fix(runpod): align sampling params with model card (top_k=40, top_p=0.95, rep=1.0)

fix(runpod): add explicit vLLM sampling params (top_k, top_p, repetition_penalty)

fix(runpod): handle tokens array and usage input/output format in response parser

feat(app): ZeroGPU 쿼터 소진 시 RunPod Serverless fallback 추가

fix(agent): step.query 누락 시 원본 쿼리로 fallback (모델이 Query 필드 생략 시 검색 스킵 방지)

fix(reasoning): EOD 래퍼 헤더 제거 및 결론→투자 의견 정규화

fix(reasoning): EOD 형식에서 _remove_boilerplate 제거로 섹션 손실 방지

fix(prompts): synthesis 프롬프트를 모델 학습 포맷(h3)에 맞게 단순화

fix(zerogpu): PicklingError 수정 - GPU 함수 내부에서 agent 생성

fix(zerogpu): 단일 GPU 컨텍스트로 전체 research 실행

fix(app): ZeroGPU 호환성 - 백그라운드 스레드 제거 및 메인 스레드 동기 실행

fix(agent): sentence-transformers 추가 및 ActionType 직렬화 수정

fix(synthesis): remove Key Findings suffix from synthesis prompt and fix ActionType serialization

feat(vectors): revert to KR-SBERT 768-dim with CPU lazy load

feat(vectors): replace KR-SBERT with MiniLM for HF Serverless compatibility

fix(vector_filter): HF_TOKEN 명시적 전달 (API 401 수정)

fix(vector_filter): sentence-transformers→HF InferenceClient로 교체 (ZeroGPU CUDA 충돌 수정)

fix(search): source_type.value 오류 수정 - use_enum_values=True로 문자열 저장됨

feat(search): KR-SBERT 기업벡터 뉴스 필터 추가 (CPU 모드, HF Spaces용)

fix(reasoning): EOD 형식 synthesis 파싱 수정 - conclusion 빈 문자열 버그

feat: fork pykrx naver OHLCV into naver_finance_client (HF Spaces safe)

feat: align synthesis output with STONKS EOD deep report format

docs: note sampling config alignment with llama-cpp-python defaults

fix: handle Conclusions plural, add key takeaways to junk, improve content dedup

fix: add h2 truncation, content dedup, and training leak cleanup in conclusion

fix: fix boilerplate body length check and add table debris removal

fix: improve conclusion truncation and boilerplate removal for 7B model

fix: expand conclusion pattern to match Conclusions and Key Takeaways variant

fix: improve conclusion truncation with separator and length limits

fix: truncate junk sections generated after conclusion

fix: expand boilerplate detection patterns and add inline junk removal

fix: remove 7B model boilerplate sections from synthesis output

fix: add stop sequences for chat pattern and remove hallucinated metadata

fix: section dedup bypass for Korean headers and reduce synthesis token limit

fix: rollback ZeroGPU to minimal generation params matching MLX approach

fix: prevent prompt template leakage and harden inference API path

fix: improve ZeroGPU output quality — generation params, repetition suppression, content filtering

fix: prevent stock data duplication in synthesis prompt

feat: add Naver Finance stock data integration (price/valuation/investor)

fix: synthesis confidence parser handles bold format

docs: add Production Enhancements section and Production Note column

docs: add Known Limitations section to README and Gradio demo

fix: query diversification and synthesis confidence fallback

fix(zerogpu): add paragraph-level repetition detection to postprocess

fix(reasoning): handle missing Thought prefix in model output

fix(zerogpu): align gen_params with model generation_config

fix(zerogpu): remove no_repeat_ngram_size that blocks RT format

fix: replace explicit EOS with repetition prevention for ZeroGPU

fix: add EOS token handling and output postprocessing for ZeroGPU

docs: add HF Spaces YAML metadata to README