[DANNA-001] add pykrx vendor and Parquet disk cache

[bbuduck]

Summary

• @simple_parquet_cache 데코레이터 추가 — sha256 키 기반 Parquet 캐시, TTL,
  fail-open, atomic write
• cache_admin.py CLI 추가 — stats / clear 서브커맨드
• pykrx_vendor.py 추가 — KRX 직접 조회: OHLCV, 시점 universe, 투자자별
  매매동향, PER/PBR 가치지표
• interface.py 등록 — kr_market_data 카테고리, route_to_vendor 통합
• conftest.py + .env.example — KRX_ID/KRX_PW 로컬 인증 가이드 포함

New tools via route_to_vendor

tool	vendor	설명
get_stock_data	pykrx	KRX OHLCV (인증 불필요)
get_kr_universe	pykrx	시점 KOSPI/KOSDAQ 종목 리스트
get_kr_investor_trading	pykrx	외국인/기관/개인 매매동향
get_kr_value_factors	pykrx	PER, PBR, EPS, BPS, DIV, DPS

Test plan

• pytest tests/test_dataflows_cache.py — 캐시 unit 테스트 6개 (네트워크
  불필요)
• RUN_NETWORK_TESTS=1 pytest tests/test_pykrx_vendor.py — 9개 통과 (unit
  3 + smoke 4 + integration 2)

Setup

.env.example 참고하여 .env 생성:

• KRX_ID / KRX_PW — data.krx.co.kr 무료 가입
  (OHLCV는 미필요)
• OPENDART_API_KEY — opendart.fss.or.kr 무료
  발급

[gemini-code-assist]

Code Review

This pull request significantly expands the TradingAgents framework by introducing a comprehensive backtesting engine, a performance tracking system, and an HTML dashboard for results visualization. Key additions include support for the Korean market via OpenDART and pykrx, real-time trade execution through the KIS broker with multi-layer safety guards, and an investment persona system. Architectural enhancements feature a disk-persisted Parquet cache, improved yfinance rate-limit handling, and a two-layer technical analysis approach. Feedback from the review highlights several critical technical issues: a concurrency bottleneck in the rate limiter caused by holding a lock during sleep, a memory leak in the cache path-locking mechanism, inefficient object instantiation in the OpenDART provider, redundant data processing in the technical calculator, and potential race conditions in file caching.

[gemini-code-assist]

The RateLimiter.acquire method holds the instance lock while sleeping. In a multi-threaded environment, this serializes all requests and prevents other threads from even checking the rate limit status, effectively negating the benefit of a sliding window for concurrent callers. The sleep should occur outside the lock.

[gemini-code-assist]

The _path_locks dictionary grows indefinitely as new unique file paths are accessed, leading to a memory leak. Since each entry is a threading.Lock object associated with a specific cache file, this can become significant over long-running sessions with many tickers or dates. Consider using a weakref.WeakValueDictionary to allow locks to be garbage collected when no longer in use.

[gemini-code-assist]

The _get_dart function creates a new OpenDartReader instance on every call. This is inefficient as it involves redundant initialization and potentially multiple authentication checks. It should be implemented as a singleton or cached at the module level.

[gemini-code-assist]

This loop calls _get_stock_stats_bulk for each indicator individually. If the underlying implementation fetches or processes the entire stock history for each call, this results in significant redundant work (e.g., multiple CSV reads or API calls). It is recommended to fetch the full technical dataset once and then extract the required indicators from the resulting DataFrame.

[gemini-code-assist]

Using a fixed temporary filename (.parquet.tmp) can lead to race conditions if multiple processes or threads attempt to write to the same cache entry simultaneously. While the in-process lock mitigates this for threads, it does not protect against multiple processes. Including the PID or a unique identifier in the temporary filename would be safer.