Wire C30 peer returns into production + add C29 MetaLabel log

Improve dashboard latency caching

Integrate Old Wang trading strategy overlay

Backup current stock predictor strategies

Promote hotspot recommendation runtime config

Merge remote-tracking branch 'hf/main' into codex/self-refine-validation-gate

feat: integrate local architecture with HF Space

fix(news): add 12h disk cache to prevent Gemini 429 quota exhaustion

fix(query_history): record queries from /batch and /recommend endpoints

Fix HF queue prediction placeholders

fix: HF Space always returns Mac ML result on first call

fix: unpack 4 values from fetch_cross_asset_tw in backtest_service + line_webhook

fix: reduce STALE_REFRESH_HOURS 4→2 to keep HF queue signals fresh

perf: 4h cache TTL on local Mac, 30min on HF Space

perf: cache queue.json in-memory (90s TTL) + run yfinance in thread

fix: correct HF dataset repo namespace to DennisChan0909

perf: return quick estimate before slow feature fetching

feat: quick-estimate + background ML + auto-poll

feat: HF Dataset queue — Mac worker pushes fresh predictions to HF Space

feat(C2): add PTT sentiment scraper and backtest; FAILED — PTT rows=0, features not added to model

feat: Phase 4 — securities lending, block trades, large holder features

feat: add TAIFEX put/call ratio and large trader features

feat: add chip_score, MA120/240, margin flow, and TA distillation

fix: remove responseMimeType when using google_search grounding (Gemini 400)

feat: add institutional signals and HF-friendly deploy

feat: ML ensemble upgrade — LightGBM, cross-asset features, walk-forward embargo

feat: stock chart image for LINE Bot + new /chart-image endpoint

feat(ml): Optuna walk-forward tuning + /tune endpoint

fix(metal): Phase 3 — import fix + predictor threshold kwargs + smoke test

feat(metal): Phase 1 backend — gold/silver analysis API

feat: typed NewsServiceError for precise news endpoint diagnostics

fix: upgrade Gemini news model to 2.5-flash (2.0-flash free tier is 0)

feat: show degraded mode warning in LINE prediction messages

fix: graceful degradation for stocks with < 60 rows (fallback to 30)

perf: incremental data fetch — only download delta after first load

perf: add history data cache + reduce workers to 1

fix: replace unbounded caches with TTLCache to prevent memory leaks

feat: comprehensive optimization — 30 improvements across ML, perf, frontend, security, architecture

perf: eliminate redundant _build_features in train/predict pipeline

perf: indicator cache, quote TTL, twstock cache, yfinance timeout

feat: walk-forward backtest, US currency, mobile RWD, Gemini news summary

refactor: improve ML features, LINE Flex Messages, watchlist panel, health endpoint

fix+refactor: LINE webhook key bug, fetcher DRY, feature importance, async predictor service