feat(presegment): G2 PR-3 — LLM 경계 폴백 (flag-gated, 기본 OFF, scaffold-first)

ToC 없는/게이트 미달 대형 PDF(>=60p)에 한해 off-card Qwen(맥북, call_deep_or_defer,
StageDeferred-safe) 경계 제안 → 동일 검증게이트(_is_clear_bundle) 통과 시에만 deterministic 과
공유하는 _create_children 로 분할. is_bundle=false/파싱·검증 실패=단일문서(오늘과 동일)+로깅.
- env PRESEGMENT_LLM_FALLBACK 기본 false → 배포 동작 무변(LLM 미호출, 검증=unit test)
- 자식생성 _create_children 공유 헬퍼로 리팩터(deterministic+LLM 단일 경로, 동작 동일)
- SegmentationOutput Pydantic + parse_json_response(house 패턴) + per-page heading 샘플(본문 미전송)
- prompt app/prompts/presegment_boundaries.txt + tests/test_presegment_llm.py(14, fitz/DB/LLM mock)
no direct HTTP·no silent fallback. 활성=flag ON + 실 router fixture 검증 후.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app/prompts/presegment_boundaries.txt

@@ -0,0 +1,41 @@
+You are a document-boundary detector. Output ONLY JSON {is_bundle, segments:[{start_page,end_page,title}]}.
+
+You are given a single PDF that may be a "bundle" — several independent logical documents
+concatenated into one file (for example: multiple laws, multiple reports, or multiple papers
+scanned together). Your job is to decide whether it is a bundle and, if so, where each logical
+document starts and ends.
+
+You receive only a compact sample per page: the page number and the first line / heading of that
+page (text may be truncated). Use these heading/first-line signals to detect where a new logical
+document begins (a new title page, a new cover, a clearly new document title, a restart of
+numbering, etc.). You do NOT receive the full text.
+
+Output rules:
+- Respond with STRICT JSON only. No prose, no markdown, no code fence.
+- Schema:
+  {
+    "is_bundle": true | false,
+    "segments": [
+      {"start_page": <int>, "end_page": <int>, "title": "<string or null>"}
+    ]
+  }
+- Page numbers are 1-based and INCLUSIVE. start_page=1 is the first page; end_page equals the last
+  page of that segment.
+- Segments MUST fully cover every page with NO gaps and NO overlaps:
+  - the first segment MUST start at page 1,
+  - each next segment MUST start exactly one page after the previous segment's end_page,
+  - the last segment MUST end at the final page (page_count).
+- Order segments by start_page ascending.
+- title = a short title for that logical document if you can infer one from its first page,
+  otherwise null.
+
+If the file is NOT a bundle (it is a single logical document), respond:
+  {"is_bundle": false, "segments": []}
+
+Be conservative: only report is_bundle=true when the heading signals clearly indicate separate
+logical documents. When unsure, return is_bundle=false.
+
+page_count: {page_count}
+
+Per-page samples (one per line, "p{n}: {first line}"):
+{page_samples}

app/workers/presegment_worker.py

@@ -4,11 +4,19 @@
   - 非PDF(file_format != pdf · suffix != .pdf) = 즉시 fast-exit → enqueue_next_stage 가 extract 로 흘림.
   - PDF = PyMuPDF ToC(level-1) deterministic 분석. '명확한 번들' 만 자식 분할, 나머지는 단일문서로 extract.
 
-이 PR 은 **deterministic 만** (LLM fallback = 후속 PR). 판정이 애매하면 보수적으로 분할하지 않고
-단일문서로 둔다(bias to NOT splitting). 분할 = '확실한 번들' 만:
+deterministic 경로(PR-G2-2): 판정이 애매하면 보수적으로 분할하지 않고 단일문서로 둔다
+(bias to NOT splitting). 분할 = '확실한 번들' 만:
   - page_count >= MIN_BUNDLE_PAGES AND level-1 ToC 항목 >= 2 AND 모든 자식 >= MIN_CHILD_PAGES
     AND 단조 증가·비중첩 AND [1, page_count] 전 범위 커버 AND 2 <= N <= MAX_CHILDREN.
 
+LLM 경계 폴백(PR-G2-3, env PRESEGMENT_LLM_FALLBACK, 기본 OFF — scaffold-first): deterministic
+이 '명확한 번들' 을 못 만든 대형 PDF(ToC 없음/level-1 없음/게이트 미달)에 한해, OFF 면 오늘과
+동일(단일문서)이고 ON 이면 off-card Qwen(맥북, 라우터 :8890, model=qwen-macbook)에게 경계를
+제안받는다. compact per-page heading 샘플만 전송(본문 미전송). LLM 출력은 **동일 검증 게이트
+(_is_clear_bundle)** 통과 시에만 deterministic 과 같은 _create_children 경로로 분할 —
+is_bundle=false / 파싱·검증 실패 = 단일문서(오늘과 동일) + presegment_llm_rejected 로깅.
+맥북 불가(503/연결/절단)는 StageDeferred 로 큐 재시도(백오프, no silent fallback).
+
 분할 시 ★후보 A(물리분할 없음, uq_documents_file_path 해소): 자식 file_path = unique 합성값
 `{부모경로}#p{start}-{end}` (UNIQUE 제약 통과), 실파일은 `bundle_source_path()` 로 부모 경로 복원.
 자식은 bundle_page_start/end(1-based inclusive) 로 부모 파일의 자기 page 범위만 가리킨다.
@@ -32,9 +40,11 @@ import re
 import unicodedata
 from pathlib import Path
 
+from pydantic import BaseModel, ValidationError
 from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from ai.client import AIClient, call_deep_or_defer, parse_json_response
 from core.config import settings
 from core.utils import setup_logger
 from models.document import Document
@@ -54,6 +64,40 @@ MAX_CHILDREN = int(os.getenv("PRESEGMENT_MAX_CHILDREN", "50"))
 # marker_worker._to_marker_path 와 동일 — NAS 상대경로 → 컨테이너 절대경로 prefix.
 CONTAINER_PATH_PREFIX = os.getenv("MARKER_CONTAINER_PATH_PREFIX", "/documents")
 
+# ─── PR-G2-3 LLM 경계 폴백 (scaffold-first, 기본 OFF) ───
+# PRESEGMENT_LLM_FALLBACK: 기본 "false". OFF 면 deterministic 경로만(=오늘과 동일 — 애매하면
+# 단일문서). ON 이면 deterministic 이 '명확한 번들' 을 못 만든 대형 PDF(page_count >=
+# MIN_BUNDLE_PAGES) 에 한해 off-card Qwen(맥북, 라우터 :8890 경유)에게 경계를 제안받아
+# **동일 검증 게이트(_is_clear_bundle)** 통과 시에만 deterministic 과 같은 자식 생성 경로로 분할.
+# 검증 실패/파싱 실패/is_bundle=false = 단일문서(오늘과 동일) + presegment_llm_rejected 로깅.
+PRESEGMENT_LLM_FALLBACK = os.getenv("PRESEGMENT_LLM_FALLBACK", "false").lower() in (
+    "1", "true", "yes", "on",
+)
+# LLM 에 보내는 per-page 샘플의 page 당 char 상한 (heading/첫줄만 — 본문 미전송).
+PRESEGMENT_LLM_PAGE_CHARS = int(os.getenv("PRESEGMENT_LLM_PAGE_CHARS", "80"))
+# 전체 page-sample 블록의 char 상한 (수 KB 가드 — 초과 시 잘라냄, 본문 누출/페이로드 폭발 방지).
+PRESEGMENT_LLM_SAMPLE_CHARS = int(os.getenv("PRESEGMENT_LLM_SAMPLE_CHARS", "12000"))
+
+# 경계 폴백 프롬프트 (app/prompts/presegment_boundaries.txt). system 지시 + 1-based inclusive·
+# 전범위 커버·무중첩 규칙. {page_count}/{page_samples} 를 str.replace 로 주입.
+_PRESEGMENT_PROMPT_PATH = Path(__file__).parent.parent / "prompts" / "presegment_boundaries.txt"
+
+
+class Segment(BaseModel):
+    """LLM 이 제안하는 1-based inclusive page 범위 한 조각."""
+
+    start_page: int
+    end_page: int
+    title: str | None = None
+
+
+class SegmentationOutput(BaseModel):
+    """presegment_boundaries 응답 스키마. parse_json_response → model_validate."""
+
+    is_bundle: bool = False
+    segments: list[Segment] = []
+    confidence: float | None = None
+
 
 def _resolve_path(file_path: str) -> Path | None:
     """NFC(DB) vs NFD(NFS) 한글 경로 차이 흡수. thumbnail_worker._resolve_path 와 동일 패턴."""
@@ -205,6 +249,216 @@ async def _ensure_child_extract(session: AsyncSession, child_id: int) -> None:
     await enqueue_stage(session, child_id, "extract")
 
 
+async def _create_children(
+    doc: Document, segments: list[dict], session: AsyncSession
+) -> int:
+    """검증된 segments 로 자식 N개 생성 + lineage + extract enqueue + 부모 표식 (멱등).
+
+    deterministic '명확한 번들' 경로와 LLM 폴백 경로가 공유하는 단일 자식 생성 경로.
+    호출 전 segments 는 반드시 _is_clear_bundle 검증을 통과해야 한다(여기선 재검증 X).
+    commit 까지 수행. 반환값 = 실제 생성한 자식 수(이미 존재해 수렴만 한 경우 0).
+    """
+    # ─── 멱등 체크: 이미 자식이 있으면 수렴만 (재생성 금지) ───
+    existing_children = (
+        await session.execute(
+            select(DocumentLineage.derived_document_id).where(
+                DocumentLineage.source_document_id == doc.id,
+                DocumentLineage.relation_type == "segmented_from",
+            )
+        )
+    ).scalars().all()
+
+    if existing_children:
+        # 부모 표식이 누락된 경우 보정(이전 부분실패 복구).
+        if doc.presegment_role != "parent":
+            doc.presegment_role = "parent"
+        for child_id in existing_children:
+            await _ensure_child_extract(session, child_id)
+        await session.commit()
+        logger.info(
+            f"[presegment] id={doc.id} children already exist "
+            f"(n={len(existing_children)}) → converge(ensure extract), no re-create"
+        )
+        return 0
+
+    # ─── 자식 N개 생성 + lineage + extract enqueue ───
+    created_ids: list[int] = []
+    for seg in segments:
+        start, end = seg["start_page"], seg["end_page"]
+        child = Document(
+            # 후보 A: 자식 file_path = unique 합성값 `{부모경로}#p{s}-{e}` (uq_documents_file_path
+            # 충돌 회피). 실파일은 bundle_source_path() 로 복원(부모 경로). 물리 분할 없음 —
+            # 자식은 bundle_page_start/end 로 부모 파일을 슬라이스.
+            file_path=f"{doc.file_path}#p{start}-{end}",
+            file_hash=_child_file_hash(doc.file_hash, start, end),
+            file_format=doc.file_format,
+            file_size=doc.file_size,
+            file_type=doc.file_type,
+            import_source=doc.import_source,
+            original_filename=doc.original_filename,
+            source_channel=doc.source_channel,
+            category=doc.category,
+            data_origin=doc.data_origin,
+            doc_purpose=doc.doc_purpose,
+            # 안전 자료실 축은 부모에서 상속(분할이 자료유형/관할을 바꾸지 않음).
+            material_type=doc.material_type,
+            jurisdiction=doc.jurisdiction,
+            title=_child_title(doc, seg),
+            bundle_page_start=start,
+            bundle_page_end=end,
+            presegment_role="child",
+        )
+        session.add(child)
+        await session.flush()  # child.id 확보
+        created_ids.append(child.id)
+
+        session.add(
+            DocumentLineage(
+                source_document_id=doc.id,
+                derived_document_id=child.id,
+                relation_type="segmented_from",
+                meta={"start_page": start, "end_page": end},
+            )
+        )
+        # 자식 extract 는 워커가 직접 enqueue (부모는 'parent' 라 extract 로 흐르지 않음).
+        await enqueue_stage(session, child.id, "extract")
+
+    # 부모 = 파일 홀더. presegment→extract 전이는 enqueue_next_stage 가 'parent' 면 억제.
+    doc.presegment_role = "parent"
+    await session.commit()
+
+    logger.info(
+        f"[presegment] id={doc.id} SPLIT into {len(created_ids)} children "
+        f"child_ids={created_ids}"
+    )
+    return len(created_ids)
+
+
+def _segments_from_output(out: "SegmentationOutput") -> list[dict]:
+    """SegmentationOutput.segments(Pydantic) → _is_clear_bundle / _create_children 가 쓰는 dict 형태."""
+    return [
+        {"start_page": s.start_page, "end_page": s.end_page, "title": (s.title or "")}
+        for s in out.segments
+    ]
+
+
+def _page_samples(pdf, page_count: int) -> str:
+    """LLM 입력용 compact per-page 샘플 — page 당 heading/첫줄만(`p{n}: {firstline}`).
+
+    PyMuPDF page.get_text() 로 page 별 텍스트를 스트리밍하되 page 당 첫 비공백 줄만,
+    PRESEGMENT_LLM_PAGE_CHARS 로 잘라 본문 누출 차단. 전체 블록은 PRESEGMENT_LLM_SAMPLE_CHARS
+    가드로 상한(수 KB) — 초과 시 그 지점에서 중단(앞쪽 페이지 우선 보존).
+    """
+    lines: list[str] = []
+    total = 0
+    for i in range(page_count):
+        try:
+            text = pdf[i].get_text() or ""
+        except Exception:
+            text = ""
+        first = ""
+        for ln in text.splitlines():
+            ln = ln.strip()
+            if ln:
+                first = ln
+                break
+        first = first[:PRESEGMENT_LLM_PAGE_CHARS]
+        entry = f"p{i + 1}: {first}"
+        if total + len(entry) + 1 > PRESEGMENT_LLM_SAMPLE_CHARS:
+            break
+        lines.append(entry)
+        total += len(entry) + 1
+    return "\n".join(lines)
+
+
+async def _llm_boundary_fallback(
+    doc: Document, source: Path, page_count: int, session: AsyncSession
+) -> bool:
+    """애매 + 대형(ToC-less 등) PDF 에 대해 off-card Qwen 으로 경계 제안 → 검증 → 분할.
+
+    반환 True = LLM 경로가 분할을 수행(또는 멱등 수렴)했으므로 호출자는 추가 처리 없이 return.
+    반환 False = is_bundle=false / 파싱 실패 / 검증 실패 → 호출자는 단일문서(오늘과 동일) 처리.
+    맥북 불가(503/연결/절단)는 call_deep_or_defer 가 StageDeferred 로 raise → 큐 재시도(백오프).
+    silent fallback 금지 — deep 슬롯 외 다른 backend 자동 호출 안 함.
+    """
+    import fitz  # PyMuPDF — deterministic 경로와 동일 의존
+
+    # per-page 샘플은 파일을 다시 열어 스트리밍(deterministic with 블록과 분리해 그 경로 무회귀).
+    try:
+        with fitz.open(str(source)) as pdf:
+            samples = _page_samples(pdf, page_count)
+    except Exception as exc:
+        logger.warning(
+            f"[presegment] id={doc.id} llm fallback sample 실패 "
+            f"({type(exc).__name__}: {exc}) → single doc(extract)"
+        )
+        return False
+
+    try:
+        template = _PRESEGMENT_PROMPT_PATH.read_text(encoding="utf-8")
+    except Exception as exc:
+        logger.warning(
+            f"[presegment] id={doc.id} prompt 로드 실패 ({type(exc).__name__}: {exc}) "
+            f"→ single doc(extract)"
+        )
+        return False
+
+    prompt = template.replace("{page_count}", str(page_count)).replace(
+        "{page_samples}", samples
+    )
+
+    # off-card 호출 — call_deep_or_defer 가 deep 슬롯(맥북, 라우터 :8890, model=qwen-macbook)
+    # 으로 라우팅. 맥북 불가는 StageDeferred 로 전파(여기서 잡지 않음 → 큐가 보류/백오프).
+    # classify_worker 와 동일하게 AIClient() 인스턴스화.
+    client = AIClient()
+    try:
+        raw = await call_deep_or_defer(client, prompt)
+    finally:
+        await client.close()
+
+    parsed = parse_json_response(raw)
+    if not parsed:
+        logger.info(
+            f"[presegment] presegment_llm_rejected id={doc.id} "
+            f"reason=parse_failed raw={raw[:160]!r} → single doc(extract)"
+        )
+        return False
+
+    try:
+        out = SegmentationOutput.model_validate(parsed)
+    except (ValidationError, ValueError, TypeError) as exc:
+        logger.info(
+            f"[presegment] presegment_llm_rejected id={doc.id} "
+            f"reason=schema_invalid({type(exc).__name__}) → single doc(extract)"
+        )
+        return False
+
+    if not out.is_bundle:
+        logger.info(
+            f"[presegment] presegment_llm_rejected id={doc.id} "
+            f"reason=is_bundle_false → single doc(extract)"
+        )
+        return False
+
+    segments = _segments_from_output(out)
+    clear, reason = _is_clear_bundle(segments, page_count)
+    if not clear:
+        # LLM 출력을 그대로 믿지 않음 — deterministic 과 동일 게이트 미달이면 단일문서.
+        logger.info(
+            f"[presegment] presegment_llm_rejected id={doc.id} "
+            f"reason={reason} n={len(segments)} pages={page_count} → single doc(extract)"
+        )
+        return False
+
+    n = await _create_children(doc, segments, session)
+    logger.info(
+        f"[presegment] id={doc.id} LLM-SPLIT accepted "
+        f"(pages={page_count} n={len(segments)} created={n} "
+        f"confidence={out.confidence})"
+    )
+    return True
+
+
 async def process(document_id: int, session: AsyncSession) -> None:
     """presegment stage 워커 진입점. queue_consumer 가 호출.
 
@@ -269,8 +523,17 @@ async def process(document_id: int, session: AsyncSession) -> None:
     segments = _level1_segments(toc, page_count)
 
     if not segments:
-        # 큰 PDF 인데 ToC 없음/level-1 없음 = 애매(LLM fallback 대상, 후속 PR).
-        # 이 PR 은 기본 = 단일문서로 처리하고 사유를 남긴다.
+        # 큰 PDF 인데 ToC 없음/level-1 없음 = 애매. flag ON 이면 LLM 경계 폴백(PR-G2-3),
+        # OFF(기본) 이면 오늘과 동일 — 단일문서로 처리하고 사유를 남긴다.
+        if PRESEGMENT_LLM_FALLBACK:
+            logger.info(
+                f"[presegment] presegment_ambiguous id={document_id} "
+                f"reason=no_level1_toc pages={page_count} → LLM fallback"
+            )
+            if await _llm_boundary_fallback(doc, source, page_count, session):
+                return
+            # LLM 이 분할하지 않음(is_bundle=false / 검증·파싱 실패) — 단일문서.
+            return
         logger.info(
             f"[presegment] presegment_ambiguous id={document_id} "
             f"reason=no_level1_toc pages={page_count} → single doc(extract)"
@@ -279,84 +542,21 @@ async def process(document_id: int, session: AsyncSession) -> None:
 
     clear, reason = _is_clear_bundle(segments, page_count)
     if not clear:
-        # 큰 PDF + ToC 는 있으나 '명확한 번들' 기준 미달 = 애매 → 단일문서(분할 안 함).
+        # 큰 PDF + ToC 는 있으나 '명확한 번들' 기준 미달 = 애매. flag ON 이면 LLM 경계 폴백,
+        # OFF(기본) 이면 오늘과 동일 — 단일문서(분할 안 함).
+        if PRESEGMENT_LLM_FALLBACK:
+            logger.info(
+                f"[presegment] presegment_ambiguous id={document_id} "
+                f"reason={reason} pages={page_count} level1={len(segments)} → LLM fallback"
+            )
+            if await _llm_boundary_fallback(doc, source, page_count, session):
+                return
+            return
         logger.info(
             f"[presegment] presegment_ambiguous id={document_id} "
             f"reason={reason} pages={page_count} level1={len(segments)} → single doc(extract)"
         )
         return
 
-    # ─── (5) 명확한 번들 — 멱등 체크: 이미 자식이 있으면 수렴만 ───
-    existing_children = (
-        await session.execute(
-            select(DocumentLineage.derived_document_id).where(
-                DocumentLineage.source_document_id == doc.id,
-                DocumentLineage.relation_type == "segmented_from",
-            )
-        )
-    ).scalars().all()
-
-    if existing_children:
-        # 부모 표식이 누락된 경우 보정(이전 부분실패 복구).
-        if doc.presegment_role != "parent":
-            doc.presegment_role = "parent"
-        for child_id in existing_children:
-            await _ensure_child_extract(session, child_id)
-        await session.commit()
-        logger.info(
-            f"[presegment] id={document_id} children already exist "
-            f"(n={len(existing_children)}) → converge(ensure extract), no re-create"
-        )
-        return
-
-    # ─── (6) 자식 N개 생성 + lineage + extract enqueue ───
-    n = len(segments)
-    created_ids: list[int] = []
-    for seg in segments:
-        start, end = seg["start_page"], seg["end_page"]
-        child = Document(
-            # 후보 A: 자식 file_path = unique 합성값 `{부모경로}#p{s}-{e}` (uq_documents_file_path
-            # 충돌 회피). 실파일은 bundle_source_path() 로 복원(부모 경로). 물리 분할 없음 —
-            # 자식은 bundle_page_start/end 로 부모 파일을 슬라이스.
-            file_path=f"{doc.file_path}#p{start}-{end}",
-            file_hash=_child_file_hash(doc.file_hash, start, end),
-            file_format=doc.file_format,
-            file_size=doc.file_size,
-            file_type=doc.file_type,
-            import_source=doc.import_source,
-            original_filename=doc.original_filename,
-            source_channel=doc.source_channel,
-            category=doc.category,
-            data_origin=doc.data_origin,
-            doc_purpose=doc.doc_purpose,
-            # 안전 자료실 축은 부모에서 상속(분할이 자료유형/관할을 바꾸지 않음).
-            material_type=doc.material_type,
-            jurisdiction=doc.jurisdiction,
-            title=_child_title(doc, seg),
-            bundle_page_start=start,
-            bundle_page_end=end,
-            presegment_role="child",
-        )
-        session.add(child)
-        await session.flush()  # child.id 확보
-        created_ids.append(child.id)
-
-        session.add(
-            DocumentLineage(
-                source_document_id=doc.id,
-                derived_document_id=child.id,
-                relation_type="segmented_from",
-                meta={"start_page": start, "end_page": end},
-            )
-        )
-        # 자식 extract 는 워커가 직접 enqueue (부모는 'parent' 라 extract 로 흐르지 않음).
-        await enqueue_stage(session, child.id, "extract")
-
-    # 부모 = 파일 홀더. presegment→extract 전이는 enqueue_next_stage 가 'parent' 면 억제.
-    doc.presegment_role = "parent"
-    await session.commit()
-
-    logger.info(
-        f"[presegment] id={document_id} SPLIT into {n} children "
-        f"(pages={page_count}) child_ids={created_ids}"
-    )
+    # ─── (5) 명확한 번들 (deterministic) — 공유 자식 생성 경로 (멱등 수렴 포함) ───
+    await _create_children(doc, segments, session)

tests/test_presegment_llm.py

@@ -0,0 +1,400 @@
+"""PR-G2-3 — presegment LLM 경계 폴백 단위 테스트.
+
+scaffold-first 안전성 박제:
+  (a) parse_json_response + SegmentationOutput 가 대표 fixture(ToC-less 120p → 3 segments) 검증
+  (b) 검증 게이트(_is_clear_bundle)가 정상 응답 수락 / 비정상(중첩·gap·tiny child·N>MAX) 거부
+  (c) flag OFF(기본) → LLM 절대 호출 안 함(call_deep count==0), flag ON → 호출됨(positive control)
+
+DB·PyMuPDF 불요(unit) — AsyncSession 은 최소 fake, fitz 는 sys.modules 주입 fake.
+라이브 LLM 호출 없음(call_deep 는 fixture 반환 monkeypatch). worker-process 레벨 E2E(실 PDF
+번들 분할, 보류 백오프 DB 기록)는 GPU 라이브 게이트에서 별도 실측.
+[[feedback_external_api_fixture_first]] / [[feedback_scaffold_first_for_external_cost_pr]]
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "app"))
+
+from ai.client import parse_json_response  # noqa: E402
+import workers.presegment_worker as pw  # noqa: E402
+from workers.presegment_worker import (  # noqa: E402
+    SegmentationOutput,
+    _is_clear_bundle,
+    _segments_from_output,
+)
+
+# ─── 대표 fixture: ToC-less 120p 번들 → 3 segments (1-based inclusive, 전범위·무중첩) ───
+GOOD_LLM_JSON = json.dumps(
+    {
+        "is_bundle": True,
+        "segments": [
+            {"start_page": 1, "end_page": 40, "title": "문서 A"},
+            {"start_page": 41, "end_page": 85, "title": "문서 B"},
+            {"start_page": 86, "end_page": 120, "title": "문서 C"},
+        ],
+        "confidence": 0.82,
+    },
+    ensure_ascii=False,
+)
+
+PAGE_COUNT = 120
+
+
+# ─── (a) parse_json_response + SegmentationOutput 검증 ──────────────────────
+
+
+def test_parse_and_validate_good_fixture():
+    parsed = parse_json_response(GOOD_LLM_JSON)
+    assert parsed is not None
+    out = SegmentationOutput.model_validate(parsed)
+    assert out.is_bundle is True
+    assert len(out.segments) == 3
+    assert out.segments[0].start_page == 1
+    assert out.segments[-1].end_page == PAGE_COUNT
+    assert out.confidence == pytest.approx(0.82)
+
+
+def test_parse_tolerates_think_and_fence():
+    """house parse_json_response 가 <think> + ```json fence 를 벗겨낸다."""
+    wrapped = f"<think>분석중...</think>\n```json\n{GOOD_LLM_JSON}\n```"
+    parsed = parse_json_response(wrapped)
+    out = SegmentationOutput.model_validate(parsed)
+    assert out.is_bundle is True and len(out.segments) == 3
+
+
+# ─── (b) 검증 게이트 accept / reject ────────────────────────────────────────
+
+
+def _segments(*spans):
+    return [{"start_page": s, "end_page": e, "title": ""} for (s, e) in spans]
+
+
+def test_gate_accepts_good():
+    out = SegmentationOutput.model_validate(parse_json_response(GOOD_LLM_JSON))
+    segs = _segments_from_output(out)
+    clear, reason = _is_clear_bundle(segs, PAGE_COUNT)
+    assert clear is True, reason
+    assert reason == ""
+
+
+def test_gate_rejects_overlap():
+    # 41 이어야 할 두번째 start 가 40 으로 중첩
+    clear, reason = _is_clear_bundle(_segments((1, 40), (40, 85), (86, 120)), PAGE_COUNT)
+    assert clear is False
+    assert "non_contiguous" in reason
+
+
+def test_gate_rejects_gap():
+    # 40 다음이 42 로 시작 → 41 빈틈 (non_contiguous 로 검출)
+    clear, reason = _is_clear_bundle(_segments((1, 40), (42, 85), (86, 120)), PAGE_COUNT)
+    assert clear is False
+    assert "non_contiguous" in reason
+
+
+def test_gate_rejects_tiny_child():
+    # 두번째 자식 41..43 = 3p < MIN_CHILD_PAGES(5)
+    clear, reason = _is_clear_bundle(_segments((1, 40), (41, 43), (44, 120)), PAGE_COUNT)
+    assert clear is False
+    assert "child_too_small" in reason
+
+
+def test_gate_rejects_coverage_not_full():
+    # 마지막이 page_count 에 못 미침
+    clear, reason = _is_clear_bundle(_segments((1, 40), (41, 85), (86, 110)), PAGE_COUNT)
+    assert clear is False
+    assert "last_end_not_page_count" in reason
+
+
+def test_gate_rejects_too_many_children():
+    # N > MAX_CHILDREN — 각 자식 MIN_CHILD_PAGES 만족시키되 개수만 초과
+    n = pw.MAX_CHILDREN + 1
+    pc = n * pw.MIN_CHILD_PAGES
+    spans = [
+        (i * pw.MIN_CHILD_PAGES + 1, (i + 1) * pw.MIN_CHILD_PAGES) for i in range(n)
+    ]
+    clear, reason = _is_clear_bundle(_segments(*spans), pc)
+    assert clear is False
+    assert "too_many_children" in reason
+
+
+def test_gate_rejects_single_segment():
+    clear, reason = _is_clear_bundle(_segments((1, 120)), PAGE_COUNT)
+    assert clear is False
+    assert "too_few_level1_entries" in reason
+
+
+# ─── 공통 fake (DB / PyMuPDF) ──────────────────────────────────────────────
+
+
+class _FakeDoc:
+    """presegment 가 읽는 Document 필드만 가진 최소 stand-in."""
+
+    def __init__(self, doc_id=1):
+        self.id = doc_id
+        self.file_path = "PKM/bundle.pdf"
+        self.file_hash = "deadbeef"
+        self.file_format = "pdf"
+        self.file_size = 123
+        self.file_type = "document"
+        self.import_source = "upload"
+        self.original_filename = "bundle.pdf"
+        self.source_channel = None
+        self.category = None
+        self.data_origin = None
+        self.doc_purpose = None
+        self.material_type = None
+        self.jurisdiction = None
+        self.title = "번들"
+        self.presegment_role = None
+        self.bundle_page_start = None
+        self.bundle_page_end = None
+        self.extracted_at = None
+        self.extracted_text = None
+
+
+class _ScalarResult:
+    def __init__(self, rows):
+        self._rows = rows
+
+    def scalars(self):
+        return self
+
+    def all(self):
+        return list(self._rows)
+
+
+class _FakeSession:
+    """_create_children / process 가 쓰는 AsyncSession 표면만 구현.
+
+    execute() = 기존 자식 lineage 조회 → 빈 결과(첫 분할). add/flush 로 child.id 부여.
+    get() = document_id → 미리 등록한 doc, child_id → 생성된 child.
+    """
+
+    def __init__(self, doc):
+        self._docs = {doc.id: doc}
+        self.added = []
+        self.commits = 0
+        self.enqueued = []  # enqueue_stage monkeypatch 가 채움
+        self._next_id = 1000
+
+    async def get(self, _model, oid):
+        return self._docs.get(oid)
+
+    async def execute(self, _stmt):
+        # _create_children 의 기존 자식 조회 → 항상 빈(첫 분할). enqueue_stage 는 monkeypatch.
+        return _ScalarResult([])
+
+    def add(self, obj):
+        self.added.append(obj)
+        # child Document 에 id 부여 (flush 대용 — _FakeDoc/실 Document 모두 setattr 가능)
+        if getattr(obj, "id", None) is None and hasattr(obj, "presegment_role"):
+            self._next_id += 1
+            obj.id = self._next_id
+            self._docs[obj.id] = obj
+
+    async def flush(self):
+        for obj in self.added:
+            if getattr(obj, "id", None) is None and hasattr(obj, "presegment_role"):
+                self._next_id += 1
+                obj.id = self._next_id
+                self._docs[obj.id] = obj
+
+    async def commit(self):
+        self.commits += 1
+
+
+def _install_fake_fitz(monkeypatch, *, page_count=PAGE_COUNT, toc=None, first_lines=None):
+    """sys.modules['fitz'] 에 fake 주입 — worker 의 `import fitz` 가 이걸 받게 한다."""
+    toc = toc or []
+
+    class _FakePage:
+        def __init__(self, idx):
+            self._idx = idx
+
+        def get_text(self):
+            if first_lines and self._idx < len(first_lines):
+                return first_lines[self._idx]
+            return f"page {self._idx + 1} body text"
+
+    class _FakePdf:
+        def __init__(self):
+            self.page_count = page_count
+
+        def get_toc(self, simple=True):
+            return list(toc)
+
+        def __getitem__(self, idx):
+            return _FakePage(idx)
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *exc):
+            return False
+
+    fake = types.ModuleType("fitz")
+    fake.open = lambda *_a, **_k: _FakePdf()
+    monkeypatch.setitem(sys.modules, "fitz", fake)
+    return fake
+
+
+class _SpyClient:
+    """AIClient stand-in — call_deep 호출 횟수 카운트 + 지정 응답 반환."""
+
+    calls = 0
+    response = GOOD_LLM_JSON
+
+    def __init__(self):
+        type(self).calls += 1  # 인스턴스화 자체는 비용 아님 — 호출 카운트는 call_deep 기준
+
+    async def call_deep(self, prompt, system=None):
+        type(self)._deep_calls += 1
+        return type(self).response
+
+    async def close(self):
+        pass
+
+
+@pytest.fixture(autouse=True)
+def _reset_spy():
+    _SpyClient.calls = 0
+    _SpyClient._deep_calls = 0
+    _SpyClient.response = GOOD_LLM_JSON
+    yield
+
+
+# ─── (b) _llm_boundary_fallback 수락/거부 (mocked LLM) ──────────────────────
+
+
+@pytest.mark.asyncio
+async def test_fallback_accepts_good_and_creates_children(monkeypatch):
+    """정상 LLM 응답 → 게이트 통과 → _create_children 가 3 자식 + parent 표식."""
+    _install_fake_fitz(monkeypatch)
+    monkeypatch.setattr(pw, "AIClient", _SpyClient)
+    # enqueue_stage 는 DB 의존 — no-op 으로 대체 (호출 인자만 기록)
+    enq = []
+
+    async def _fake_enqueue(session, doc_id, stage, **kw):
+        enq.append((doc_id, stage))
+        return True
+
+    monkeypatch.setattr(pw, "enqueue_stage", _fake_enqueue)
+
+    doc = _FakeDoc()
+    session = _FakeSession(doc)
+    ok = await pw._llm_boundary_fallback(doc, Path("/tmp/bundle.pdf"), PAGE_COUNT, session)
+
+    assert ok is True
+    assert _SpyClient._deep_calls == 1
+    # 자식 3개 생성 + parent 표식 + lineage 3 + commit
+    children = [o for o in session.added if getattr(o, "presegment_role", None) == "child"]
+    assert len(children) == 3
+    assert doc.presegment_role == "parent"
+    assert sum(1 for o in session.added if o.__class__.__name__ == "DocumentLineage") == 3
+    assert {s for (_id, s) in enq} == {"extract"}
+
+
+@pytest.mark.asyncio
+async def test_fallback_rejects_bad_segments(monkeypatch):
+    """LLM 이 중첩 경계 반환 → 게이트 거부 → False + 자식 0 (단일문서)."""
+    _install_fake_fitz(monkeypatch)
+    bad = json.dumps({
+        "is_bundle": True,
+        "segments": [
+            {"start_page": 1, "end_page": 40},
+            {"start_page": 40, "end_page": 85},   # 중첩
+            {"start_page": 86, "end_page": 120},
+        ],
+    })
+    _SpyClient.response = bad
+    monkeypatch.setattr(pw, "AIClient", _SpyClient)
+
+    async def _fake_enqueue(*a, **k):
+        return True
+
+    monkeypatch.setattr(pw, "enqueue_stage", _fake_enqueue)
+
+    doc = _FakeDoc()
+    session = _FakeSession(doc)
+    ok = await pw._llm_boundary_fallback(doc, Path("/tmp/b.pdf"), PAGE_COUNT, session)
+
+    assert ok is False
+    assert _SpyClient._deep_calls == 1
+    assert [o for o in session.added if getattr(o, "presegment_role", None) == "child"] == []
+    assert doc.presegment_role is None
+
+
+@pytest.mark.asyncio
+async def test_fallback_rejects_is_bundle_false(monkeypatch):
+    """is_bundle=false → 호출은 했으나 분할 안 함(False, 자식 0)."""
+    _install_fake_fitz(monkeypatch)
+    _SpyClient.response = json.dumps({"is_bundle": False, "segments": []})
+    monkeypatch.setattr(pw, "AIClient", _SpyClient)
+
+    async def _fake_enqueue(*a, **k):
+        return True
+
+    monkeypatch.setattr(pw, "enqueue_stage", _fake_enqueue)
+
+    doc = _FakeDoc()
+    session = _FakeSession(doc)
+    ok = await pw._llm_boundary_fallback(doc, Path("/tmp/b.pdf"), PAGE_COUNT, session)
+    assert ok is False
+    assert _SpyClient._deep_calls == 1
+    assert doc.presegment_role is None
+
+
+# ─── (c) flag gating — OFF=호출 0 (deployed default 무변), ON=호출됨 ───────────
+
+
+@pytest.mark.asyncio
+async def test_flag_off_never_calls_llm(monkeypatch):
+    """PRESEGMENT_LLM_FALLBACK=False(기본) → 큰 ToC-less PDF 도 LLM 미호출 = 오늘과 동일."""
+    monkeypatch.setattr(pw, "PRESEGMENT_LLM_FALLBACK", False)
+    _install_fake_fitz(monkeypatch, page_count=120, toc=[])  # 대형 + level-1 ToC 없음 = 애매
+    monkeypatch.setattr(pw, "AIClient", _SpyClient)
+    monkeypatch.setattr(pw, "_resolve_path", lambda raw: Path("/tmp/bundle.pdf"))
+
+    async def _fake_enqueue(*a, **k):
+        return True
+
+    monkeypatch.setattr(pw, "enqueue_stage", _fake_enqueue)
+
+    doc = _FakeDoc()
+    session = _FakeSession(doc)
+    await pw.process(doc.id, session)
+
+    assert _SpyClient._deep_calls == 0          # ★ LLM 절대 호출 안 됨
+    assert doc.presegment_role is None          # 단일문서 (분할 안 함)
+    assert session.commits == 0
+
+
+@pytest.mark.asyncio
+async def test_flag_on_calls_llm_and_splits(monkeypatch):
+    """positive control — flag ON 이면 같은 입력에 LLM 호출 + 게이트 통과 시 분할."""
+    monkeypatch.setattr(pw, "PRESEGMENT_LLM_FALLBACK", True)
+    _install_fake_fitz(monkeypatch, page_count=120, toc=[])
+    _SpyClient.response = GOOD_LLM_JSON
+    monkeypatch.setattr(pw, "AIClient", _SpyClient)
+    monkeypatch.setattr(pw, "_resolve_path", lambda raw: Path("/tmp/bundle.pdf"))
+
+    async def _fake_enqueue(*a, **k):
+        return True
+
+    monkeypatch.setattr(pw, "enqueue_stage", _fake_enqueue)
+
+    doc = _FakeDoc()
+    session = _FakeSession(doc)
+    await pw.process(doc.id, session)
+
+    assert _SpyClient._deep_calls == 1          # LLM 호출됨
+    assert doc.presegment_role == "parent"      # 분할 수행
+    children = [o for o in session.added if getattr(o, "presegment_role", None) == "child"]
+    assert len(children) == 3