAPI 调用采集器抓取/复用缓存 → 分析引
 **核心文件过大问题已明显缓解，但边界仍需继续稳定**：`WeatherDataCollector` 与 `web/app.py` 的超大文件问题已完成第一阶段拆分；当前风险已从“文件过大”转为“跨模块兼容与边界稳定性”，例如旧调用路径、兼容导出、跨层 helper 仍需持续清理。
 **可复现性已从“缺模板”进入“模板与生产对齐”的阶段**：`.env.example`、`.env.secrets.example`、中文配置文档、前端部署文档、运行时配置校验器都已存在；当前风险主要在于线上历史 `.env` 与新模板并存、旧变量命名残留、以及密钥轮换与分层是否真正落实。
 **CI 已建立，但组织级质量门禁未必完全收口**：CI 现已覆盖 Python、前端与 Docker build。当前问题不再是“缺 CI”，而是是否把这些 status check 绑定到 `main` 保护策略，以及是否逐步引入更严格的 pre-merge 审查。
-**运行态状态/缓存迁移已完成主切换，但离线脚本仍待收口**：`daily_records`、`telegram_alert_state`、`probability_training_snapshots`、`open_meteo` 缓存已经支持并可在生产中主读 SQLite，迁移/校验脚本也已验证可用；当前残留问题不再是线上是否能切，而是部分训练、回填、报表脚本仍默认读取 `data/*.json` / `data/*.jsonl`，容易形成“线上一套数据、离线一套输入”的维护成本。
+**运行态状态/缓存与核心离线链路的 SQLite 收口已完成**：`daily_records`、`telegram_alert_state`、`probability_training_snapshots`、`open_meteo` 缓存已经支持并在生产中主读 SQLite，迁移/校验脚本可用；进一步地，在临时移除 `data/*.json` / `data/*.jsonl` 后，训练集导出、概率拟合、评估报告、shadow report 和关键 backfill 脚本已验证仍可运行。当前 legacy 文件路径主要是显式回退入口，而不再是默认主输入。
 **第三方服务合规与稳定性风险**：
 项目强依赖外部 API（Open-Meteo、AviationWeather、NWS、HKO、CWA、Polymarket、Supabase）。其中 AviationWeather Data API 有明确速率限制；Polymarket 官方说明 Gamma/Data/CLOB 三套 API 分属不同域，CLOB 交易端点需鉴权且策略可能变化；Supabase 明确强调 `service_role`/secret keys 绝不可暴露。若缺乏集中治理（重试/退避/熔断/降级/配额监控/密钥轮换），稳定性与合规不可控。
 **可观测性已起步，但仍不构成完整监控体系**：项目现在已有 `/healthz`、`/api/system/status`、`/metrics`，并为 HTTP 与关键第三方源增加了轻量指标；但仍缺少 Prometheus/Grafana 级别的外部抓取、告警阈值、趋势面板和运行日报。这部分现在属于“已开始，不算完成”。
@@ -172,7 +172,6 @@ Web/Telegram 请求 → FastAPI 调用采集器抓取/复用缓存 → 分析引
 下表按截至 `2026-03-21` 的真实状态重排优先级。已完成项不再继续列为“待做”，只保留当前仍需推进的事项。
 | 优先级 | 改进项 | 预估工作量 | 主要收益 | 主要风险 | 可执行步骤（建议顺序） |
 | ------ | --------------------------------------------------------------------------------------------------------------------------------- | -------------------: | ------------------------------------------------------------------- | ------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| 高 | **完成 SQLite 迁移的“第二阶段收口”**：把离线脚本、训练与回填链路统一成 SQLite 优先 | 2–5 天 | 彻底形成单一数据源，避免线上 SQLite 与离线 JSON/JSONL 双轨漂移 | 训练脚本行为变化可能影响既有报表或回填输出 | 1) 盘点所有默认读取 `data/*.json` / `data/*.jsonl` 的脚本 → 2) 改为按 `state_storage_mode` 自动优先读 SQLite → 3) 保留显式 `--history-file/--snapshot-file` 作为回退输入 → 4) 用同一批数据对比训练与报表结果 |
 | 高 | **把轻量可观测性接入外部监控与告警**：围绕 `/metrics` 建立抓取、阈值与巡检 | 3–7 天 | 不再只靠日志定位问题；可以监控第三方源错误率、缓存命中与 HTTP 延迟 | 指标不分层会导致噪音高、告警无用 | 1) 抓取 `/metrics` → 2) 先围绕 HTTP、Open-Meteo、MGM、METAR 建立最小仪表板 → 3) 为 429/403/error/stale_cache 设阈值 → 4) 增加巡检脚本或告警通道 |
 | 高 | **稳定 EMOS shadow 并收紧上线门禁** | 1–2 周 | 让概率引擎升级具备明确发布条件，避免拍脑袋切换 | 当前 shadow bucket brier 退化明显，存在误上线风险 | 1) 持续积累 snapshot 样本 → 2) 定期重训与生成 `evaluation_report` / `shadow_report` / `rollout_report` → 3) 重点压 `bucket_brier` 退化 → 4) 只有门禁从 `hold` 进入 `observe/promote` 后才考虑上线 |
 | 中 | **市场层升级为 async + 类型安全**：引入 `aiopolymarket` 或在现有层加重试/backoff/连接池 | 4–7 天 | 行情层更稳，减少短时网络抖动；更易扩展更多市场/分页 | 依赖升级带来的行为差异 | 1) 把 requests.Session 替换为 aiohttp/httpx → 2) 在 Gamma/CLOB 调用侧实现指数退避 → 3) 引入 typed models，减少解析失败 |

scripts/backfill_daily_record_model_gaps.py

@@ -16,6 +16,7 @@ from src.analysis.probability_snapshot_archive import (  # noqa: E402
     load_snapshot_rows_for_day,
 )
 from src.database.runtime_state import STATE_STORAGE_FILE, get_state_storage_mode  # noqa: E402
+from scripts.fit_probability_calibration import _default_history_arg  # noqa: E402
 
 
 def _load_daily_records(path: Path) -> Dict[str, Dict[str, Dict[str, Any]]]:
@@ -53,8 +54,8 @@ def main() -> int:
     )
     parser.add_argument(
         "--history-file",
-        default=str(Path("data") / "daily_records.json"),
+        default=_default_history_arg(),
-        help="Path to daily_records.json",
+        help="Optional legacy daily_records.json path. In sqlite mode this defaults to the runtime database.",
     )
     parser.add_argument("--city", help="Optional city filter, e.g. ankara")
     parser.add_argument("--date", help="Optional YYYY-MM-DD filter")
@@ -70,8 +71,8 @@ def main() -> int:
     )
     args = parser.parse_args()
 
-    history_path = Path(args.history_file)
+    history_path = Path(args.history_file) if args.history_file else None
-    data = _load_daily_records(history_path)
+    data = _load_daily_records(history_path or Path())
     model_name = str(args.model or "").strip()
     city_filter = str(args.city or "").strip().lower() or None
     date_filter = str(args.date or "").strip() or None
@@ -133,8 +134,8 @@ def main() -> int:
     if changed:
         previous_mode = get_state_storage_mode()
         # Reuse existing save path semantics. In sqlite-only mode, save_history would skip file write.
-        save_history(str(history_path), data)
+        save_history(str(history_path or ""), data)
-        if previous_mode == STATE_STORAGE_FILE and not history_path.exists():
+        if previous_mode == STATE_STORAGE_FILE and (history_path is None or not history_path.exists()):
             raise FileNotFoundError(history_path)
 
     return 0

scripts/backfill_probability_shadow_history.py

@@ -13,6 +13,7 @@ from src.analysis.probability_calibration import (  # noqa: E402
     apply_probability_calibration,
     build_probability_features,
 )
+from scripts.fit_probability_calibration import _default_history_arg  # noqa: E402
 
 
 def _sample_to_features(sample):
@@ -44,7 +45,7 @@ def main():
     parser = argparse.ArgumentParser(description="Backfill shadow probability snapshots into daily records.")
     parser.add_argument(
         "--history-file",
-        default=os.path.join(PROJECT_ROOT, "data", "daily_records.json"),
+        default=_default_history_arg(),
     )
     parser.add_argument(
         "--training-samples",

scripts/backfill_recent_daily_actuals_from_metar.py

@@ -9,6 +9,7 @@ if PROJECT_ROOT not in sys.path:
 
 from src.analysis.deb_algorithm import load_history, reconcile_recent_actual_highs, save_history  # noqa: E402
 from src.data_collection.city_registry import CITY_REGISTRY  # noqa: E402
+from scripts.fit_probability_calibration import _default_history_arg  # noqa: E402
 
 
 def _target_dates(city_info: dict, lookback_days: int) -> list[str]:
@@ -50,7 +51,7 @@ def main() -> None:
     )
     args = parser.parse_args()
 
-    history_file = os.path.join(PROJECT_ROOT, "data", "daily_records.json")
+    history_file = _default_history_arg() or ""
     data = load_history(history_file)
 
     selected = {str(item).strip().lower() for item in args.cities if str(item).strip()}

scripts/build_probability_shadow_report.py

@@ -11,6 +11,7 @@ if PROJECT_ROOT not in sys.path:
 
 from src.analysis.deb_algorithm import load_history  # noqa: E402
 from src.analysis.settlement_rounding import apply_city_settlement  # noqa: E402
+from scripts.fit_probability_calibration import _default_history_arg  # noqa: E402
 
 
 def _sf(value):
@@ -112,7 +113,7 @@ def main():
     parser = argparse.ArgumentParser(description="Build live shadow probability report from daily records.")
     parser.add_argument(
         "--history-file",
-        default=os.path.join(PROJECT_ROOT, "data", "daily_records.json"),
+        default=_default_history_arg(),
     )
     parser.add_argument(
         "--output",

scripts/evaluate_probability_calibration.py

@@ -16,6 +16,7 @@ from src.analysis.probability_calibration import (  # noqa: E402
 )
 from src.analysis.settlement_rounding import apply_city_settlement  # noqa: E402
 from scripts.fit_probability_calibration import (  # noqa: E402
+    _default_history_arg,
     _extract_samples,
     _load_json_if_exists,
 )
@@ -61,7 +62,7 @@ def main():
     parser = argparse.ArgumentParser(description="Evaluate legacy vs EMOS probability calibration.")
     parser.add_argument(
         "--history-file",
-        default=os.path.join(PROJECT_ROOT, "data", "daily_records.json"),
+        default=_default_history_arg(),
     )
     parser.add_argument(
         "--settlement-history",

scripts/export_probability_training_dataset.py

@@ -8,6 +8,8 @@ if PROJECT_ROOT not in sys.path:
     sys.path.insert(0, PROJECT_ROOT)
 
 from scripts.fit_probability_calibration import (  # noqa: E402
+    _default_history_arg,
+    _default_snapshot_arg,
     _extract_samples,
     _load_history_with_fallback,
     _load_json_if_exists,
@@ -19,7 +21,7 @@ def main():
     parser = argparse.ArgumentParser(description="Export normalized probability calibration training samples.")
     parser.add_argument(
         "--history-file",
-        default=os.path.join(PROJECT_ROOT, "data", "daily_records.json"),
+        default=_default_history_arg(),
     )
     parser.add_argument(
         "--settlement-history",
@@ -41,7 +43,7 @@ def main():
     )
     parser.add_argument(
         "--snapshot-file",
-        default=os.path.join(PROJECT_ROOT, "data", "probability_training_snapshots.jsonl"),
+        default=_default_snapshot_arg(),
     )
     args = parser.parse_args()
 

scripts/fit_probability_calibration.py

@@ -16,6 +16,7 @@ from src.analysis.probability_calibration import (  # noqa: E402
 from src.analysis.deb_algorithm import load_history  # noqa: E402
 from src.database.runtime_state import (  # noqa: E402
     ProbabilitySnapshotRepository,
+    STATE_STORAGE_FILE,
     STATE_STORAGE_SQLITE,
     get_state_storage_mode,
 )
@@ -38,6 +39,22 @@ def _load_json_if_exists(path):
     return data if isinstance(data, dict) else {}
 
 
+def _legacy_history_path():
+    return os.path.join(PROJECT_ROOT, "data", "daily_records.json")
+
+
+def _legacy_snapshot_path():
+    return os.path.join(PROJECT_ROOT, "data", "probability_training_snapshots.jsonl")
+
+
+def _default_history_arg():
+    return _legacy_history_path() if get_state_storage_mode() == STATE_STORAGE_FILE else None
+
+
+def _default_snapshot_arg():
+    return _legacy_snapshot_path() if get_state_storage_mode() == STATE_STORAGE_FILE else None
+
+
 def _load_history_with_fallback(path):
     data = load_history(path)
     if data:
@@ -253,8 +270,8 @@ def main():
     parser = argparse.ArgumentParser(description="Fit PolyWeather probability calibration parameters.")
     parser.add_argument(
         "--history-file",
-        default=os.path.join(PROJECT_ROOT, "data", "daily_records.json"),
+        default=_default_history_arg(),
-        help="Path to the historical daily_records.json file.",
+        help="Optional legacy daily_records.json path. In sqlite mode this defaults to the runtime database.",
     )
     parser.add_argument(
         "--output",
@@ -273,8 +290,8 @@ def main():
     )
     parser.add_argument(
         "--snapshot-file",
-        default=os.path.join(PROJECT_ROOT, "data", "probability_training_snapshots.jsonl"),
+        default=_default_snapshot_arg(),
-        help="Optional JSONL file with archived probability snapshots.",
+        help="Optional legacy JSONL snapshot archive path. In sqlite mode this defaults to the runtime database.",
     )
     parser.add_argument(
         "--version",

src/bot/analysis/deb_analysis_service.py

@@ -4,13 +4,18 @@ import os
 from datetime import datetime as _dt
 from datetime import timedelta as _td
 
+from src.database.runtime_state import STATE_STORAGE_FILE, get_state_storage_mode
+
 
 class DebAnalysisService:
     """DEB analytics adapter with lazy imports to trim cold startup."""
 
     def __init__(self, project_root: str):
         self.project_root = project_root
-        self.history_file = os.path.join(project_root, "data", "daily_records.json")
+        if get_state_storage_mode() == STATE_STORAGE_FILE:
+            self.history_file = os.path.join(project_root, "data", "daily_records.json")
+        else:
+            self.history_file = ""
 
     @staticmethod
     def _load_aliases() -> dict[str, str]:

src/models/lgbm_features.py

@@ -8,6 +8,13 @@ from typing import Any, Dict, List, Optional, Tuple
 
 from src.analysis.deb_algorithm import load_history
 from src.data_collection.city_registry import ALIASES
+from src.database.runtime_state import (
+    DailyRecordRepository,
+    ProbabilitySnapshotRepository,
+    STATE_STORAGE_FILE,
+    STATE_STORAGE_SQLITE,
+    get_state_storage_mode,
+)
 
 
 BASE_MODEL_COLUMNS: List[Tuple[str, str]] = [
@@ -82,12 +89,16 @@ def _parse_timestamp(value: Any) -> Optional[datetime]:
 
 def _history_file_path() -> str:
     root = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-    return os.path.join(root, "data", "daily_records.json")
+    if get_state_storage_mode() == STATE_STORAGE_FILE:
+        return os.path.join(root, "data", "daily_records.json")
+    return ""
 
 
 def _snapshot_archive_path() -> str:
     root = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-    return os.path.join(root, "data", "probability_training_snapshots.jsonl")
+    if get_state_storage_mode() == STATE_STORAGE_FILE:
+        return os.path.join(root, "data", "probability_training_snapshots.jsonl")
+    return ""
 
 
 def _normalized_city_key(city_name: str) -> str:
@@ -125,8 +136,26 @@ def _compute_model_summary(features: Dict[str, Optional[float]]) -> Tuple[Option
 
 
 def load_snapshot_index(archive_path: Optional[str] = None) -> Dict[Tuple[str, str], Dict[str, Any]]:
+    mode = get_state_storage_mode()
+    if mode == STATE_STORAGE_SQLITE:
+        latest_rows: Dict[Tuple[str, str], Dict[str, Any]] = {}
+        for row in ProbabilitySnapshotRepository().load_all_rows():
+            if not isinstance(row, dict):
+                continue
+            city = _normalized_city_key(str(row.get("city") or ""))
+            date_str = str(row.get("date") or "").strip()
+            if not city or not date_str:
+                continue
+            key = (city, date_str)
+            current_best = latest_rows.get(key)
+            if current_best is None or str(row.get("timestamp") or "") >= str(
+                current_best.get("timestamp") or ""
+            ):
+                latest_rows[key] = row
+        return latest_rows
+
     path = archive_path or _snapshot_archive_path()
-    if not os.path.exists(path):
+    if not path or not os.path.exists(path):
         return {}
 
     latest_rows: Dict[Tuple[str, str], Dict[str, Any]] = {}
@@ -258,7 +287,12 @@ def build_training_samples(
     history_data: Optional[Dict[str, Any]] = None,
     snapshot_index: Optional[Dict[Tuple[str, str], Dict[str, Any]]] = None,
 ) -> List[Dict[str, Any]]:
-    data = history_data if isinstance(history_data, dict) else load_history(_history_file_path())
+    if isinstance(history_data, dict):
+        data = history_data
+    elif get_state_storage_mode() == STATE_STORAGE_SQLITE:
+        data = DailyRecordRepository().load_all()
+    else:
+        data = load_history(_history_file_path())
     snapshots = snapshot_index if isinstance(snapshot_index, dict) else load_snapshot_index()
     samples: List[Dict[str, Any]] = []