@oh-story/dsh 0.1.8 → 0.1.10

package/lib/oh-story/skills/story-long-analyze/scripts/manage_analysis_run.py

@@ -0,0 +1,1106 @@
+#!/usr/bin/env python3
+"""Plan, commit, split, and recover long-analysis batches.
+
+Runtime state lives only in the managed block of ``_progress.md``. Plans are
+printed as JSON and are never persisted. Batch caches are complete recovery
+evidence, not a second state database. An existing chapter summary is never
+overwritten: to redo a chapter, delete its summary and plan again.
+"""
+
+from __future__ import annotations
+
+import argparse
+import codecs
+import csv
+import hashlib
+import json
+import os
+import re
+import sys
+import tempfile
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Optional, Sequence, Set, Tuple
+
+from inspect_existing_assets import inspect
+
+
+STATE_START = "<!-- story-long-analyze:runtime-state:start -->"
+STATE_END = "<!-- story-long-analyze:runtime-state:end -->"
+CACHE_START = "<!-- story-long-analyze:cache:start -->"
+CACHE_END = "<!-- story-long-analyze:cache:end -->"
+MODEL_START = "<!-- MODEL_OUTPUT_START -->"
+MODEL_END = "<!-- MODEL_OUTPUT_END -->"
+CHAPTER_TOKEN_RE = re.compile(r"<!--\s*CHAPTER_(START|END):(\d+)\s*-->")
+CHAPTER_BLOCK_RE = re.compile(
+    r"<!--\s*CHAPTER_START:(\d+)\s*-->\s*(.*?)\s*<!--\s*CHAPTER_END:\1\s*-->", re.DOTALL
+)
+BATCH_ID_RE = re.compile(r"^(RAW|REUSE)-(\d+)-(\d+)$")
+COMPACT_FIELDS = (
+    "概要", "因果", "关键行动", "局面结果", "涉及人物", "信息变化", "状态变化",
+    "三维节奏", "章尾钩子", "证据",
+)
+POINT_HEADER_RE = re.compile(r"^P(\d+)\s+\*\*(.+?)\*\*\s*[：:]\s*(.+)$")
+POINT_TAG_RE = re.compile(r"^主题标签\s*[：:]?\s*([^|｜]*?)\s*[|｜]\s*基调\s*[：:]?\s*(.*?)\s*$")
+THEMES = ("爱情", "亲情", "友情", "权力", "金钱", "成长", "复仇", "悬念", "搞笑", "热血", "日常", "其他")
+TONES = ("紧张", "轻松", "悲伤", "热血", "爽", "甜", "温馨", "恐怖", "压抑", "其他")
+POINT_TYPES = ("转折点", "信息揭示", "冲突", "解决", "铺垫", "行动", "对话", "状态变化")
+THEME_ALIASES = {"恋爱": "爱情", "权谋": "权力", "政治": "权力", "幽默": "搞笑"}
+TONE_ALIASES = {"悲痛": "悲伤", "伤感": "悲伤", "悲愤": "悲伤", "痛快": "爽", "解气": "爽", "惊悚": "恐怖",
+                "恐惧": "恐怖", "危险": "紧张", "危急": "紧张", "绝望": "压抑", "无力": "压抑", "释然": "轻松"}
+POINT_ALIASES = {"揭示": "信息揭示", "转折": "转折点", "变化": "状态变化", "动作": "行动",
+                 "交谈": "对话", "化解": "解决"}
+MIN_PLOT_POINTS = 10
+MAX_PLOT_POINTS = 30
+MAX_CHAPTERS = 3
+MAX_CHARS = 25_000
+# Stage 3-6 each run once over complete Stage 2 output; their rows decide 最终状态.
+FINAL_STAGES = ("stage3", "stage4", "stage5", "stage6")
+LEGACY_FINAL_RE = re.compile(r"(?m)^([ \t]*(?:[-*][ \t]*)?最终状态[ \t]*[：:][ \t]*)([A-Za-z0-9_]+)")
+
+
+class RunError(ValueError):
+    def __init__(self, code: str, detail: str, author_message: Optional[str] = None) -> None:
+        super().__init__(detail)
+        self.code = code
+        self.detail = detail
+        self.author_message = author_message
+
+
+def sha256(data: bytes) -> str:
+    return hashlib.sha256(data).hexdigest()
+
+
+def normalized(text: str) -> str:
+    return text.replace("\r\n", "\n").replace("\r", "\n")
+
+
+def atomic_write(path: Path, data: bytes) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    fd, temporary = tempfile.mkstemp(prefix=".%s." % path.name, suffix=".tmp", dir=str(path.parent))
+    try:
+        with os.fdopen(fd, "wb") as handle:
+            handle.write(data)
+            handle.flush()
+            os.fsync(handle.fileno())
+        os.replace(temporary, str(path))
+    except BaseException:
+        try:
+            os.unlink(temporary)
+        except OSError:
+            pass
+        raise
+
+
+def require_root(root: Path) -> Path:
+    root = root.resolve()
+    if not root.exists():
+        raise RunError("root_not_found", str(root))
+    if not root.is_dir():
+        raise RunError("root_is_not_directory", str(root))
+    try:
+        next(root.iterdir(), None)
+    except OSError as exc:
+        raise RunError("root_unreadable", str(exc)) from exc
+    return root
+
+
+def read_index(root: Path, index_arg: Optional[Path] = None) -> List[Dict[str, Any]]:
+    path = (index_arg.resolve() if index_arg else root / "chapter_index.csv")
+    if not path.is_file():
+        raise RunError("chapter_index_required", str(path))
+    rows = []  # type: List[Dict[str, Any]]
+    try:
+        with path.open("r", encoding="utf-8-sig", newline="") as handle:
+            reader = csv.DictReader(handle)
+            required = {"chapter", "start_line", "end_line", "char_count", "source_locator", "chapter_sha256"}
+            missing = required - set(reader.fieldnames or ())
+            if missing:
+                raise RunError("chapter_index_invalid", "missing columns: %s" % ", ".join(sorted(missing)))
+            for raw in reader:
+                row = dict(raw)
+                try:
+                    row["chapter"] = int(row["chapter"])
+                    row["start_line"] = int(row["start_line"])
+                    row["end_line"] = int(row["end_line"])
+                    row["char_count"] = int(row["char_count"])
+                except (TypeError, ValueError) as exc:
+                    raise RunError("chapter_index_invalid", "numeric column invalid") from exc
+                if not re.fullmatch(r"[0-9a-f]{64}", str(row["chapter_sha256"])):
+                    raise RunError("chapter_index_invalid", "chapter_sha256 invalid")
+                rows.append(row)
+    except (OSError, UnicodeError, csv.Error) as exc:
+        raise RunError("chapter_index_unreadable", str(exc)) from exc
+    if [row["chapter"] for row in rows] != list(range(1, len(rows) + 1)):
+        raise RunError("chapter_index_invalid", "chapter ids must be continuous from 1")
+    return rows
+
+
+def range_sha256(rows: Sequence[Dict[str, Any]], start: int, end: int) -> str:
+    selected = [row for row in rows if start <= row["chapter"] <= end]
+    if [row["chapter"] for row in selected] != list(range(start, end + 1)):
+        raise RunError("range_not_in_index", "%s-%s" % (start, end))
+    payload = "range-v1\n" + "".join(
+        "%s:%s\n" % (row["chapter"], row["chapter_sha256"]) for row in selected
+    )
+    return sha256(payload.encode("ascii"))
+
+
+def decode_progress(raw: bytes) -> Tuple[str, bool, str]:
+    bom = raw.startswith(codecs.BOM_UTF8)
+    text = raw.decode("utf-8-sig")
+    newline = "\r\n" if "\r\n" in text else "\n"
+    return text, bom, newline
+
+
+def empty_state() -> Dict[str, Any]:
+    return {"batches": {}, "stages": {}}
+
+
+def table_cells(line: str) -> List[str]:
+    return [cell.strip() for cell in line.strip().strip("|").split("|")]
+
+
+def load_state(progress: Path) -> Tuple[Dict[str, Any], bytes]:
+    raw = progress.read_bytes() if progress.is_file() else b""
+    text, _, _ = decode_progress(raw)
+    state = empty_state()
+    start_count = text.count(STATE_START)
+    end_count = text.count(STATE_END)
+    if start_count != end_count or start_count > 1:
+        raise RunError(
+            "progress_state_block_invalid",
+            "expected zero or one complete managed state block; found start=%s end=%s"
+            % (start_count, end_count),
+        )
+    match = re.search(re.escape(STATE_START) + r"(.*?)" + re.escape(STATE_END), text, re.DOTALL)
+    if not match:
+        return state, raw
+    section = None
+    for line in match.group(1).splitlines():
+        if line.strip() == "### 批次状态":
+            section = "batches"
+            continue
+        if line.strip() == "### 阶段状态":
+            section = "stages"
+            continue
+        if not line.lstrip().startswith("|") or set(line.replace("|", "").replace("-", "").replace(":", "").strip()) == set():
+            continue
+        cells = table_cells(line)
+        if section == "batches" and cells and cells[0] not in {"批次ID", "---"} and len(cells) >= 7:
+            try:
+                start, end = [int(value) for value in cells[1].split("-", 1)]
+            except (ValueError, IndexError):
+                continue
+            state["batches"][cells[0]] = {
+                "batch_id": cells[0], "start": start, "end": end, "input_kind": cells[2],
+                "range_sha256": cells[3], "status": cells[4],
+                "parent": "" if cells[5] == "-" else cells[5],
+                "cache": "" if cells[6] == "-" else cells[6],
+            }
+        elif section == "stages" and cells and cells[0] not in {"阶段", "---"} and len(cells) >= 3:
+            state["stages"][cells[0]] = {"status": cells[1], "output": "" if cells[2] == "-" else cells[2]}
+    return state, raw
+
+
+def final_status(stages: Dict[str, Dict[str, str]]) -> str:
+    """Value for the ``最终状态`` line that session hooks read (Stage 3-6 rows)."""
+    statuses = [stages.get(stage, {}).get("status") for stage in FINAL_STAGES]
+    if all(status in {"completed", "completed_with_errors"} for status in statuses):
+        return "completed_with_errors" if "completed_with_errors" in statuses else "completed"
+    return "pending"
+
+
+def render_state(state: Dict[str, Any], newline: str, final_line: bool = True) -> str:
+    lines = [STATE_START, "## 长篇拆文运行状态", ""]
+    if final_line and state["stages"]:
+        lines.extend(["- 最终状态：%s" % final_status(state["stages"]), ""])
+    lines.extend(["### 批次状态",
+                  "| 批次ID | 章节范围 | 输入 | 原文范围hash | 状态 | 父批次 | 缓存 |",
+                  "|---|---|---|---|---|---|---|"])
+    batches = list(state["batches"].values())
+    batches.sort(key=lambda row: (row["start"], row["end"], row["batch_id"]))
+    for row in batches:
+        lines.append("| %s | %s-%s | %s | %s | %s | %s | %s |" % (
+            row["batch_id"], row["start"], row["end"], row["input_kind"],
+            row["range_sha256"], row["status"], row.get("parent") or "-", row.get("cache") or "-"))
+    lines.extend(["", "### 阶段状态", "| 阶段 | 状态 | 产物 |", "|---|---|---|"])
+    for stage in sorted(state["stages"]):
+        row = state["stages"][stage]
+        lines.append("| %s | %s | %s |" % (stage, row["status"], row.get("output") or "-"))
+    lines.append(STATE_END)
+    return newline.join(lines)
+
+
+def write_state(progress: Path, state: Dict[str, Any]) -> bool:
+    raw = progress.read_bytes() if progress.is_file() else b""
+    text, bom, newline = decode_progress(raw)
+    pattern = re.compile(re.escape(STATE_START) + r".*?" + re.escape(STATE_END), re.DOTALL)
+    # Hooks read the first 最终状态 in the file. A legacy project keeps its own
+    # line as that single value; the runtime only promotes it once Stage 3-6
+    # are all complete and never demotes it.
+    existing = pattern.search(text)
+    head = text[:existing.start()] if existing else text
+    legacy = LEGACY_FINAL_RE.search(head)
+    status = final_status(state["stages"])
+    if legacy and status != "pending" and legacy.group(2) != status:
+        head = head[:legacy.start(2)] + status + head[legacy.end(2):]
+        text = head + (text[existing.start():] if existing else "")
+    block = render_state(state, newline, final_line=legacy is None)
+    if pattern.search(text):
+        updated = pattern.sub(lambda _: block, text, count=1)
+    else:
+        if not text:
+            text = "# 深度拆解进度" + newline + "- schema_version: 2" + newline
+        separator = "" if text.endswith(newline + newline) else (newline if text.endswith(newline) else newline + newline)
+        updated = text + separator + block + newline
+    data = ((codecs.BOM_UTF8 if bom else b"") + updated.encode("utf-8"))
+    if data == raw:
+        return False
+    atomic_write(progress, data)
+    return True
+
+
+def summary_path(root: Path, chapter: int) -> Path:
+    return root / "章节" / ("第%s章_摘要.md" % chapter)
+
+
+def cache_complete(path: Path) -> bool:
+    try:
+        text = normalized(path.read_text(encoding="utf-8-sig"))
+    except (OSError, UnicodeError):
+        return False
+    return text.rstrip().endswith(CACHE_END) and text.count(MODEL_START) == 1 and text.count(MODEL_END) == 1
+
+
+def cache_path(root: Path, batch_id: str) -> Path:
+    return root / "_analysis_cache" / ("批次-%s.md" % batch_id)
+
+
+def completed_batch(root: Path, row: Dict[str, Any], rows: Optional[Sequence[Dict[str, Any]]]) -> bool:
+    if row.get("status") not in {"completed", "success"} or not all(summary_path(root, chapter).is_file() for chapter in range(row["start"], row["end"] + 1)):
+        return False
+    path = root / row.get("cache", "") if row.get("cache") else cache_path(root, row["batch_id"])
+    if not cache_complete(path):
+        return False
+    try:
+        metadata = parse_cache(path)
+    except (OSError, UnicodeError, RunError):
+        return False
+    if (metadata.get("batch_id"), metadata.get("start"), metadata.get("end")) != (
+        row["batch_id"], row["start"], row["end"]
+    ):
+        return False
+    if metadata.get("range_sha256") != row.get("range_sha256"):
+        return False
+    if row["input_kind"] == "raw-original":
+        return rows is not None and row.get("range_sha256") == range_sha256(rows, row["start"], row["end"])
+    return True
+
+
+def compact_ranges(chapters: Iterable[int]) -> List[Tuple[int, int]]:
+    values = sorted(set(chapters))
+    if not values:
+        return []
+    result = []
+    start = previous = values[0]
+    for value in values[1:]:
+        if value == previous + 1:
+            previous = value
+        else:
+            result.append((start, previous))
+            start = previous = value
+    result.append((start, previous))
+    return result
+
+
+def chunk_range(start: int, end: int, index_by_chapter: Optional[Dict[int, Dict[str, Any]]]) -> List[Tuple[int, int]]:
+    result = []
+    current_start = start
+    count = 0
+    chars = 0
+    previous = start - 1
+    for chapter in range(start, end + 1):
+        chapter_chars = index_by_chapter[chapter]["char_count"] if index_by_chapter else 0
+        if count and (count + 1 > MAX_CHAPTERS or chars + chapter_chars > MAX_CHARS):
+            result.append((current_start, previous))
+            current_start = chapter
+            count = 0
+            chars = 0
+        count += 1
+        chars += chapter_chars
+        previous = chapter
+    result.append((current_start, previous))
+    return result
+
+
+def invalid_batch_targets(root: Path, rows: Optional[Sequence[Dict[str, Any]]],
+                          state: Dict[str, Any]) -> Tuple[Set[int], Set[int]]:
+    """Chapters of recorded batches that no longer verify (cache lost or source changed)."""
+    raw = set()
+    reuse = set()
+    for batch in state["batches"].values():
+        chapters = set(range(batch["start"], batch["end"] + 1))
+        # Split children stay owed until committed, even when summaries exist.
+        pending_child = batch.get("parent") and batch.get("status") == "planned"
+        if not pending_child and (
+            batch.get("status") not in {"completed", "success"} or completed_batch(root, batch, rows)
+        ):
+            continue
+        if batch.get("input_kind") == "raw-original":
+            raw.update(chapters)
+        elif batch.get("input_kind") == "existing-results":
+            reuse.update(chapters)
+    return raw, reuse
+
+
+def add_recoverable_caches(root: Path, state: Dict[str, Any],
+                           rows: Optional[Sequence[Dict[str, Any]]]) -> Tuple[List[Dict[str, Any]], Set[int]]:
+    """Use complete caches in this read-only plan without persisting recovery."""
+    cache_dir = root / "_analysis_cache"
+    recoverable = []
+    covered = set()  # type: Set[int]
+    for path in sorted(cache_dir.glob("批次-*.md")) if cache_dir.is_dir() else []:
+        try:
+            metadata = parse_cache(path)
+            batch_id = metadata["batch_id"]
+            input_kind, start, end = parse_batch_id(batch_id)
+            if metadata["input_kind"] != input_kind:
+                continue
+            current_hash = metadata["range_sha256"]
+            if input_kind == "raw-original":
+                if rows is None or current_hash != range_sha256(rows, start, end):
+                    continue
+            records = parse_model_output(metadata["model_output"], start, end, input_kind)
+            missing = [chapter for chapter in range(start, end + 1) if not summary_path(root, chapter).is_file()]
+            if missing and not all(chapter in records for chapter in missing):
+                continue
+            covered.update(range(start, end + 1))
+            current = state["batches"].get(batch_id)
+            needs_state_repair = current is None or not completed_batch(root, current, rows)
+            if missing or needs_state_repair:
+                recoverable.append({"batch_id": batch_id, "cache": path.relative_to(root).as_posix(),
+                                    "missing_summary_chapters": missing,
+                                    "state_repair_required": needs_state_repair})
+                if missing:
+                    continue
+            recovered = {
+                "batch_id": batch_id, "start": start, "end": end, "input_kind": input_kind,
+                "range_sha256": current_hash, "status": "completed", "parent": "",
+                "cache": path.relative_to(root).as_posix(),
+            }
+            if current is None or not completed_batch(root, current, rows):
+                state["batches"][batch_id] = recovered
+        except (OSError, UnicodeError, RunError, ValueError):
+            continue
+    return recoverable, covered
+
+
+def plan_command(args: argparse.Namespace) -> Dict[str, Any]:
+    root = require_root(args.root)
+    report = inspect(root, args.expected_chapters)
+    state, _ = load_state(root / "_progress.md")
+    index_rows = None  # type: Optional[List[Dict[str, Any]]]
+    index_path = args.index.resolve() if args.index else root / "chapter_index.csv"
+    if index_path.is_file():
+        index_rows = read_index(root, index_path)
+    recoverable_caches, cache_covered = add_recoverable_caches(root, state, index_rows)
+    expected = report.get("expected_chapters") or (len(index_rows) if index_rows else None)
+    if not expected:
+        raise RunError("expected_chapters_unknown", "build the chapter index or pass --expected-chapters")
+    semantic = set(report["completed_semantic_chapters"])
+    summaries = set(report["completed_summary_chapters"])
+    all_chapters = set(range(1, int(expected) + 1))
+    invalid_raw, invalid_reuse = invalid_batch_targets(root, index_rows, state)
+
+    raw_targets = (all_chapters - semantic) | invalid_raw
+    if args.intent == "enhance":
+        reuse_targets = (semantic - raw_targets) | invalid_reuse
+    else:
+        reuse_targets = ((semantic - summaries) | invalid_reuse) - raw_targets
+    raw_targets -= cache_covered
+    reuse_targets -= cache_covered
+    if raw_targets and index_rows is None:
+        raise RunError("chapter_index_required", "raw-original work remains")
+    if index_rows is not None:
+        absent = raw_targets - {row["chapter"] for row in index_rows}
+        if absent:
+            raise RunError("range_not_in_index", ",".join(map(str, sorted(absent))))
+    mapping_blocked = set(report.get("chapter_mapping_blocked_chapters", []))
+    if mapping_blocked & (raw_targets | reuse_targets):
+        raise RunError(
+            "chapter_mapping_ambiguous",
+            "; ".join(report.get("chapter_mapping_conflicts", [])) or "legacy chapter identity is unresolved",
+            report.get("chapter_mapping_author_message"),
+        )
+
+    selected = []  # type: List[Dict[str, Any]]
+    index_by_chapter = {row["chapter"]: row for row in index_rows or []}
+    preferred_paths = report["chapter_sources"]["preferred_paths"]
+    for kind, targets in (("raw-original", raw_targets), ("existing-results", reuse_targets)):
+        for range_start, range_end in compact_ranges(targets):
+            for start, end in chunk_range(range_start, range_end, index_by_chapter if kind == "raw-original" else None):
+                selected.append({"input_kind": kind, "start": start, "end": end})
+
+    # Persisted split children replace any recombined parent range on later plans.
+    split_children = [
+        row for row in state["batches"].values()
+        if row.get("parent") and row.get("status") != "superseded"
+    ]
+    for child in split_children:
+        for item in list(selected):
+            if item["input_kind"] == child["input_kind"] and item["start"] <= child["start"] and child["end"] <= item["end"]:
+                selected.remove(item)
+                if item["start"] < child["start"]:
+                    selected.append({"input_kind": item["input_kind"], "start": item["start"], "end": child["start"] - 1})
+                selected.append({"input_kind": child["input_kind"], "start": child["start"], "end": child["end"]})
+                if child["end"] < item["end"]:
+                    selected.append({"input_kind": item["input_kind"], "start": child["end"] + 1, "end": item["end"]})
+                break
+
+    batches = []
+    raw_reads = 0
+    result_reads = 0
+    for item in sorted(selected, key=lambda value: (value["start"], value["end"], value["input_kind"])):
+        prefix = "RAW" if item["input_kind"] == "raw-original" else "REUSE"
+        batch_id = "%s-%s-%s" % (prefix, item["start"], item["end"])
+        current_range_hash = range_sha256(index_rows, item["start"], item["end"]) if item["input_kind"] == "raw-original" else "existing-results"
+        prior = state["batches"].get(batch_id)
+        if prior and completed_batch(root, prior, index_rows):
+            continue
+        if item["input_kind"] == "raw-original":
+            sources = [index_by_chapter[chapter]["source_locator"] for chapter in range(item["start"], item["end"] + 1)]
+            raw_reads += len(sources)
+        else:
+            sources = sorted(set(preferred_paths.get(str(chapter), "") for chapter in range(item["start"], item["end"] + 1)) - {""})
+            result_reads += len(sources)
+        batches.append({
+            "batch_id": batch_id, "chapter_range": [item["start"], item["end"]],
+            "input_kind": item["input_kind"], "range_sha256": current_range_hash,
+            "source_files": sources, "cache": "_analysis_cache/批次-%s.md" % batch_id,
+            "chapter_chars": [index_by_chapter[chapter]["char_count"] for chapter in range(item["start"], item["end"] + 1)]
+            if item["input_kind"] == "raw-original" else [],
+        })
+    required_stages = list(report.get("stage_repairs", []))
+    if batches or recoverable_caches:
+        required_stages = sorted(set(required_stages + ["stage2"]))
+    summary_gaps = sorted(all_chapters - summaries)
+    payload = {
+        "ok": True, "root": str(root), "intent": args.intent,
+        "classification": report["classification"], "recommended_path": report["recommended_path"],
+        "mixed_sources": report["mixed_sources"], "batches": batches,
+        "remaining_batches": len(batches),
+        "recoverable_caches": recoverable_caches,
+        "summary_gaps": summary_gaps,
+        "read_counts": {"raw_chapters": raw_reads, "existing_result_files": result_reads},
+        "required_stages": required_stages,
+        "state_written": False,
+    }
+    if args.next is not None:
+        if args.next < 0:
+            raise RunError("invalid_next", str(args.next))
+        # Dispatch one batch at a time without re-reading the whole remaining plan;
+        # counts above still describe everything that is left.
+        payload["batches"] = batches[:args.next]
+        payload["summary_gaps"] = ["%s-%s" % pair if pair[0] != pair[1] else str(pair[0])
+                                   for pair in compact_ranges(summary_gaps)]
+    return payload
+
+
+def parse_batch_id(batch_id: str) -> Tuple[str, int, int]:
+    match = BATCH_ID_RE.fullmatch(batch_id)
+    if not match:
+        raise RunError("invalid_batch_id", batch_id)
+    start, end = int(match.group(2)), int(match.group(3))
+    if start < 1 or end < start:
+        raise RunError("invalid_batch_range", batch_id)
+    return ("raw-original" if match.group(1) == "RAW" else "existing-results", start, end)
+
+
+def compact_field(body: str, name: str) -> str:
+    match = re.search(r"(?m)^\*\*%s\*\*\s*[：:]\s*(\S.*)$" % re.escape(name), body)
+    if not match:
+        raise RunError("chapter_schema_incomplete", "missing field: %s" % name)
+    value = match.group(1).strip()
+    if "{" in value or "}" in value:
+        raise RunError("template_placeholder", name)
+    return value
+
+
+def map_enum(value: str, allowed: Sequence[str], aliases: Dict[str, str], default: str = "其他") -> str:
+    """Map one enum value; with several listed values, the first that maps wins."""
+    for token in [value.strip()] + re.split(r"[、/，,；;\s]+", value.strip()):
+        if token in allowed:
+            return token
+        if token in aliases:
+            return aliases[token]
+        for item in allowed:
+            if item != "其他" and token.startswith(item):
+                return item
+    return default
+
+
+def parse_plot_points(body: str, chapter: int, minimum: int) -> List[Dict[str, Any]]:
+    """Parse the repeated ``P{n}`` blocks that follow ``**情节点**：``."""
+    match = re.search(r"(?m)^\*\*情节点\*\*\s*[：:]\s*$", body)
+    if not match:
+        raise RunError("chapter_schema_incomplete", "chapter %s missing field: 情节点" % chapter)
+    points = []  # type: List[Dict[str, Any]]
+    for line in body[match.end():].split("\n"):
+        stripped = line.strip()
+        header = POINT_HEADER_RE.match(stripped)
+        if header:
+            segments = [item.strip() for item in re.split(r"[|｜]", header.group(3))]
+            # Tolerate the tag written inline at the end of the P line.
+            inline = POINT_TAG_RE.match(" | ".join(segments[-2:])) if len(segments) >= 3 else None
+            rest = segments[1:-2] if inline else segments[1:]
+            if not segments[0].startswith("类型") or not rest or not rest[0]:
+                raise RunError("plot_point_invalid", "chapter %s P%s needs 类型 and 白描" % (chapter, header.group(1)))
+            points.append({"number": int(header.group(1)), "title": header.group(2).strip(),
+                           "type": segments[0][2:].lstrip(" ：:"), "rest": rest, "quote": [],
+                           "tag": (inline.group(1), inline.group(2)) if inline else None})
+            continue
+        if not stripped or stripped == "---":
+            continue
+        if not points or points[-1]["tag"] is not None:
+            raise RunError("plot_point_invalid", "chapter %s: unexpected line %r" % (chapter, stripped[:40]))
+        tag = POINT_TAG_RE.match(stripped)
+        if tag:
+            points[-1]["tag"] = (tag.group(1), tag.group(2))
+        else:
+            points[-1]["quote"].append(stripped)
+    if [point["number"] for point in points] != list(range(1, len(points) + 1)):
+        raise RunError("plot_point_invalid", "chapter %s: points must be numbered P1..Pn" % chapter)
+    if not minimum <= len(points) <= MAX_PLOT_POINTS:
+        raise RunError("plot_point_count", "chapter %s has %s points; expected %s-%s"
+                       % (chapter, len(points), minimum, MAX_PLOT_POINTS))
+    for point in points:
+        if point["tag"] is None:
+            raise RunError("plot_point_invalid", "chapter %s P%s missing 主题标签/基调 line" % (chapter, point["number"]))
+        text = "%s %s %s" % (point["title"], " ".join(point["rest"]), " ".join(point["quote"]))
+        if "{" in text or "}" in text:
+            raise RunError("template_placeholder", "chapter %s P%s" % (chapter, point["number"]))
+    return points
+
+
+def parse_model_output(text: str, start: int, end: int, input_kind: str) -> Dict[int, Dict[str, Any]]:
+    text = normalized(text)
+    if "BATCH_ERROR:" in text:
+        raise RunError("extractor_reported_error", "model returned BATCH_ERROR")
+    if any(marker in text for marker in (CACHE_START, CACHE_END, MODEL_START, MODEL_END)):
+        raise RunError("reserved_marker_in_output", "model output contains a runtime cache marker")
+    expected_tokens = [token for chapter in range(start, end + 1) for token in (("START", chapter), ("END", chapter))]
+    actual_tokens = [(match.group(1), int(match.group(2))) for match in CHAPTER_TOKEN_RE.finditer(text)]
+    if input_kind == "raw-original" and actual_tokens != expected_tokens:
+        raise RunError("chapter_marker_mismatch", "expected %s; received %s" % (expected_tokens, actual_tokens))
+    if actual_tokens and actual_tokens != expected_tokens:
+        raise RunError("chapter_marker_mismatch", "expected %s; received %s" % (expected_tokens, actual_tokens))
+    # Projections rebuilt from old results may hold fewer beats than a fresh reading.
+    minimum = MIN_PLOT_POINTS if input_kind == "raw-original" else 1
+    records = {}  # type: Dict[int, Dict[str, Any]]
+    for match in CHAPTER_BLOCK_RE.finditer(text):
+        chapter = int(match.group(1))
+        body = match.group(2).strip()
+        if not re.search(r"(?m)^##\s+第%s章(?:\s+.*)?$" % chapter, body):
+            raise RunError("chapter_schema_incomplete", "chapter %s heading missing" % chapter)
+        fields = {name: compact_field(body, name) for name in COMPACT_FIELDS}  # type: Dict[str, Any]
+        fields["情节点"] = parse_plot_points(body, chapter, minimum)
+        records[chapter] = fields
+    if input_kind == "raw-original" and set(records) != set(range(start, end + 1)):
+        raise RunError("chapter_block_missing", "%s-%s" % (start, end))
+    if records and set(records) != set(range(start, end + 1)):
+        raise RunError("chapter_block_missing", "partial reuse projection is not allowed")
+    if not records and input_kind == "existing-results":
+        marker = re.search(r"<!--\s*REUSED_CHAPTERS:(\d+)-(\d+)\s*-->", text)
+        if not marker or (int(marker.group(1)), int(marker.group(2))) != (start, end):
+            raise RunError("reused_range_mismatch", "%s-%s" % (start, end))
+    if text.count("<!-- BATCH_OBSERVATIONS_START -->") != 1 or text.count("<!-- BATCH_OBSERVATIONS_END -->") != 1:
+        raise RunError("batch_marker_mismatch", "one complete cross-chapter observation block is required")
+    observation_start = text.index("<!-- BATCH_OBSERVATIONS_START -->")
+    observation_end = text.index("<!-- BATCH_OBSERVATIONS_END -->")
+    last_source_marker = max((match.end() for match in CHAPTER_TOKEN_RE.finditer(text)), default=0)
+    reused_marker = re.search(r"<!--\s*REUSED_CHAPTERS:\d+-\d+\s*-->", text)
+    if reused_marker:
+        last_source_marker = max(last_source_marker, reused_marker.end())
+    if observation_start < last_source_marker or observation_end <= observation_start:
+        raise RunError("batch_marker_order", "cross-chapter observations must follow source coverage")
+    return records
+
+
+def render_plot_point(point: Dict[str, Any]) -> str:
+    point_type = map_enum(point["type"], POINT_TYPES, POINT_ALIASES, default="行动")
+    theme = map_enum(point["tag"][0], THEMES, THEME_ALIASES)
+    tone = map_enum(point["tag"][1], TONES, TONE_ALIASES)
+    lines = ["P%s **%s**：%s" % (point["number"], point["title"], " | ".join(["类型" + point_type] + point["rest"]))]
+    lines.extend(point["quote"])
+    lines.extend(["", "主题标签%s | 基调：%s" % (theme, tone)])
+    return "\n".join(lines)
+
+
+def render_summary(chapter: int, fields: Dict[str, Any], source_kind: str,
+                   chapter_hash: str, batch_id: str) -> bytes:
+    text = (
+        "<!-- story-long-analyze:projection runtime=single-state-v1 source=%s chapter_sha256=%s batch=%s -->\n"
+        "## 第%s章\n\n**概要**：%s\n\n**关键事件**：\n1. %s\n\n"
+        "**因果**：%s\n\n**局面结果**：%s\n\n**涉及**：%s\n\n"
+        "**信息变化**：%s\n\n**状态变化**：%s\n\n**三维节奏**：%s\n\n"
+        "**章尾钩子**：%s\n\n**证据**：%s\n\n**情节点**：\n\n%s\n"
+    ) % (
+        source_kind, chapter_hash, batch_id, chapter, fields["概要"], fields["关键行动"],
+        fields["因果"], fields["局面结果"], fields["涉及人物"], fields["信息变化"],
+        fields["状态变化"], fields["三维节奏"], fields["章尾钩子"], fields["证据"],
+        "\n\n---\n\n".join(render_plot_point(point) for point in fields["情节点"]),
+    )
+    return text.encode("utf-8")
+
+
+def render_cache(batch_id: str, start: int, end: int, input_kind: str,
+                 range_hash: str, source_files: Sequence[str], model_output: str,
+                 projection_schema: str = "compact-v3") -> bytes:
+    text = (
+        "%s\n# 批次 %s\n- batch_id: %s\n- chapters: %s-%s\n- input_kind: %s\n"
+        "- range_sha256: %s\n- projection_schema: %s\n"
+        "- source_files: %s\n%s\n%s\n%s\n%s\n"
+    ) % (CACHE_START, batch_id, batch_id, start, end, input_kind, range_hash, projection_schema,
+           json.dumps(list(source_files), ensure_ascii=False), MODEL_START,
+           normalized(model_output).strip(), MODEL_END, CACHE_END)
+    return text.encode("utf-8")
+
+
+def parse_cache(path: Path) -> Dict[str, Any]:
+    text = normalized(path.read_text(encoding="utf-8-sig"))
+    if not text.rstrip().endswith(CACHE_END):
+        raise RunError("cache_incomplete", str(path))
+    metadata = {}  # type: Dict[str, Any]
+    for key in ("batch_id", "chapters", "input_kind", "range_sha256", "projection_schema"):
+        match = re.search(r"(?m)^- %s:\s*(.+)$" % key, text)
+        if not match:
+            raise RunError("cache_invalid", "missing %s" % key)
+        metadata[key] = match.group(1).strip()
+    try:
+        start, end = [int(value) for value in metadata["chapters"].split("-", 1)]
+    except ValueError:
+        raise RunError("cache_invalid", "chapters: %s" % metadata["chapters"])
+    model_match = re.search(re.escape(MODEL_START) + r"\n(.*?)\n" + re.escape(MODEL_END), text, re.DOTALL)
+    if not model_match:
+        raise RunError("cache_invalid", "model output markers missing")
+    metadata.update({"start": start, "end": end, "model_output": model_match.group(1)})
+    return metadata
+
+
+def source_files_from_args(values: Optional[Sequence[str]]) -> List[str]:
+    return list(values or [])
+
+
+def commit_from_cache(root: Path, metadata: Dict[str, Any], cache: Path,
+                      index_rows: Optional[Sequence[Dict[str, Any]]]) -> Dict[str, Any]:
+    batch_id = metadata["batch_id"]
+    input_kind, id_start, id_end = parse_batch_id(batch_id)
+    start, end = metadata["start"], metadata["end"]
+    if (start, end, input_kind) != (id_start, id_end, metadata["input_kind"]):
+        raise RunError("cache_invalid", "batch metadata mismatch")
+    if input_kind == "raw-original":
+        if index_rows is None:
+            raise RunError("chapter_index_required", batch_id)
+        current_hash = range_sha256(index_rows, start, end)
+        if metadata["range_sha256"] != current_hash:
+            raise RunError("range_hash_mismatch", batch_id)
+    else:
+        current_hash = metadata["range_sha256"]
+    records = parse_model_output(metadata["model_output"], start, end, input_kind)
+    hashes = {row["chapter"]: row["chapter_sha256"] for row in index_rows or []}
+    created = []
+    kept = []
+    for chapter, fields in sorted(records.items()):
+        path = summary_path(root, chapter)
+        if path.exists():
+            kept.append(chapter)
+            continue
+        chapter_hash = hashes.get(chapter, "0" * 64)
+        atomic_write(path, render_summary(chapter, fields, input_kind, chapter_hash, batch_id))
+        created.append(path.relative_to(root).as_posix())
+        failure_after = os.environ.get("STORY_ANALYZE_FAIL_AFTER_SUMMARIES")
+        if failure_after and len(created) >= int(failure_after):
+            raise OSError("injected_failure_after_%s_summaries" % failure_after)
+    missing = [chapter for chapter in range(start, end + 1) if not summary_path(root, chapter).is_file()]
+    if missing:
+        raise RunError("summary_projection_missing", ",".join(map(str, missing)))
+    state, _ = load_state(root / "_progress.md")
+    state["batches"][batch_id] = {
+        "batch_id": batch_id, "start": start, "end": end, "input_kind": input_kind,
+        "range_sha256": current_hash, "status": "completed", "parent": "",
+        "cache": cache.relative_to(root).as_posix(),
+    }
+    changed = write_state(root / "_progress.md", state)
+    return {"batch_id": batch_id, "created_summaries": created,
+            "kept_existing_summary_chapters": kept, "progress_updated": changed}
+
+
+def commit_command(args: argparse.Namespace) -> Dict[str, Any]:
+    root = require_root(args.root)
+    input_kind, start, end = parse_batch_id(args.batch_id)
+    if end - start + 1 > MAX_CHAPTERS:
+        raise RunError("batch_too_large", "%s exceeds %s chapters" % (args.batch_id, MAX_CHAPTERS))
+    text = args.input.read_text(encoding="utf-8-sig")
+    records = parse_model_output(text, start, end, input_kind)
+    index_rows = read_index(root, args.index) if input_kind == "raw-original" or (args.index or root / "chapter_index.csv").is_file() else None
+    if input_kind == "raw-original" and end > start:
+        total_chars = sum(row["char_count"] for row in index_rows if start <= row["chapter"] <= end)
+        if total_chars > MAX_CHARS:
+            raise RunError("batch_too_large", "%s exceeds %s characters" % (args.batch_id, MAX_CHARS))
+    current_hash = range_sha256(index_rows, start, end) if input_kind == "raw-original" else "existing-results"
+    if input_kind == "raw-original" and not args.range_sha256:
+        raise RunError("range_hash_required", "pass the value printed by plan")
+    if args.range_sha256 and args.range_sha256 != current_hash:
+        raise RunError("range_hash_mismatch", args.batch_id)
+    # Parsing above validates the whole result before the first write.
+    path = cache_path(root, args.batch_id)
+    data = render_cache(args.batch_id, start, end, input_kind, current_hash,
+                        source_files_from_args(args.source_file), text)
+    if not path.is_file() or path.read_bytes() != data:
+        if path.is_file():
+            old_data = path.read_bytes()
+            history = root / "_analysis_cache" / "legacy" / (
+                "%s.%s.md" % (path.stem, sha256(old_data)[:12])
+            )
+            if not history.exists():
+                atomic_write(history, old_data)
+        atomic_write(path, data)
+    if os.environ.get("STORY_ANALYZE_FAIL_AFTER_CACHE") == "1":
+        raise OSError("injected_failure_after_cache")
+    metadata = parse_cache(path)
+    result = commit_from_cache(root, metadata, path, index_rows)
+    result.update({"ok": True, "cache": path.relative_to(root).as_posix(), "validated_chapters": sorted(records)})
+    return result
+
+
+def repair_command(args: argparse.Namespace) -> Dict[str, Any]:
+    root = require_root(args.root)
+    index_rows = read_index(root, args.index) if (args.index or root / "chapter_index.csv").is_file() else None
+    paths = [cache_path(root, args.batch_id)] if args.batch_id else sorted((root / "_analysis_cache").glob("批次-*.md"))
+    state, _ = load_state(root / "_progress.md")
+    repaired = []
+    skipped = []
+    errors = []
+    for path in paths:
+        # A cache replaced by a split or by other committed batches is history, not a repair target.
+        batch_id = path.stem[len("批次-"):]
+        if BATCH_ID_RE.fullmatch(batch_id):
+            _, start, end = parse_batch_id(batch_id)
+            covered = set()
+            for other in state["batches"].values():
+                if other["batch_id"] != batch_id and completed_batch(root, other, index_rows):
+                    covered.update(range(other["start"], other["end"] + 1))
+            if state["batches"].get(batch_id, {}).get("status") == "superseded" or set(range(start, end + 1)) <= covered:
+                skipped.append({"cache": path.relative_to(root).as_posix(), "reason": "superseded"})
+                continue
+        try:
+            metadata = parse_cache(path)
+            repaired.append(commit_from_cache(root, metadata, path, index_rows))
+        except (OSError, UnicodeError, RunError) as exc:
+            errors.append({"cache": path.as_posix(), "error": getattr(exc, "code", str(exc)), "detail": str(exc)})
+    return {"ok": not errors, "repaired": repaired, "skipped": skipped, "errors": errors}
+
+
+def split_command(args: argparse.Namespace) -> Dict[str, Any]:
+    root = require_root(args.root)
+    input_kind, start, end = parse_batch_id(args.batch_id)
+    if start >= end:
+        raise RunError("batch_not_splittable", args.batch_id)
+    index_rows = read_index(root, args.index) if input_kind == "raw-original" else None
+    split_at = args.at
+    if split_at is None:
+        if index_rows:
+            counts = {row["chapter"]: row["char_count"] for row in index_rows}
+            total = sum(counts[chapter] for chapter in range(start, end + 1))
+            running = 0
+            split_at = start
+            for chapter in range(start, end):
+                running += counts[chapter]
+                split_at = chapter
+                if running >= total / 2:
+                    break
+        else:
+            split_at = (start + end) // 2
+    if split_at < start or split_at >= end:
+        raise RunError("invalid_split_point", str(split_at))
+    state, _ = load_state(root / "_progress.md")
+    parent_hash = range_sha256(index_rows, start, end) if index_rows else "existing-results"
+    state["batches"][args.batch_id] = {
+        "batch_id": args.batch_id, "start": start, "end": end, "input_kind": input_kind,
+        "range_sha256": parent_hash, "status": "superseded", "parent": "", "cache": "",
+    }
+    prefix = "RAW" if input_kind == "raw-original" else "REUSE"
+    children = []
+    for child_start, child_end in ((start, split_at), (split_at + 1, end)):
+        child_id = "%s-%s-%s" % (prefix, child_start, child_end)
+        child_hash = range_sha256(index_rows, child_start, child_end) if index_rows else "existing-results"
+        existing = state["batches"].get(child_id)
+        if not existing or existing.get("status") not in {"completed", "success"}:
+            state["batches"][child_id] = {
+                "batch_id": child_id, "start": child_start, "end": child_end,
+                "input_kind": input_kind, "range_sha256": child_hash, "status": "planned",
+                "parent": args.batch_id, "cache": "",
+            }
+        children.append(child_id)
+    write_state(root / "_progress.md", state)
+    return {"ok": True, "parent": args.batch_id, "status": "superseded", "children": children}
+
+
+def stage_required_outputs(root: Path, stage: str) -> List[str]:
+    if stage == "stage1":
+        expected = inspect(root, None).get("expected_chapters") or 3
+        required = ["章节/第%s章_深度拆解.md" % chapter for chapter in range(1, min(3, int(expected)) + 1)]
+        required.append("快速预览.md")
+        return required
+    if stage == "stage2":
+        expected = inspect(root, None).get("expected_chapters")
+        if not expected:
+            raise RunError("expected_chapters_unknown", "Stage 2 completion requires a known chapter count")
+        return ["章节/第%s章_摘要.md" % chapter for chapter in range(1, int(expected) + 1)]
+    if stage == "stage3":
+        return ["剧情/情绪模块.md", "剧情/节奏.md"]
+    if stage == "stage4":
+        character_files = sorted(path for path in (root / "角色").glob("*.md") if path.is_file() and path.stat().st_size)
+        setting_files = sorted(path for path in (root / "设定").rglob("*.md") if path.is_file() and path.stat().st_size)
+        missing = []
+        if not character_files:
+            missing.append("角色/*.md")
+        if not setting_files:
+            missing.append("设定/**/*.md")
+        if missing:
+            raise RunError("stage_output_missing", ",".join(missing))
+        return [character_files[0].relative_to(root).as_posix(), setting_files[0].relative_to(root).as_posix()]
+    if stage == "stage5":
+        return ["拆文报告.md"]
+    if stage == "stage6":
+        return ["文风.md"]
+    raise RunError("stage_unknown", stage)
+
+
+def mark_stage_command(args: argparse.Namespace) -> Dict[str, Any]:
+    root = require_root(args.root)
+    if args.prepare:
+        if args.stage != "stage5":
+            raise RunError("prepare_stage_unsupported", args.stage)
+        report = root / "拆文报告.md"
+        backup = root / "_analysis_cache" / "legacy" / "拆文报告.md"
+        created = False
+        selected_backup = backup
+        if report.is_file():
+            report_data = report.read_bytes()
+            if backup.exists() and backup.read_bytes() != report_data:
+                selected_backup = backup.with_name("拆文报告.%s.md" % sha256(report_data)[:12])
+            if not selected_backup.exists():
+                atomic_write(selected_backup, report_data)
+                created = True
+        return {"ok": True, "stage": args.stage, "prepared": True,
+                "legacy_report_backup": selected_backup.relative_to(root).as_posix() if selected_backup.is_file() else None,
+                "backup_created": created, "progress_updated": False}
+    required_outputs = stage_required_outputs(root, args.stage) if args.status == "completed" else []
+    missing_required = [
+        name for name in required_outputs
+        if not (root / name).is_file() or (root / name).stat().st_size == 0
+    ]
+    if missing_required:
+        raise RunError("stage_output_missing", ",".join(missing_required))
+    output = args.output
+    relative_output = ""
+    if output:
+        path = output if output.is_absolute() else root / output
+        if args.status == "completed" and (not path.is_file() or path.stat().st_size == 0):
+            raise RunError("stage_output_missing", str(path))
+        try:
+            relative_output = path.resolve().relative_to(root).as_posix()
+        except ValueError:
+            raise RunError("stage_output_outside_root", str(path))
+    elif required_outputs:
+        relative_output = ";".join(required_outputs)
+    state, _ = load_state(root / "_progress.md")
+    state["stages"][args.stage] = {"status": args.status, "output": relative_output}
+    changed = write_state(root / "_progress.md", state)
+    return {"ok": True, "stage": args.stage, "status": args.status, "progress_updated": changed}
+
+
+OBS_START = "<!-- BATCH_OBSERVATIONS_START -->"
+OBS_END = "<!-- BATCH_OBSERVATIONS_END -->"
+# Labels as projected by render_summary (legacy summaries use the same labels).
+SUMMARY_FIELDS = ("概要", "关键事件", "因果", "局面结果", "涉及", "信息变化", "状态变化",
+                  "三维节奏", "章尾钩子", "证据")
+
+
+def parse_chapter_window(value: Optional[str]) -> Optional[Tuple[int, int]]:
+    if not value:
+        return None
+    match = re.fullmatch(r"(\d+)(?:-(\d+))?", value.strip())
+    if not match:
+        raise RunError("invalid_chapter_window", value)
+    start = int(match.group(1))
+    end = int(match.group(2) or start)
+    if start < 1 or end < start:
+        raise RunError("invalid_chapter_window", value)
+    return start, end
+
+
+def digest_caches(root: Path) -> List[Tuple[int, int, str, Path]]:
+    """Complete caches of committed batches, falling back to every complete cache on disk."""
+    state, _ = load_state(root / "_progress.md")
+    chosen = {}  # type: Dict[str, Tuple[int, int, str, Path]]
+    for row in state["batches"].values():
+        if row.get("status") not in {"completed", "success"}:
+            continue
+        path = root / row["cache"] if row.get("cache") else cache_path(root, row["batch_id"])
+        if cache_complete(path):
+            chosen[row["batch_id"]] = (row["start"], row["end"], row["batch_id"], path)
+    if not chosen:
+        for path in sorted((root / "_analysis_cache").glob("批次-*.md")):
+            try:
+                _, start, end = parse_batch_id(path.stem[len("批次-"):])
+            except RunError:
+                continue
+            if cache_complete(path):
+                chosen[path.stem] = (start, end, path.stem[len("批次-"):], path)
+    return sorted(chosen.values())
+
+
+def summary_field(text: str, name: str) -> str:
+    if name == "关键事件":
+        match = re.search(r"(?ms)^\*\*关键事件\*\*\s*[：:]\s*\n(.*?)(?=^\*\*|\Z)", text)
+        return " ".join(line.strip() for line in match.group(1).splitlines() if line.strip()) if match else ""
+    match = re.search(r"(?m)^\*\*%s\*\*\s*[：:]\s*(\S.*)$" % re.escape(name), text)
+    return match.group(1).strip() if match else ""
+
+
+def summary_points(text: str, mode: str) -> List[str]:
+    lines = text.split("\n")
+    points = []  # type: List[str]
+    for index, line in enumerate(lines):
+        header = POINT_HEADER_RE.match(line.strip())
+        if not header:
+            continue
+        segments = [item.strip() for item in re.split(r"[|｜]", header.group(3))]
+        tone = ""
+        for follow in lines[index + 1:index + 8]:
+            tag = POINT_TAG_RE.match(follow.strip())
+            if tag:
+                tone = tag.group(2)
+                break
+            if POINT_HEADER_RE.match(follow.strip()):
+                break
+        point_type = segments[0][2:].lstrip(" ：:") if segments[0].startswith("类型") else segments[0]
+        if mode == "brief":
+            points.append("P%s %s｜%s｜%s" % (header.group(1), header.group(2).strip(), point_type, tone or "—"))
+        else:
+            points.append("%s ｜基调：%s" % (line.strip(), tone or "—"))
+    return points
+
+
+def digest_command(args: argparse.Namespace) -> Dict[str, Any]:
+    root = require_root(args.root)
+    window = parse_chapter_window(args.chapters)
+    parts = []  # type: List[str]
+    if args.part == "observations":
+        for start, end, batch_id, path in digest_caches(root):
+            if window and (end < window[0] or start > window[1]):
+                continue
+            model = parse_cache(path)["model_output"]
+            if OBS_START not in model or OBS_END not in model:
+                continue
+            body = model.split(OBS_START, 1)[1].split(OBS_END, 1)[0].strip()
+            body = re.sub(r"(?m)^##\s*跨章观察\s*\n", "", body).strip()
+            parts.append("## %s（第%s-%s章）\n\n%s" % (batch_id, start, end, body))
+    else:
+        fields = [item.strip() for item in (args.fields or "").split(",") if item.strip()]
+        unknown = [item for item in fields if item not in SUMMARY_FIELDS]
+        if unknown:
+            raise RunError("unknown_summary_field", ",".join(unknown))
+        if not fields and args.points == "none":
+            raise RunError("digest_empty_request", "pass --fields and/or --points")
+        chapters = sorted(
+            int(match.group(1)) for match in
+            (re.fullmatch(r"第(\d+)章_摘要", path.stem) for path in (root / "章节").glob("第*章_摘要.md"))
+            if match
+        )
+        for chapter in chapters:
+            if window and not window[0] <= chapter <= window[1]:
+                continue
+            text = normalized(summary_path(root, chapter).read_text(encoding="utf-8-sig"))
+            lines = ["### 第%s章" % chapter]
+            lines.extend("- %s：%s" % (name, summary_field(text, name) or "未提供") for name in fields)
+            if args.points != "none":
+                lines.extend(summary_points(text, args.points))
+            parts.append("\n".join(lines))
+    if not parts:
+        raise RunError("digest_nothing_found", "no committed batch caches or summaries in range")
+    return {"ok": True, "text": "\n\n".join(parts) + "\n"}
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description=__doc__)
+    sub = parser.add_subparsers(dest="command", required=True)
+    plan = sub.add_parser("plan", help="print a read-only in-memory plan")
+    plan.add_argument("--root", required=True, type=Path)
+    plan.add_argument("--index", type=Path)
+    plan.add_argument("--expected-chapters", type=int)
+    plan.add_argument("--intent", choices=("continue", "enhance"), default="continue")
+    plan.add_argument("--next", type=int, metavar="N",
+                      help="only print the first N batches (0 = counts only); remaining_batches stays total")
+    plan.set_defaults(handler=plan_command)
+    commit = sub.add_parser("commit", help="validate and atomically commit one batch")
+    commit.add_argument("--root", required=True, type=Path)
+    commit.add_argument("--input", required=True, type=Path)
+    commit.add_argument("--batch-id", required=True)
+    commit.add_argument("--range-sha256")
+    commit.add_argument("--index", type=Path)
+    commit.add_argument("--source-file", action="append")
+    commit.set_defaults(handler=commit_command)
+    split = sub.add_parser("split", help="persist a failed batch split")
+    split.add_argument("--root", required=True, type=Path)
+    split.add_argument("--batch-id", required=True)
+    split.add_argument("--at", type=int)
+    split.add_argument("--index", type=Path)
+    split.set_defaults(handler=split_command)
+    repair = sub.add_parser("repair-progress", help="recover missing projections/state from complete caches")
+    repair.add_argument("--root", required=True, type=Path)
+    repair.add_argument("--batch-id")
+    repair.add_argument("--index", type=Path)
+    repair.set_defaults(handler=repair_command)
+    stage = sub.add_parser("mark-stage", help="mark a stage after its output is present")
+    stage.add_argument("--root", required=True, type=Path)
+    stage.add_argument("--stage", required=True, choices=("stage1", "stage2", "stage3", "stage4", "stage5", "stage6"))
+    stage.add_argument("--status", choices=("completed", "completed_with_errors"), default="completed")
+    stage.add_argument("--output", type=Path)
+    stage.add_argument("--prepare", action="store_true", help="before Stage 5, preserve the existing report")
+    stage.set_defaults(handler=mark_stage_command)
+    digest = sub.add_parser("digest", help="print Stage 3-5 reading material without whole-cache reads")
+    digest.add_argument("--root", required=True, type=Path)
+    digest.add_argument("--part", required=True, choices=("observations", "chapters"))
+    digest.add_argument("--chapters", help="chapter window such as 1-40")
+    digest.add_argument("--fields", help="comma-separated summary fields, e.g. 三维节奏,涉及,状态变化")
+    digest.add_argument("--points", choices=("none", "brief", "full"), default="none")
+    digest.set_defaults(handler=digest_command)
+    return parser
+
+
+def main() -> int:
+    for stream in (sys.stdout, sys.stderr):
+        if hasattr(stream, "reconfigure"):
+            stream.reconfigure(encoding="utf-8")
+    parser = build_parser()
+    args = parser.parse_args()
+    try:
+        payload = args.handler(args)
+        if args.command == "digest":
+            sys.stdout.write(payload["text"])
+            return 0
+        print(json.dumps(payload, ensure_ascii=False, indent=2, sort_keys=True))
+        return 0 if payload.get("ok", True) else 2
+    except (OSError, UnicodeError, RunError) as exc:
+        payload = {"ok": False, "error": getattr(exc, "code", "io_error"), "detail": str(exc)}
+        if getattr(exc, "author_message", None):
+            payload["author_message"] = exc.author_message
+        print(json.dumps(payload, ensure_ascii=False, indent=2))
+        return 2
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())