C.E.L_Slide_test2/tests/matching/pipeline_17b_extend_missing_sections.py

"""Pipeline Step 17b — V4 full-32 평가를 누락 섹션에 확장 (GitHub issue #17).

배경: v4_full32_result.yaml (2026-04-29) 은 mdx 01~04 의 10개 섹션만 평가 —
`01-intro` / `05-1` / `05-2` 는 V4 source 자체가 없어 런타임에서 무조건
generic fallback (design_readiness=not_ready) 이 됨 (emergency.md C3).

원칙: "결과물을 고치지 말고 프로세스를 고친다" — yaml 손편집이 아니라
pipeline_17 과 동일한 평가 코드(detect_mdx_analysis + compute_template_fit
+ route)로 누락 섹션만 평가해 병합한다. 기존 섹션 엔트리는 무접촉.

실행: python tests/matching/pipeline_17b_extend_missing_sections.py
출력: v4_full32_result.yaml (기존 10 + 신규 3 = 13 sections)
"""
import datetime
import sys
from pathlib import Path

import yaml

HERE = Path(__file__).parent
sys.path.insert(0, str(HERE))

from phase_common import load_32_frames, load_frame_index
from detect_mdx import detect_mdx_analysis
from template_fit import (
    load_templates_v1, collect_anchor_vocab, compute_template_fit, route,
)
from embeddings import embed_texts, cosine
from pipeline_01_extract_nodes import MDX_DIR

OUT_PATH = HERE / 'v4_full32_result.yaml'

# 신규 평가 대상 — 런타임 section_id 와 동일한 키 (lookup_v4 exact match).
# 01-intro 는 heading 없는 pre-intro 블록이라 시작 라인을 본문 첫 그룹으로 지정.
NEW_SECTIONS = {
    '01-intro': {'file': '01.mdx', 'start': '* **용어의 혼용**', 'end_prefix': '## 1.',
                 'title': '건설산업 DX의 올바른 이해 — 도입(용어의 혼용)'},
    '05-1':     {'file': '05.mdx', 'start': '## 1. 설계의 자동화', 'end_prefix': '## 2.',
                 'title': None},
    '05-2':     {'file': '05.mdx', 'start': '## 2. S/W 중심 설계 방식', 'end_prefix': None,
                 'title': None},
}


def extract_raw(cfg):
    lines = (MDX_DIR / cfg['file']).read_text(encoding='utf-8').split('\n')
    start = None
    for i, ln in enumerate(lines):
        if ln.strip() == cfg['start'].strip():
            start = i
            break
    if start is None:
        raise RuntimeError(f"start line not found: {cfg['start']!r} in {cfg['file']}")
    end = len(lines)
    if cfg.get('end_prefix'):
        for i in range(start + 1, len(lines)):
            if lines[i].strip().startswith(cfg['end_prefix']):
                end = i
                break
    section = lines[start:end]
    title = cfg.get('title') or section[0].lstrip('#').strip()
    return title, '\n'.join(section)


def main():
    existing = yaml.safe_load(OUT_PATH.read_text(encoding='utf-8'))
    already = set(existing['mdx_sections'].keys())
    todo = {sid: cfg for sid, cfg in NEW_SECTIONS.items() if sid not in already}
    if not todo:
        print('신규 평가 대상 없음 — 모두 존재.')
        return

    templates = load_templates_v1()
    anchor_vocab = collect_anchor_vocab(templates)
    frames = load_32_frames()
    idx_data, frame_to_short = load_frame_index()
    fids = list(frames.keys())
    frame_num_map = {fid: int(frame_to_short[fid]) for fid in fids}
    frame_contents = [frames[fid].get('content', '') for fid in fids]

    print(f'[17b] 32 frame content 임베딩 중... (신규 섹션: {sorted(todo)})')
    frame_vecs = embed_texts(frame_contents)

    for sid, cfg in sorted(todo.items()):
        title, raw_text = extract_raw(cfg)
        mdx_analysis = detect_mdx_analysis(raw_text, title, anchor_vocab=anchor_vocab)
        mdx_vec = embed_texts([mdx_analysis['summary']])[0]

        judgments = []
        for i, fid in enumerate(fids):
            if fid not in templates:
                continue
            template = templates[fid]
            content_emb = max(0.0, min(1.0, float(cosine(mdx_vec, frame_vecs[i]))))
            fit = compute_template_fit(mdx_analysis, template, content_emb)
            label = route(fit['confidence'], fit['axes'], fit['adaptation'], fit['not_suits'])
            judgments.append({
                'frame_id': fid,
                'frame_number': frame_num_map[fid],
                'template_id': template.get('template_id'),
                'confidence': round(float(fit['confidence']), 4),
                'base': round(float(fit['base']), 4),
                'penalty': round(float(fit['total_penalty']), 4),
                'label': label,
                'content_embedding': round(content_emb, 4),
                'axes': {
                    'anchor': round(float(fit['axes']['anchor']['score']), 4),
                    'cardinality': round(float(fit['axes']['cardinality']), 4),
                    'relation': round(float(fit['axes']['relation']), 4),
                    'slot': round(float(fit['axes']['slot']), 4),
                    'content': round(float(fit['axes']['content']), 4),
                },
            })
        judgments.sort(key=lambda x: -x['confidence'])
        for new_rank, item in enumerate(judgments, start=1):
            item['v4_full_rank'] = new_rank

        existing['mdx_sections'][sid] = {
            'mdx_title': title,
            'answer_frame_number': None,   # blind — ANSWER_MAP 변경 금지 (Holdout 원칙 1)
            'is_holdout': True,
            'judgments_full32': judgments,
            'usable_count': sum(1 for j in judgments if j['label'] != 'reject'),
            'reject_count': sum(1 for j in judgments if j['label'] == 'reject'),
        }
        top = judgments[0]
        print(f"  {sid}: rank1 = F{top['frame_number']} {top['template_id']} "
              f"({top['label']}, {top['confidence']}) usable={existing['mdx_sections'][sid]['usable_count']}")

    meta = existing.setdefault('meta', {})
    ext = meta.setdefault('extensions', [])
    ext.append({
        'step': '17b_extend_missing_sections',
        'added_sections': sorted(todo),
        'reason': 'GitHub issue #17 — generic fallback 탈출 (C3: V4 source 누락 해소)',
        'generated_at': datetime.datetime.now().isoformat(timespec='seconds'),
    })
    holdout = meta.setdefault('holdout_sections', [])
    for sid in sorted(todo):
        if sid not in holdout:
            holdout.append(sid)

    OUT_PATH.write_text(
        yaml.safe_dump(existing, allow_unicode=True, sort_keys=False, width=1000),
        encoding='utf-8',
    )
    print(f'병합 완료: {OUT_PATH} (총 {len(existing["mdx_sections"])} sections)')


if __name__ == '__main__':
    main()