[2026-05-27] 포맷 후 복구 설치 스크립트 추가

CLAUDE.md

@@ -4,6 +4,43 @@
 > 인프라: 로컬 Windows → Synology NAS Docker 이전 예정  
 > 현재 모드: 모의투자 (KIS_MOCK=true, DRY_RUN=true)
 
+## Current safety policy
+
+- AI is for market review and proposals only; code decides entries and exits.
+- Evening Claude must not edit `app/config.py` or execution/strategy/risk code directly.
+- Parameter changes must be written to `reports/proposals/YYYY-MM-DD_strategy_proposal.md` with evidence and manual approval required.
+- `FORCE_EXIT = "14:50"` remains immutable.
+
+## Current implementation status - 2026-05-27
+
+- Mode: paper trading / dry-run focused. Real-cash trading is not approved yet.
+- Data layer: `entry_snapshots` and `post_entry_snapshots` are active for training data.
+- Post-entry sampling: 60s, 180s, 300s, and 600s after successful entry.
+- Training data export: `scripts/export_training_dataset.py`.
+- External data collection:
+  - Daily market features: `scripts/collect_daily_features.py`.
+  - KIS minute bars: `scripts/collect_minute_data.py`.
+  - External dataset builder: `scripts/build_external_training_dataset.py`.
+- ML engine:
+  - Training: `scripts/train_ai_model.py`.
+  - Model output: `models/scalping_model.joblib`.
+  - Metrics output: `models/scalping_model.metrics.json`.
+  - Runtime loader: `app/ml/predictor.py`.
+- AI runtime mode: observation only. If a model exists, entry-time scores are logged and saved to `entry_snapshots`; they do not block or resize trades.
+- Training schedule: `StockBot_Training` runs at 16:00 on trading days via `scripts/run_training_pipeline.ps1`.
+- Dependency install:
+  - `requirements.txt` includes `app/requirements.txt`.
+  - `scripts/install_dependencies.ps1` installs from `vendor/wheels` when available.
+  - `scripts/download_dependencies.ps1` builds the local wheelhouse.
+
+## Current operational risks - 2026-05-27
+
+- KIS minute-bar endpoint must be verified with live response logs.
+- Early ML models may be meaningless until enough labeled rows exist.
+- External minute data is pretraining data, not actual bot-trade data.
+- Real-cash mode still needs stronger fill, partial-fill, unfilled-order, cancel/replace, and recovery logic.
+- Existing logs and older docs contain encoding damage; new operational notes should stay ASCII unless the file encoding is intentionally cleaned.
+
 ---
 
 ## 프로젝트 개요