-----/tools/epub_review_editor/server.py

from __future__ import annotations

import argparse
import copy
import datetime as dt
import hashlib
import html
import json
import os
import posixpath
import re
import shutil
import signal
import socket
import subprocess
import sys
import threading
import time
import urllib.error
import urllib.parse
import urllib.request
import webbrowser
import zipfile
from html.parser import HTMLParser
from pathlib import Path
from typing import Any
import xml.etree.ElementTree as ET

from flask import Flask, jsonify, make_response, request, send_from_directory

try:
    from version import VERSION_RULES, version
except ImportError:
    VERSION_RULES = {
        "PATCH": "修复 bug 或兼容性小修",
        "MINOR": "新增向后兼容能力、API 字段或可选配置",
        "MAJOR": "破坏兼容的 API、配置或行为变更",
    }
    version = "0.12.2"


P_RE = re.compile(r"(<p\b[^>]*>)(.*?)(</p>)", re.S | re.I)
TAG_RE = re.compile(r"<[^>]+>")
IMG_TAG_RE = re.compile(r"<(?:img|image)\b[^>]*>", re.S | re.I)
ATTR_RE = re.compile(r"([:\w-]+)\s*=\s*(['\"])(.*?)\2", re.S)
JP_RE = re.compile(r"[\u3040-\u30ff\u3400-\u9fff々〆〤]")
KANA_RE = re.compile(r"[\u3040-\u30ff]")
GRAY_RE = re.compile(
    r"(?:color\s*:\s*(?:#?808080|gray|grey|rgb\(\s*128\s*,\s*128\s*,\s*128\s*\))|class\s*=\s*['\"][^'\"]*(?:gray|grey|jp|source)[^'\"]*['\"])",
    re.I,
)
EXTERNAL_URI_RE = re.compile(r"^(?:[a-z][a-z0-9+.-]*:|//)", re.I)
IMAGE_EXTENSIONS = {".apng", ".avif", ".bmp", ".gif", ".jpeg", ".jpg", ".png", ".svg", ".webp"}
RASTER_IMAGE_EXTENSIONS = {".apng", ".avif", ".bmp", ".gif", ".jpeg", ".jpg", ".png", ".webp"}

DEFAULT_REVIEW_ROOT = "epub_review_sessions"
STATE_VERSION = 1
UPLOAD_DIR_NAME = "uploads"
DEFAULT_GPT_BASE_URL = "https://api.openai.com/v1"
DEFAULT_GPT_MODEL = "gpt-4o-mini"
DEFAULT_TRANSLATION_PROMPT = (
    "你是日语轻小说到简体中文的审校重翻助手。只重翻当前目标语句/段落。"
    "目标是自然、流畅、忠实的简体中文轻小说文风。意义优先，再调整为中文语序；"
    "不要贴着日语句序硬译，也不要补写原文没有的信息。"
    "术语、人名、地名、TRPG/系统词只在本段命中时服从给定术语；没有命中的术语不要写进译文。"
)
LEGACY_TRANSLATION_PROMPT_WITHOUT_GLOBAL_TERM = (
    "你是日语轻小说到简体中文的审校重翻助手。目标是把当前目标语句译成自然、流畅、忠实的简体中文轻小说文风。"
    "意义优先，再调整为中文语序；不要贴着日语句序硬译，也不要补写原文没有的信息。"
    "术语、人名、地名、TRPG/系统词优先服从给定名词表。"
)
DEFAULT_FORMAT_PROMPT = (
    "只输出目标语句的新中文译文 HTML 片段，不要解释、不要 Markdown、不要代码围栏。"
    "不得合并或拆分段落。保留必要的 HTML 内联标签，尤其是有意义的 ruby，必须写成"
    "<ruby><rb>中文</rb><rt>读音/外语</rt></ruby>，不要改成括号。"
    "标点统一为简中出版格式：对话引号用「」；省略号用……；破折号用——；问号用？，感叹号用！，问叹连用用？！或！？；"
    "逗号、句号、顿号、冒号、分号用，。、：；。不要用英文半角标点替代中文标点。"
)
DEFAULT_CHARACTER_PROMPT = (
    "根据目标语句和少量前后文判断说话人、情绪和角色口吻。"
    "本系列以第一人称轻小说叙述为主，主角聪明、爱吐槽、有TRPG玩家思维；对话要符合身份差异，"
    "不要把所有角色写成同一种口吻，也不要过度网络化。"
)
GPT_PROMPT_DEFAULTS = {
    "translation_prompt": DEFAULT_TRANSLATION_PROMPT,
    "format_prompt": DEFAULT_FORMAT_PROMPT,
    "character_prompt": DEFAULT_CHARACTER_PROMPT,
}
GPT_PROMPT_MAX_CHARS = {
    "translation_prompt": 5000,
    "format_prompt": 5000,
    "character_prompt": 3000,
}
GLOSSARY_ENTRY_LIMIT = 80
GLOSSARY_PROMPT_LIMIT = 16
TRANSLATION_JOB_DIR_NAME = "translation_jobs"
TRANSLATION_OUTPUT_DIR_NAME = "translated_outputs"
TRANSLATION_LOG_LIMIT = 80
DEFAULT_TRANSLATION_LIMIT = 20
TRANSLATION_RANGE_LIMIT_MAX = 5000
TRANSLATION_CONTEXT_RADIUS = 1


def now_iso() -> str:
    return dt.datetime.now(dt.timezone.utc).astimezone().isoformat(timespec="seconds")


def safe_slug(value: str, max_len: int = 80) -> str:
    value = re.sub(r'[\\/:*?"<>|]+', "_", value).strip().strip(".")
    value = re.sub(r"\s+", "_", value)
    return value[:max_len] or "epub"


def find_free_port(start: int, host: str = "127.0.0.1") -> int:
    connect_host = "127.0.0.1" if host in {"localhost", "0.0.0.0", "::"} else host
    bind_host = "127.0.0.1" if host in {"localhost", "::"} else host
    port_in_use = getattr(socket, "SO_EXCLUSIVEADDRUSE", None)
    for port in range(start, start + 100):
        try:
            with socket.create_connection((connect_host, port), timeout=0.2):
                continue
        except OSError:
            pass
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
            if port_in_use is not None:
                sock.setsockopt(socket.SOL_SOCKET, port_in_use, 1)
            try:
                sock.bind((bind_host, port))
            except OSError:
                continue
            return port
    raise RuntimeError(f"no free port found from {start}")


def display_host(host: str) -> str:
    return "localhost" if host in {"127.0.0.1", "0.0.0.0", "::"} else host


def read_json(path: Path, default: Any) -> Any:
    if not path.exists():
        return copy.deepcopy(default)
    return json.loads(path.read_text(encoding="utf-8-sig"))


def write_json(path: Path, data: Any) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")


def write_json_atomic(path: Path, data: Any) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    tmp = path.with_name(f"{path.name}.tmp")
    tmp.write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    tmp.replace(path)


def append_jsonl(path: Path, record: dict[str, Any]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("a", encoding="utf-8", newline="\n") as fh:
        fh.write(json.dumps(record, ensure_ascii=False) + "\n")


def strip_tags(fragment: str) -> str:
    fragment = re.sub(r"<rt\b[^>]*>.*?</rt>", "", fragment, flags=re.S | re.I)
    return html.unescape(TAG_RE.sub("", fragment)).strip()


def strip_html_keep_ruby_text(fragment: str) -> str:
    return html.unescape(TAG_RE.sub("", fragment or "")).strip()


def compact_text(value: str) -> str:
    return re.sub(r"\s+", " ", html.unescape(value or "")).strip()


def file_sort_key(value: str) -> list[Any]:
    parts = re.split(r"(\d+)", value.lower())
    return [int(part) if part.isdigit() else part for part in parts]


def entry_dir(entry_name: str) -> str:
    directory = posixpath.dirname(entry_name)
    return f"{directory}/" if directory else ""


def normalize_href(base_dir: str, href: str) -> str:
    href = urllib.parse.unquote((href or "").split("#", 1)[0])
    if not href:
        return ""
    return posixpath.normpath(posixpath.join(base_dir, href)).lstrip("./")


def normalize_href_with_fragment(base_dir: str, href: str) -> tuple[str, str]:
    href = urllib.parse.unquote(href or "")
    path, fragment = (href.split("#", 1) + [""])[:2] if "#" in href else (href, "")
    path = posixpath.normpath(posixpath.join(base_dir, path)).lstrip("./") if path else ""
    return path, fragment


def read_xml_root(path: Path) -> ET.Element | None:
    if not path.exists():
        return None
    try:
        return ET.fromstring(path.read_text(encoding="utf-8", errors="replace"))
    except ET.ParseError:
        try:
            return ET.fromstring(path.read_bytes())
        except ET.ParseError:
            return None


def local_name(tag: str) -> str:
    return tag.rsplit("}", 1)[-1].lower()


def element_text(elem: ET.Element) -> str:
    return compact_text("".join(elem.itertext()))


def is_japaneseish_fragment(fragment: str) -> bool:
    return bool(KANA_RE.search(fragment or ""))


def is_japaneseish(fragment: str) -> bool:
    text = strip_tags(fragment)
    return bool(KANA_RE.search(text))


def is_gray_source(attrs: str, inner: str) -> bool:
    if GRAY_RE.search(attrs):
        return True
    return is_japaneseish(inner)


class HtmlSanitizer(HTMLParser):
    allowed_tags = {
        "a",
        "b",
        "br",
        "code",
        "em",
        "i",
        "mark",
        "rb",
        "rp",
        "rt",
        "ruby",
        "small",
        "span",
        "strong",
        "sub",
        "sup",
    }
    allowed_attrs = {
        "a": {"href", "title"},
        "mark": {"class", "data-review-mark"},
        "span": {"class", "style", "data-review-mark"},
        "ruby": {"class"},
        "rt": {"class"},
        "rb": {"class"},
        "rp": {"class"},
    }
    void_tags = {"br"}
    style_safe = re.compile(r"^(?:\s*(?:color|background-color|font-weight|font-style)\s*:\s*[-#(),.%\w\s]+\s*;?)*$", re.I)

    def __init__(self) -> None:
        super().__init__(convert_charrefs=False)
        self.parts: list[str] = []

    def handle_starttag(self, tag: str, attrs: list[tuple[str, str | None]]) -> None:
        tag = tag.lower()
        if tag not in self.allowed_tags:
            return
        safe_attrs = []
        for key, value in attrs:
            key = key.lower()
            value = value or ""
            if key not in self.allowed_attrs.get(tag, set()):
                continue
            if key.startswith("on"):
                continue
            if key == "href" and re.match(r"\s*(?:javascript|data):", value, re.I):
                continue
            if key == "style" and not self.style_safe.fullmatch(value):
                continue
            safe_attrs.append(f'{key}="{html.escape(value, quote=True)}"')
        attr_text = (" " + " ".join(safe_attrs)) if safe_attrs else ""
        if tag in self.void_tags:
            self.parts.append(f"<{tag}{attr_text}/>")
        else:
            self.parts.append(f"<{tag}{attr_text}>")

    def handle_endtag(self, tag: str) -> None:
        tag = tag.lower()
        if tag in self.allowed_tags and tag not in self.void_tags:
            self.parts.append(f"</{tag}>")

    def handle_data(self, data: str) -> None:
        self.parts.append(html.escape(data, quote=False))

    def handle_entityref(self, name: str) -> None:
        self.parts.append(f"&{name};")

    def handle_charref(self, name: str) -> None:
        self.parts.append(f"&#{name};")


def sanitize_cn_html(value: str) -> str:
    value = re.sub(r"(?is)<\s*(script|style|iframe|object|embed)\b.*?</\s*\1\s*>", "", value)
    sanitizer = HtmlSanitizer()
    sanitizer.feed(value)
    sanitizer.close()
    return "".join(sanitizer.parts).strip()


def normalize_gpt_base_url(value: str) -> str:
    value = (value or DEFAULT_GPT_BASE_URL).strip().rstrip("/")
    if not value:
        value = DEFAULT_GPT_BASE_URL
    parsed = urllib.parse.urlparse(value)
    if parsed.scheme not in {"http", "https"} or not parsed.netloc:
        raise ValueError("Base URL 必须是 http(s) 地址")
    return value


def normalize_gpt_prompt(name: str, value: Any) -> str:
    text = str(value or "").strip()
    if not text:
        text = GPT_PROMPT_DEFAULTS[name]
    if name == "translation_prompt":
        text = normalize_translation_prompt(text)
    return text[: GPT_PROMPT_MAX_CHARS[name]]


def gpt_prompt_defaults() -> dict[str, str]:
    return dict(GPT_PROMPT_DEFAULTS)


def compact_prompt_text(text: str) -> str:
    return re.sub(r"\s+", "", text or "")


def normalize_translation_prompt(text: str) -> str:
    text = strip_obsolete_global_term_rules(text)
    if not text:
        return DEFAULT_TRANSLATION_PROMPT
    if compact_prompt_text(text) == compact_prompt_text(LEGACY_TRANSLATION_PROMPT_WITHOUT_GLOBAL_TERM):
        return DEFAULT_TRANSLATION_PROMPT
    scoped_clause = "术语、人名、地名、TRPG/系统词只在本段命中时服从给定术语；没有命中的术语不要写进译文。"
    text = re.sub(
        r"术语、人名、地名、TRPG/系统词优先服从给定名词表[。；;]?",
        scoped_clause,
        text,
    )
    return text


def strip_obsolete_global_term_rules(text: str) -> str:
    """Old saved prompts hard-coded one term into every request; keep terms scoped."""
    parts = re.split(r"([。！？\n；;]+)", text)
    kept: list[str] = []
    for idx in range(0, len(parts), 2):
        segment = parts[idx]
        delimiter = parts[idx + 1] if idx + 1 < len(parts) else ""
        has_falchion = re.search(r"Falchion|ファルシオン", segment, flags=re.I)
        has_translation = "大砍刀" in segment or "偃月刀" in segment
        if has_falchion and has_translation:
            continue
        kept.append(segment + delimiter)
    text = "".join(kept)
    text = re.sub(r"[；;]\s*([。！？])", r"\1", text)
    text = re.sub(r"[；;]\s*$", "。", text)
    text = re.sub(r"。{2,}", "。", text)
    return text.strip()


def project_root() -> Path:
    return Path(__file__).resolve().parents[2]


def default_glossary_path() -> Path:
    root = project_root()
    candidates = sorted(
        path
        for path in root.glob("*/mingcibiao.json")
        if "epub_review_sessions" not in path.parts
    )
    if len(candidates) == 1:
        return candidates[0]
    return root / "mingcibiao.json"


def normalize_glossary_path_value(value: str = "") -> str:
    raw = str(value or "").strip()
    path = Path(raw).expanduser() if raw else default_glossary_path()
    if not path.is_absolute():
        path = project_root() / path
    if path.suffix.lower() != ".json":
        raise ValueError("术语表路径必须指向 .json 文件")
    return str(path.resolve())


def update_gpt_config_fields(review_root: Path, fields: dict[str, Any]) -> None:
    existing = read_json(gpt_config_path(review_root), {})
    if not isinstance(existing, dict):
        existing = {}
    existing.update(fields)
    existing["updated_at"] = now_iso()
    write_json(gpt_config_path(review_root), existing)


def read_gpt_config(review_root: Path) -> dict[str, Any]:
    data = read_json(gpt_config_path(review_root), {})
    api_key = str(data.get("api_key") or os.environ.get("OPENAI_API_KEY") or "").strip()
    base_url = str(data.get("base_url") or os.environ.get("OPENAI_BASE_URL") or DEFAULT_GPT_BASE_URL).strip()
    model = str(data.get("model") or os.environ.get("OPENAI_MODEL") or DEFAULT_GPT_MODEL).strip()
    prompts = {
        name: normalize_gpt_prompt(name, data.get(name, default))
        for name, default in GPT_PROMPT_DEFAULTS.items()
    }
    return {
        "base_url": normalize_gpt_base_url(base_url),
        "model": model or DEFAULT_GPT_MODEL,
        "api_key": api_key,
        "configured": bool(api_key),
        "key_source": "config" if str(data.get("api_key") or "").strip() else ("env" if os.environ.get("OPENAI_API_KEY") else ""),
        "updated_at": data.get("updated_at", ""),
        "glossary_path": normalize_glossary_path_value(str(data.get("glossary_path") or "")),
        **prompts,
    }


def public_gpt_config(review_root: Path) -> dict[str, Any]:
    config = read_gpt_config(review_root)
    return {
        "configured": config["configured"],
        "base_url": config["base_url"],
        "model": config["model"],
        "key_source": config["key_source"],
        "updated_at": config.get("updated_at", ""),
        "translation_prompt": config["translation_prompt"],
        "format_prompt": config["format_prompt"],
        "character_prompt": config["character_prompt"],
        "glossary_path": config["glossary_path"],
        "prompt_defaults": gpt_prompt_defaults(),
    }


def save_gpt_config(review_root: Path, data: dict[str, Any]) -> dict[str, Any]:
    base_url = normalize_gpt_base_url(str(data.get("base_url") or DEFAULT_GPT_BASE_URL))
    model = str(data.get("model") or DEFAULT_GPT_MODEL).strip() or DEFAULT_GPT_MODEL
    api_key = str(data.get("api_key") or "").strip()
    keep_existing = bool(data.get("keep_existing", False))
    existing = read_json(gpt_config_path(review_root), {})
    if keep_existing and not api_key:
        api_key = str(existing.get("api_key") or "").strip()
    prompts = {
        name: normalize_gpt_prompt(
            name,
            data.get(name) if name in data else existing.get(name, default),
        )
        for name, default in GPT_PROMPT_DEFAULTS.items()
    }
    record = {
        "base_url": base_url,
        "model": model,
        "api_key": api_key,
        "glossary_path": normalize_glossary_path_value(str(data.get("glossary_path") or existing.get("glossary_path") or "")),
        **prompts,
        "updated_at": now_iso(),
    }
    write_json(gpt_config_path(review_root), record)
    return public_gpt_config(review_root)


def glossary_path(review_root: Path) -> Path:
    data = read_json(gpt_config_path(review_root), {})
    configured = str(data.get("glossary_path") or "").strip()
    return Path(normalize_glossary_path_value(configured))


def read_glossary(review_root: Path) -> dict[str, str]:
    path = glossary_path(review_root)
    if not path.exists():
        return {}
    try:
        data = read_json(path, {})
    except json.JSONDecodeError as exc:
        raise ValueError(f"术语表 JSON 解析失败：{exc}") from exc
    if not isinstance(data, dict):
        raise ValueError("术语表必须是 JSON 对象")
    glossary: dict[str, str] = {}
    for key, value in data.items():
        if isinstance(key, str) and isinstance(value, str):
            glossary[key] = value
    return glossary


def public_glossary(review_root: Path) -> dict[str, Any]:
    path = glossary_path(review_root)
    glossary = read_glossary(review_root)
    entries = [
        {
            "source": key,
            "target": value,
            "clean_target": value.split("#", 1)[0].strip(),
        }
        for key, value in glossary.items()
    ]
    return {
        "path": str(path),
        "exists": path.exists(),
        "count": len(entries),
        "entries": entries,
        "updated_at": dt.datetime.fromtimestamp(path.stat().st_mtime).astimezone().isoformat(timespec="seconds") if path.exists() else "",
    }


def save_glossary(review_root: Path, entries: list[dict[str, Any]]) -> dict[str, Any]:
    path = glossary_path(review_root)
    if path.exists() and not path.is_file():
        raise ValueError("术语表路径不是文件")
    glossary: dict[str, str] = {}
    for entry in entries:
        if not isinstance(entry, dict):
            raise ValueError("术语条目必须是对象")
        source = str(entry.get("source") or "").strip()
        target = str(entry.get("target") or "").strip()
        if not source:
            continue
        if not target:
            raise ValueError(f"术语“{source}”缺少译名")
        if len(source) > 300:
            raise ValueError(f"术语“{source[:80]}...”过长")
        if len(target) > 1200:
            raise ValueError(f"术语“{source}”的译名/备注过长")
        if source in glossary:
            raise ValueError(f"术语“{source}”重复")
        glossary[source] = target
    path.parent.mkdir(parents=True, exist_ok=True)
    write_json(path, glossary)
    return public_glossary(review_root)


def search_glossary_entries(glossary: dict[str, str], term: str, limit: int = GLOSSARY_ENTRY_LIMIT) -> list[str]:
    needle = term.strip().lower()
    if not needle:
        return []
    selected: list[str] = []
    for key, value in glossary.items():
        clean_value = value.split("#", 1)[0].strip()
        if needle in key.lower() or needle in value.lower():
            selected.append(f"{key} => {clean_value or value}")
        if len(selected) >= limit:
            break
    return selected


def parenthetical_aliases(value: str) -> list[str]:
    aliases: list[str] = []
    for match in re.finditer(r"\(([^()]{1,80})\)|（([^（）]{1,80})）", value):
        alias = (match.group(1) or match.group(2) or "").strip()
        if alias:
            aliases.append(alias)
    return aliases


def strip_parenthetical_aliases(value: str) -> str:
    return re.sub(r"\([^()]{1,80}\)|（[^（）]{1,80}）", "", value).strip()


def glossary_match_variants(key: str, value: str) -> list[str]:
    variants: list[str] = []
    key = key.strip()
    clean_value = value.split("#", 1)[0].strip()
    if key:
        variants.append(key)
    variants.extend(parenthetical_aliases(key))
    variants.extend(parenthetical_aliases(clean_value))
    deduped: list[str] = []
    for item in variants:
        if len(item) < 2 or item in deduped:
            continue
        deduped.append(item)
    return deduped


def glossary_prompt_line(key: str, value: str, source: str) -> str:
    clean_value = value.split("#", 1)[0].strip()
    value_aliases = parenthetical_aliases(clean_value)
    keep_parenthetical = bool(parenthetical_aliases(key)) or any(alias in source for alias in value_aliases)
    translation = clean_value if keep_parenthetical else strip_parenthetical_aliases(clean_value) or clean_value
    return f"{key} => {translation}"


def glossary_matches_for_prompt(review_root: Path, row: dict[str, Any], limit: int = GLOSSARY_PROMPT_LIMIT) -> list[str]:
    try:
        glossary = read_glossary(review_root)
    except ValueError:
        raise
    except Exception as exc:
        raise ValueError(f"读取术语表失败：{exc}") from exc
    if not glossary:
        return []
    source = "\n".join(
        [
            str(row.get("jp_text") or ""),
            strip_html_keep_ruby_text(str(row.get("jp_html") or "")),
        ]
    )
    selected: list[str] = []
    for key, value in glossary.items():
        if not isinstance(key, str) or not isinstance(value, str):
            continue
        clean_value = value.split("#", 1)[0].strip()
        if not clean_value:
            continue
        if any(variant in source for variant in glossary_match_variants(key, value)):
            selected.append(glossary_prompt_line(key, value, source))
        if len(selected) >= limit:
            break
    return selected[:limit]


def sanitize_source_html_for_prompt(fragment: str) -> str:
    allowed = {"p", "br", "ruby", "rb", "rt", "rp", "span", "em", "strong", "b", "i"}

    def clean_tag(match: re.Match[str]) -> str:
        raw = match.group(0)
        closing = raw.startswith("</")
        name_match = re.match(r"</?\s*([a-zA-Z0-9:-]+)", raw)
        if not name_match:
            return ""
        name = name_match.group(1).lower()
        if name not in allowed:
            return ""
        if closing:
            return f"</{name}>"
        if name == "br" or raw.rstrip().endswith("/>"):
            return f"<{name} />"
        return f"<{name}>"

    return re.sub(r"<[^>]+>", clean_tag, fragment or "")


def glossary_mapping_from_data(data: Any) -> dict[str, str]:
    glossary: dict[str, str] = {}
    if isinstance(data, dict) and isinstance(data.get("terms"), list):
        for item in data.get("terms", []):
            if not isinstance(item, dict):
                continue
            source = str(item.get("source") or "").strip()
            target = str(item.get("target") or "").strip()
            if not source or not target:
                continue
            target_rt = str(item.get("target_rt") or "").strip()
            mode = str(item.get("mode") or "").strip()
            value = f"{target}({target_rt})" if target_rt and mode == "ruby" else target
            glossary[source] = value
            aliases = item.get("aliases") or []
            if isinstance(aliases, list):
                for alias in aliases:
                    alias_text = str(alias or "").strip()
                    if alias_text and alias_text not in glossary:
                        glossary[alias_text] = value
            rt = str(item.get("rt") or "").strip()
            if rt and rt not in glossary:
                glossary[rt] = value
        return glossary
    if isinstance(data, dict):
        for key, value in data.items():
            if isinstance(key, str) and isinstance(value, str):
                glossary[key] = value
    return glossary


def read_glossary_mapping_from_path(path: Path) -> dict[str, str]:
    if not path.exists():
        return {}
    try:
        data = read_json(path, {})
    except json.JSONDecodeError as exc:
        raise ValueError(f"术语表 JSON 解析失败：{exc}") from exc
    return glossary_mapping_from_data(data)


def glossary_matches_for_source(
    glossary: dict[str, str],
    source_html: str,
    source_text: str = "",
    limit: int = GLOSSARY_PROMPT_LIMIT,
) -> list[str]:
    if not glossary:
        return []
    source = "\n".join([str(source_text or ""), strip_html_keep_ruby_text(source_html or "")])
    selected: list[str] = []
    for key, value in glossary.items():
        if not isinstance(key, str) or not isinstance(value, str):
            continue
        clean_value = value.split("#", 1)[0].strip()
        if not clean_value:
            continue
        if any(variant in source for variant in glossary_match_variants(key, value)):
            selected.append(glossary_prompt_line(key, value, source))
        if len(selected) >= limit:
            break
    return selected[:limit]


def truncate_prompt_text(value: Any, limit: int = 1600) -> str:
    text = compact_text(str(value or ""))
    if len(text) <= limit:
        return text
    return text[: limit - 1].rstrip() + "…"


def row_translation_html(row: dict[str, Any]) -> str:
    return str(row.get("current_html") or row.get("cn_html") or row.get("cn_text") or "")


def row_source_html(row: dict[str, Any]) -> str:
    return str(row.get("jp_html") or row.get("jp_text") or "")


def row_source_html_for_prompt(row: dict[str, Any]) -> str:
    allowed = {"p", "br", "ruby", "rb", "rt", "rp", "span", "em", "strong", "b", "i"}

    def clean_tag(match: re.Match[str]) -> str:
        raw = match.group(0)
        closing = raw.startswith("</")
        name_match = re.match(r"</?\s*([a-zA-Z0-9:-]+)", raw)
        if not name_match:
            return ""
        name = name_match.group(1).lower()
        if name not in allowed:
            return ""
        if closing:
            return f"</{name}>"
        if name == "br" or raw.rstrip().endswith("/>"):
            return f"<{name} />"
        return f"<{name}>"

    return re.sub(r"<[^>]+>", clean_tag, row_source_html(row))


def build_retranslate_context(row: dict[str, Any], rows: list[dict[str, Any]], radius: int = 1) -> list[str]:
    same_file_rows = [item for item in rows if item.get("file") == row.get("file")]
    index = next((idx for idx, item in enumerate(same_file_rows) if item.get("id") == row.get("id")), -1)
    if index < 0:
        return []
    start = max(0, index - radius)
    end = min(len(same_file_rows), index + radius + 1)
    context: list[str] = []
    for idx in range(start, end):
        item = same_file_rows[idx]
        if item.get("id") == row.get("id"):
            continue
        direction = "前文" if idx < index else "后文"
        context.append(f"{direction} {item.get('id', '')}: {truncate_prompt_text(item.get('jp_text'), 260)}")
    return context


def normalize_candidate_punctuation(fragment: str) -> str:
    protected: list[str] = []

    def protect(match: re.Match[str]) -> str:
        protected.append(match.group(0))
        return f"\uE000{len(protected) - 1}\uE001"

    protected_pattern = (
        r"<[^>]+>|"
        r"&(?:#[0-9]+|#x[0-9a-fA-F]+|[A-Za-z][A-Za-z0-9]+);|"
        r"https?://[^\s<>\"]+|"
        r"[\w.+-]+@[\w.-]+\.[A-Za-z]{2,}|"
        r"\b\d+(?:[.,]\d+)+\b|"
        r"\b[A-Za-z0-9]+(?:\.[A-Za-z0-9]+)+\b"
    )
    text = re.sub(protected_pattern, protect, fragment)
    text = text.replace("“", "「").replace("”", "」")
    text = text.replace("『", "「").replace("』", "」")
    text = text.replace("...", "……").replace("…", "……")
    text = re.sub(r"(?:……)+", "……", text)
    text = re.sub(r"[-]{2,}", "——", text)
    text = text.translate(str.maketrans({
        ",": "，",
        ".": "。",
        ":": "：",
        ";": "；",
    }))
    text = text.replace("?", "？").replace("!", "！")
    text = re.sub(r"？+！+", "？！", text)
    text = re.sub(r"！+？+", "！？", text)
    text = re.sub(r"？{2,}", "？", text)
    text = re.sub(r"！{2,}", "！", text)

    def restore(match: re.Match[str]) -> str:
        return protected[int(match.group(1))]

    return re.sub(r"\uE000(\d+)\uE001", restore, text)


def restore_glossary_ruby_candidates(fragment: str, glossary_lines: list[str]) -> str:
    pairs: list[tuple[str, str]] = []
    for line in glossary_lines:
        value = line.split("=>", 1)[1].strip() if "=>" in line else line
        match = re.fullmatch(r"(.+?)\(([^()<>]{1,80})\)", value)
        if match:
            rb, rt = match.group(1).strip(), match.group(2).strip()
            if rb and rt and len(rb) <= 80:
                pairs.append((rb, rt))
    if not pairs:
        return fragment
    protected: list[str] = []

    def protect(match: re.Match[str]) -> str:
        protected.append(match.group(0))
        return f"\uE010{len(protected) - 1}\uE011"

    text = re.sub(r"<ruby\b.*?</ruby>|<[^>]+>", protect, fragment, flags=re.S | re.I)
    for rb, rt in sorted(pairs, key=lambda item: len(item[0]), reverse=True):
        pattern = re.escape(f"{rb}({rt})")
        ruby = f"<ruby><rb>{html.escape(rb)}</rb><rt>{html.escape(rt)}</rt></ruby>"
        text = re.sub(pattern, ruby, text)

    def restore(match: re.Match[str]) -> str:
        return protected[int(match.group(1))]

    return re.sub(r"\uE010(\d+)\uE011", restore, text)


def build_retranslate_messages(
    row: dict[str, Any],
    rows: list[dict[str, Any]],
    config: dict[str, Any],
    glossary_lines: list[str],
    instruction: str = "",
) -> list[dict[str, str]]:
    glossary_text = "\n".join(f"- {item}" for item in glossary_lines)
    context_lines = build_retranslate_context(row, rows)
    system_prompt = "\n\n".join(
        [
            "你是日语轻小说到简体中文的审校重翻助手。只处理一个目标语句/段落。",
            "只输出新的简体中文译文 HTML 片段，不解释，不输出 Markdown。",
            "不得合并、拆分、补写或删除目标段落信息。",
        ]
    )
    user_parts = [
        "【翻译内容提示词】",
        str(config.get("translation_prompt") or DEFAULT_TRANSLATION_PROMPT).strip(),
        "",
        "【目标语句：日文原文 HTML】",
        row_source_html_for_prompt(row),
    ]
    if glossary_text:
        user_parts.extend(["", "【本段命中术语】", glossary_text])
    if context_lines:
        user_parts.extend(["", "【少量上下文】", "\n".join(f"- {item}" for item in context_lines)])
    user_parts.extend(
        [
            "",
            "【角色状态与口吻提示词】",
            str(config.get("character_prompt") or DEFAULT_CHARACTER_PROMPT).strip(),
        ]
    )
    user_parts.extend(
        [
            "【格式与标点提示词】",
            str(config.get("format_prompt") or DEFAULT_FORMAT_PROMPT).strip(),
        ]
    )
    if instruction.strip():
        user_parts.extend(["", "【本次额外要求】", instruction.strip()[:800]])
    user_parts.extend(
        [
            "",
            "请只输出目标语句的新中文译文 HTML 片段。",
        ]
    )
    return [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": "\n".join(user_parts)},
    ]


def call_openai_compatible_chat(
    config: dict[str, Any],
    messages: list[dict[str, str]],
    temperature: float = 0.2,
    timeout: int = 90,
) -> str:
    api_key = str(config.get("api_key") or "").strip()
    if not api_key:
        raise ValueError("尚未配置 GPT API Key")
    base_url = normalize_gpt_base_url(str(config.get("base_url") or DEFAULT_GPT_BASE_URL))
    url = f"{base_url}/chat/completions"
    payload = {
        "model": str(config.get("model") or DEFAULT_GPT_MODEL),
        "messages": messages,
        "temperature": temperature,
    }
    body = json.dumps(payload, ensure_ascii=False).encode("utf-8")
    req = urllib.request.Request(
        url,
        data=body,
        headers={
            "Content-Type": "application/json",
            "Authorization": f"Bearer {api_key}",
        },
        method="POST",
    )
    try:
        with urllib.request.urlopen(req, timeout=timeout) as response:
            result = json.loads(response.read().decode("utf-8", errors="replace"))
    except urllib.error.HTTPError as exc:
        detail = exc.read().decode("utf-8", errors="replace")[:1200]
        raise RuntimeError(f"GPT 请求失败：HTTP {exc.code} {detail}") from exc
    except urllib.error.URLError as exc:
        raise RuntimeError(f"GPT 请求失败：{exc.reason}") from exc
    choices = result.get("choices") or []
    if not choices:
        raise RuntimeError("GPT 返回为空")
    content = ((choices[0].get("message") or {}).get("content") or "").strip()
    if content.startswith("```"):
        content = re.sub(r"^```(?:html|xml)?\s*", "", content, flags=re.I).strip()
        content = re.sub(r"\s*```$", "", content).strip()
    if not content:
        raise RuntimeError("GPT 没有返回译文")
    return content


def session_root_for(review_root: Path, epub_path: Path) -> Path:
    digest = hashlib.sha1(str(epub_path.resolve()).encode("utf-8")).hexdigest()[:10]
    return review_root / f"{safe_slug(epub_path.stem)}_{digest}"


def app_manifest_path(review_root: Path) -> Path:
    return review_root / "app_state.json"


def gpt_config_path(review_root: Path) -> Path:
    return review_root / "gpt_config.json"


def session_manifest_path(session_root: Path) -> Path:
    return session_root / "review_state" / "session.json"


def session_public_id(session_root: Path) -> str:
    return session_root.name


def session_root_from_id(review_root: Path, session_id: str) -> Path:
    safe_id = safe_slug(session_id, max_len=240)
    if safe_id != session_id:
        raise ValueError("invalid session id")
    root = (review_root / session_id).resolve()
    try:
        root.relative_to(review_root.resolve())
    except ValueError as exc:
        raise ValueError("invalid session id") from exc
    return root


def write_session_manifest(session_root: Path, epub_path: Path, source_name: str = "") -> None:
    existing = read_json(session_manifest_path(session_root), {})
    record = {
        "id": session_public_id(session_root),
        "source_name": source_name or epub_path.name,
        "source_epub": str(epub_path),
        "session_root": str(session_root),
        "created_at": existing.get("created_at") or now_iso(),
        "updated_at": now_iso(),
    }
    write_json(session_manifest_path(session_root), record)


def update_app_manifest(review_root: Path, session_root: Path | None) -> None:
    data = read_json(app_manifest_path(review_root), {})
    data["active_session_id"] = session_public_id(session_root) if session_root else ""
    data["updated_at"] = now_iso()
    write_json(app_manifest_path(review_root), data)


def session_epub_path(session_root: Path) -> Path:
    manifest = read_json(session_manifest_path(session_root), {})
    source_epub = manifest.get("source_epub") or read_json(
        session_root / "review_state" / "state.json", {}
    ).get("source_epub")
    if source_epub:
        path = Path(source_epub)
        if path.exists():
            return path
    fallback = session_root / "source.epub"
    if fallback.exists():
        return fallback
    return Path(source_epub or fallback)


def normalize_reading_position(data: Any) -> dict[str, Any]:
    if not isinstance(data, dict):
        return {}

    def field(name: str, limit: int = 160) -> str:
        return str(data.get(name) or "").strip()[:limit]

    def nonnegative_int(name: str, limit: int = 10_000_000) -> int:
        try:
            value = int(float(data.get(name) or 0))
        except (TypeError, ValueError):
            value = 0
        return max(0, min(value, limit))

    mode = field("mode", 20)
    if mode not in {"reading", "polish"}:
        mode = "reading"
    scope = field("scope", 20)
    if scope not in {"chapter", "part", "image", "none"}:
        scope = "none"
    position = {
        "mode": mode,
        "scope": scope,
        "chapter_id": field("chapter_id"),
        "part_id": field("part_id"),
        "image_item_id": field("image_item_id"),
        "row_id": field("row_id"),
        "scroll_top": nonnegative_int("scroll_top"),
        "offset": nonnegative_int("offset", 5000),
        "updated_at": field("updated_at", 80) or now_iso(),
    }
    return position


def read_reading_position(session_root: Path) -> dict[str, Any]:
    state = read_json(session_root / "review_state" / "state.json", {})
    return normalize_reading_position(state.get("reading_position", {}))


def summarize_session(session_root: Path) -> dict[str, Any]:
    manifest = read_json(session_manifest_path(session_root), {})
    state = read_json(session_root / "review_state" / "state.json", {})
    rows = merge_rows(session_root) if (session_root / "review_state" / "rows.json").exists() else []
    touched = [row for row in rows if rowTouched_py(row)]
    source_epub = manifest.get("source_epub") or state.get("source_epub", "")
    source_name = manifest.get("source_name") or (Path(source_epub).name if source_epub else session_root.name)
    exports_dir = session_root / "exports"
    latest_export = ""
    if exports_dir.exists():
        exports = sorted(exports_dir.glob("*.epub"), key=lambda p: p.stat().st_mtime, reverse=True)
        latest_export = str(exports[0]) if exports else ""
    updated_at = state.get("updated_at") or manifest.get("updated_at") or ""
    return {
        "id": session_public_id(session_root),
        "version": version,
        "source_name": source_name,
        "source_epub": source_epub,
        "session_root": str(session_root),
        "row_count": len(rows),
        "touched_count": len(touched),
        "marked_count": sum(1 for row in rows if row.get("marked")),
        "updated_at": updated_at,
        "created_at": state.get("created_at") or manifest.get("created_at") or "",
        "feedback_md": str(session_root / "feedback" / "feedback_for_codex.md"),
        "feedback_jsonl": str(session_root / "feedback" / "translation_feedback.jsonl"),
        "latest_export": latest_export,
        "reading_position": normalize_reading_position(state.get("reading_position", {})),
    }


def rowTouched_py(row: dict[str, Any]) -> bool:
    return bool(
        row.get("edited")
        or row.get("marked")
        or row.get("comment")
        or row.get("learn_note")
        or row.get("issue_type")
        or row.get("severity")
        or row.get("tags")
    )


def list_review_sessions(review_root: Path) -> list[dict[str, Any]]:
    if not review_root.exists():
        return []
    sessions = []
    for child in review_root.iterdir():
        if not child.is_dir() or child.name == UPLOAD_DIR_NAME:
            continue
        if not (child / "review_state" / "state.json").exists():
            continue
        try:
            sessions.append(summarize_session(child))
        except Exception:
            continue
    sessions.sort(key=lambda item: item.get("updated_at") or item.get("created_at") or "", reverse=True)
    return sessions


def validate_epub_file(path: Path) -> None:
    if path.suffix.lower() != ".epub":
        raise ValueError("只支持 .epub 文件")
    if not zipfile.is_zipfile(path):
        raise ValueError("文件不是有效的 EPUB/ZIP 包")


def copy_upload_to_review_root(file_storage: Any, review_root: Path) -> Path:
    filename = safe_slug(Path(file_storage.filename or "uploaded.epub").name, max_len=140)
    if not filename.lower().endswith(".epub"):
        filename = f"{filename}.epub"
    upload_dir = review_root / UPLOAD_DIR_NAME
    upload_dir.mkdir(parents=True, exist_ok=True)
    stem = safe_slug(Path(filename).stem, max_len=100)
    target = upload_dir / f"{dt.datetime.now().strftime('%Y%m%d_%H%M%S')}_{stem}.epub"
    try:
        file_storage.save(target)
        validate_epub_file(target)
        return target
    except Exception:
        target.unlink(missing_ok=True)
        raise


def extracted_path(session_root: Path, entry_name: str) -> Path:
    root = (session_root / "extracted").resolve()
    target = (root / entry_name).resolve()
    try:
        target.relative_to(root)
    except ValueError as exc:
        raise ValueError(f"unsafe epub entry path: {entry_name}") from exc
    return target


def extract_epub(epub_path: Path, session_root: Path, reset: bool = False) -> None:
    extracted = session_root / "extracted"
    if reset and session_root.exists():
        shutil.rmtree(session_root)
    if extracted.exists():
        return
    session_root.mkdir(parents=True, exist_ok=True)
    shutil.copy2(epub_path, session_root / "source.epub")
    extracted.mkdir(parents=True, exist_ok=True)
    entries = []
    with zipfile.ZipFile(epub_path, "r") as zf:
        for info in zf.infolist():
            entries.append(
                {
                    "filename": info.filename,
                    "compress_type": int(info.compress_type),
                    "date_time": list(info.date_time),
                    "is_dir": info.is_dir(),
                }
            )
            target = extracted_path(session_root, info.filename)
            if info.is_dir():
                target.mkdir(parents=True, exist_ok=True)
                continue
            target.parent.mkdir(parents=True, exist_ok=True)
            target.write_bytes(zf.read(info.filename))
    write_json(session_root / "review_state" / "epub_entries.json", entries)


def find_opf_entry(session_root: Path) -> str:
    extracted = session_root / "extracted"
    container = extracted / "META-INF" / "container.xml"
    root = read_xml_root(container)
    if root is not None:
        for elem in root.iter():
            if local_name(elem.tag) == "rootfile":
                full_path = elem.attrib.get("full-path") or elem.attrib.get("full_path")
                if full_path:
                    return posixpath.normpath(full_path).lstrip("./")
    entries = read_json(session_root / "review_state" / "epub_entries.json", [])
    for item in entries:
        name = item.get("filename", "")
        if not item.get("is_dir") and name.lower().endswith(".opf"):
            return name
    return ""


def parse_opf_metadata(session_root: Path) -> dict[str, Any]:
    opf_entry = find_opf_entry(session_root)
    if not opf_entry:
        return {"opf_entry": "", "opf_dir": "", "spine": []}
    opf_path = extracted_path(session_root, opf_entry)
    root = read_xml_root(opf_path)
    if root is None:
        return {"opf_entry": opf_entry, "opf_dir": entry_dir(opf_entry), "spine": []}

    opf_dir = entry_dir(opf_entry)
    manifest: dict[str, dict[str, str]] = {}
    spine_idrefs: list[str] = []
    for elem in root.iter():
        name = local_name(elem.tag)
        if name == "item":
            item_id = elem.attrib.get("id", "")
            href = elem.attrib.get("href", "")
            if item_id and href:
                manifest[item_id] = {
                    "href": normalize_href(opf_dir, href),
                    "media_type": elem.attrib.get("media-type", ""),
                    "properties": elem.attrib.get("properties", ""),
                }
        elif name == "itemref":
            idref = elem.attrib.get("idref", "")
            if idref:
                spine_idrefs.append(idref)

    spine = []
    for idref in spine_idrefs:
        item = manifest.get(idref)
        if not item:
            continue
        href = item["href"]
        if href.lower().endswith((".xhtml", ".html", ".htm")):
            spine.append(href)
    return {
        "opf_entry": opf_entry,
        "opf_dir": opf_dir,
        "manifest": manifest,
        "spine": spine,
    }


def html_entries_in_reading_order(session_root: Path) -> list[str]:
    entries = read_json(session_root / "review_state" / "epub_entries.json", [])
    all_html = [
        item["filename"]
        for item in entries
        if not item.get("is_dir")
        and item["filename"].lower().endswith((".xhtml", ".html", ".htm"))
    ]
    opf = parse_opf_metadata(session_root)
    ordered: list[str] = []
    for item in opf.get("spine", []):
        if item in all_html and item not in ordered:
            ordered.append(item)
    for item in sorted(all_html, key=file_sort_key):
        basename = posixpath.basename(item).lower()
        if item in ordered or basename in {"nav.xhtml", "toc.xhtml"}:
            continue
        ordered.append(item)
    return ordered


def extract_heading_from_html(text: str) -> str:
    for pattern in (
        r"<h[1-6]\b[^>]*>(.*?)</h[1-6]>",
        r"<title\b[^>]*>(.*?)</title>",
    ):
        match = re.search(pattern, text, flags=re.S | re.I)
        if not match:
            continue
        title = compact_text(strip_tags(match.group(1)))
        if title:
            return title
    return ""


def extract_body_heading_from_html(text: str) -> str:
    match = re.search(r"<h[1-6]\b[^>]*>(.*?)</h[1-6]>", text, flags=re.S | re.I)
    if not match:
        return ""
    return compact_text(strip_tags(match.group(1)))


def read_document_title(session_root: Path, entry_name: str) -> str:
    path = extracted_path(session_root, entry_name)
    if not path.exists():
        return ""
    text = path.read_text(encoding="utf-8", errors="replace")
    return extract_heading_from_html(text)


def read_body_heading(session_root: Path, entry_name: str) -> str:
    path = extracted_path(session_root, entry_name)
    if not path.exists():
        return ""
    text = path.read_text(encoding="utf-8", errors="replace")
    return extract_body_heading_from_html(text)


def first_content_text(session_root: Path, entry_name: str) -> str:
    path = extracted_path(session_root, entry_name)
    if not path.exists():
        return ""
    text = path.read_text(encoding="utf-8", errors="replace")
    for match in P_RE.finditer(text):
        inner = match.group(2)
        if re.search(r"<img\b", inner, flags=re.I):
            continue
        label = compact_text(strip_tags(inner))
        if label:
            return label
    return ""


def tag_attrs(tag: str) -> dict[str, str]:
    attrs: dict[str, str] = {}
    for key, _quote, value in ATTR_RE.findall(tag):
        attrs[key.lower()] = html.unescape(value or "")
    return attrs


def part_display_label(entry_name: str, fallback_index: int) -> str:
    stem = Path(entry_name).stem
    for pattern in (r"part[-_]?(\d+)$", r"p[-_]?(\d+)$", r"^(\d+)$"):
        match = re.search(pattern, stem, flags=re.I)
        if match:
            return f"part{int(match.group(1)):04d}"
    if stem and stem.lower() not in {"titlepage", "cover"}:
        return stem
    return f"part{fallback_index:04d}"


def is_image_asset(entry_name: str) -> bool:
    suffix = posixpath.splitext(entry_name.split("?", 1)[0])[1].lower()
    return suffix in IMAGE_EXTENSIONS


def is_raster_image_asset(entry_name: str) -> bool:
    suffix = posixpath.splitext(entry_name.split("?", 1)[0])[1].lower()
    return suffix in RASTER_IMAGE_EXTENSIONS


def asset_url(entry_name: str) -> str:
    return f"/api/asset/{urllib.parse.quote(entry_name, safe='/')}"


def extract_images_from_html(session_root: Path, entry_name: str) -> list[dict[str, Any]]:
    path = extracted_path(session_root, entry_name)
    if not path.exists():
        return []
    text = path.read_text(encoding="utf-8", errors="replace")
    images: list[dict[str, Any]] = []
    for index, match in enumerate(IMG_TAG_RE.finditer(text), start=1):
        attrs = tag_attrs(match.group(0))
        href = attrs.get("src") or attrs.get("href") or attrs.get("xlink:href")
        if not href or EXTERNAL_URI_RE.match(href):
            continue
        image_entry = normalize_href(entry_dir(entry_name), href)
        if not image_entry or not is_raster_image_asset(image_entry):
            continue
        images.append(
            {
                "index": index,
                "file": entry_name,
                "file_label": Path(entry_name).name,
                "asset_path": image_entry,
                "asset_name": Path(image_entry).name,
                "asset_url": asset_url(image_entry),
                "alt": compact_text(attrs.get("alt") or attrs.get("title") or ""),
            }
        )
    return images


def collect_asset_paths_from_structure(structure: dict[str, Any]) -> set[str]:
    paths: set[str] = set()
    for chapter in structure.get("chapters", []):
        for image in chapter.get("images", []):
            asset_path = image.get("asset_path", "")
            if asset_path:
                paths.add(asset_path)
        for item in chapter.get("items", []):
            for image in item.get("images", []):
                asset_path = image.get("asset_path", "")
                if asset_path:
                    paths.add(asset_path)
    return paths


def image_item_count(chapters: list[dict[str, Any]]) -> int:
    return sum(1 for chapter in chapters for item in chapter.get("items", []) if item.get("kind") == "image")


def nav_target_key(base_dir: str, href: str) -> tuple[str, str]:
    path, fragment = normalize_href_with_fragment(base_dir, href)
    return path, fragment


def parse_internal_toc(session_root: Path) -> list[dict[str, Any]]:
    ordered_files = html_entries_in_reading_order(session_root)
    ordered_set = set(ordered_files)
    for entry_name in ordered_files:
        path = extracted_path(session_root, entry_name)
        if not path.exists():
            continue
        text = path.read_text(encoding="utf-8", errors="replace")
        if not re.search(r"<a\b[^>]*href=", text, flags=re.I):
            continue
        basename = posixpath.basename(entry_name).lower()
        heading = extract_heading_from_html(text)
        is_toc_like_name = basename in {"nav.xhtml", "toc.xhtml", "toc.html", "toc.htm"} or "toc" in basename
        is_toc_like_title = bool(re.search(r"(目录|目次|contents?)", heading, flags=re.I))
        matches = list(P_RE.finditer(text))
        base_dir = entry_dir(entry_name)
        candidate_items: list[dict[str, Any]] = []
        i = 0
        while i < len(matches):
            source_attrs = matches[i].group(1)[2:-1]
            source_inner = matches[i].group(2)
            link_match = re.search(r"<a\b[^>]*href=['\"]([^'\"]+)['\"][^>]*>(.*?)</a>", source_inner, flags=re.S | re.I)
            if not link_match:
                i += 1
                continue
            target_file, target_fragment = nav_target_key(base_dir, link_match.group(1))
            if not target_file:
                i += 1
                continue
            if target_file not in ordered_set:
                i += 1
                continue
            source_title = compact_text(strip_tags(link_match.group(2))) or compact_text(strip_tags(source_inner))
            title = source_title
            if is_gray_source(source_attrs, source_inner) and i + 1 < len(matches):
                next_attrs = matches[i + 1].group(1)[2:-1]
                next_inner = matches[i + 1].group(2)
                if not is_gray_source(next_attrs, next_inner):
                    next_link = re.search(r"<a\b[^>]*href=['\"]([^'\"]+)['\"][^>]*>(.*?)</a>", next_inner, flags=re.S | re.I)
                    same_target = True
                    if next_link:
                        next_target_file, next_target_fragment = nav_target_key(base_dir, next_link.group(1))
                        same_target = next_target_file == target_file and next_target_fragment == target_fragment
                    next_title = compact_text(strip_tags(next_link.group(2) if next_link else next_inner))
                    if next_title:
                        if same_target:
                            title = next_title
                            i += 1
            if title:
                candidate_items.append(
                    {
                        "title": title,
                        "href": target_file,
                        "fragment": target_fragment,
                        "children": [],
                        "source_title": source_title,
                    }
                )
            i += 1
        if candidate_items:
            internal_targets = sum(1 for item in candidate_items if item["href"] in ordered_set and item["href"] != entry_name)
            link_density = internal_targets / max(1, len(matches))
            if is_toc_like_name or is_toc_like_title or (len(candidate_items) >= 4 and link_density >= 0.35):
                return candidate_items
    return []


def parse_nav_toc(session_root: Path) -> list[dict[str, Any]]:
    entries = read_json(session_root / "review_state" / "epub_entries.json", [])
    nav_entries = [
        item["filename"]
        for item in entries
        if not item.get("is_dir")
        and posixpath.basename(item["filename"]).lower() in {"nav.xhtml", "toc.xhtml", "toc.html", "toc.htm"}
    ]
    opf = parse_opf_metadata(session_root)
    for manifest_item in opf.get("manifest", {}).values():
        if "nav" in manifest_item.get("properties", "").split():
            href = manifest_item.get("href", "")
            if href and href not in nav_entries:
                nav_entries.insert(0, href)

    for entry_name in nav_entries:
        path = extracted_path(session_root, entry_name)
        root = read_xml_root(path)
        if root is None:
            continue
        base_dir = entry_dir(entry_name)
        toc_root: ET.Element | None = None
        for elem in root.iter():
            if local_name(elem.tag) != "nav":
                continue
            nav_type = elem.attrib.get("{http://www.idpf.org/2007/ops}type", "") or elem.attrib.get("epub:type", "")
            if "toc" in nav_type.split() or toc_root is None:
                toc_root = elem
                if "toc" in nav_type.split():
                    break
        if toc_root is None:
            continue
        top_ol = next((child for child in list(toc_root) if local_name(child.tag) == "ol"), None)
        if top_ol is None:
            top_ol = next((elem for elem in toc_root.iter() if local_name(elem.tag) == "ol"), None)
        if top_ol is None:
            continue
        items = parse_nav_ol(top_ol, base_dir)
        if items:
            return items
    return []


def parse_nav_ol(ol: ET.Element, base_dir: str) -> list[dict[str, Any]]:
    items: list[dict[str, Any]] = []
    for li in list(ol):
        if local_name(li.tag) != "li":
            continue
        label = ""
        href = ""
        fragment = ""
        children: list[dict[str, Any]] = []
        for child in list(li):
            name = local_name(child.tag)
            if name == "a":
                label = element_text(child)
                href, fragment = normalize_href_with_fragment(base_dir, child.attrib.get("href", ""))
            elif name == "span" and not label:
                label = element_text(child)
            elif name == "ol":
                children = parse_nav_ol(child, base_dir)
        if not label:
            label = element_text(li)
        items.append({"title": label, "href": href, "fragment": fragment, "children": children})
    return items


def flatten_nav_items(items: list[dict[str, Any]]) -> list[dict[str, Any]]:
    flattened: list[dict[str, Any]] = []

    def walk(nodes: list[dict[str, Any]], depth: int) -> None:
        for node in nodes:
            href = node.get("href", "")
            if href:
                flattened.append({"title": node.get("title") or href, "href": href, "depth": depth})
            walk(node.get("children", []), depth + 1)

    walk(items, 0)
    return flattened


def build_structure(session_root: Path) -> dict[str, Any]:
    rows = merge_rows(session_root)
    rows_by_file: dict[str, list[dict[str, Any]]] = {}
    for row in rows:
        rows_by_file.setdefault(row["file"], []).append(row)

    all_ordered_files = html_entries_in_reading_order(session_root)
    image_by_file = {
        entry_name: images
        for entry_name in all_ordered_files
        if (images := extract_images_from_html(session_root, entry_name))
    }
    ordered_files = [file for file in all_ordered_files if file in rows_by_file]
    for file in rows_by_file:
        if file not in ordered_files:
            ordered_files.append(file)
    content_files = [file for file in all_ordered_files if file in rows_by_file or file in image_by_file]
    for file in sorted(set(rows_by_file) | set(image_by_file), key=file_sort_key):
        if file not in content_files:
            content_files.append(file)

    nav_tree = parse_nav_toc(session_root)
    nav_flat = flatten_nav_items(nav_tree)
    nav_titles: dict[str, str] = {}
    for item in nav_flat:
        href = item.get("href", "")
        title = item.get("title", "")
        if href and title and href not in nav_titles:
            nav_titles[href] = title

    start_titles: dict[str, str] = {}
    chapter_specs: list[dict[str, Any]] = []
    toc_items = parse_internal_toc(session_root)
    ordered_index = {entry_name: index for index, entry_name in enumerate(all_ordered_files)}
    content_file_set = set(content_files)

    def first_content_file_at_or_after(entry_name: str) -> str:
        start_pos = ordered_index.get(entry_name, -1)
        if start_pos < 0:
            return ""
        for candidate in all_ordered_files[start_pos:]:
            if candidate in content_file_set:
                return candidate
        return ""

    for item in toc_items:
        href = item.get("href", "")
        if href not in ordered_index:
            continue
        start = first_content_file_at_or_after(href)
        if start and item.get("title") and start not in start_titles:
            start_titles[start] = item["title"]

    if not start_titles:
        for item in nav_flat:
            href = item.get("href", "")
            if href not in ordered_index:
                continue
            start = first_content_file_at_or_after(href)
            if start and item.get("title") and start not in start_titles:
                start_titles[start] = item["title"]

    if not start_titles:
        for entry_name in ordered_files:
            heading = read_body_heading(session_root, entry_name)
            if heading and entry_name not in start_titles and len(strip_tags(heading)) <= 80:
                start_titles[entry_name] = heading

    if start_titles:
        starts = sorted(start_titles, key=lambda name: ordered_index.get(name, 10**9))
        first_start_pos = ordered_index.get(starts[0], len(all_ordered_files)) if starts else len(all_ordered_files)
        preface_files = [
            entry_name
            for entry_name in all_ordered_files[:first_start_pos]
            if entry_name in rows_by_file or entry_name in image_by_file
        ]
        if preface_files:
            chapter_specs.append({"title": "卷首", "files": preface_files, "start": preface_files[0]})
        for index, start in enumerate(starts):
            start_pos = ordered_index.get(start, -1)
            next_pos = ordered_index.get(starts[index + 1], len(all_ordered_files)) if index + 1 < len(starts) else len(all_ordered_files)
            if start_pos < 0:
                continue
            files = [
                entry_name
                for entry_name in all_ordered_files[start_pos:next_pos]
                if entry_name in rows_by_file or entry_name in image_by_file
            ]
            if files:
                chapter_specs.append({"title": start_titles.get(start, ""), "files": files, "start": files[0]})

        grouped_files = {entry_name for spec in chapter_specs for entry_name in spec["files"]}
        for entry_name in content_files:
            if entry_name not in grouped_files:
                chapter_specs.append({"title": "", "files": [entry_name], "start": entry_name})
        chapter_specs.sort(key=lambda spec: ordered_index.get(spec["start"], 10**9))

    part_serial = 0
    image_serial = 0

    def file_counts(file_rows: list[dict[str, Any]]) -> tuple[int, int]:
        touched_count = sum(
            1
            for row in file_rows
            if row["edited"]
            or row["marked"]
            or row["comment"]
            or row["learn_note"]
            or row["issue_type"]
            or row["severity"]
            or row["tags"]
        )
        marked_count = sum(1 for row in file_rows if row["marked"])
        return touched_count, marked_count

    def make_part(entry_name: str, title_override: str = "") -> dict[str, Any]:
        nonlocal part_serial
        part_serial += 1
        file_rows = rows_by_file[entry_name]
        title = part_display_label(entry_name, part_serial - 1)
        touched_count, marked_count = file_counts(file_rows)
        return {
            "id": f"PT{part_serial:04d}",
            "kind": "text",
            "title": title,
            "source_title": title_override
            or nav_titles.get(entry_name)
            or read_body_heading(session_root, entry_name)
            or Path(entry_name).stem,
            "file": entry_name,
            "file_label": Path(entry_name).name,
            "row_ids": [row["id"] for row in file_rows],
            "row_count": len(file_rows),
            "touched_count": touched_count,
            "marked_count": marked_count,
            "first_row_id": file_rows[0]["id"] if file_rows else "",
            "last_row_id": file_rows[-1]["id"] if file_rows else "",
        }

    def make_image_item(
        entry_name: str,
        image: dict[str, Any],
        title_override: str = "",
        image_index: int = 1,
        image_total: int = 1,
    ) -> dict[str, Any]:
        nonlocal image_serial
        image_serial += 1
        suffix = f"-{image_index:02d}" if image_total > 1 else ""
        label = part_display_label(entry_name, image_serial - 1)
        is_cover = Path(entry_name).stem.lower() in {"titlepage", "cover"}
        default_title = "封面" if is_cover else f"插图 {label}{suffix}"
        base_title = title_override or nav_titles.get(entry_name) or read_body_heading(session_root, entry_name)
        if is_cover:
            title = default_title
        elif title_override and base_title:
            title = f"{base_title} · 插图{suffix}"
        else:
            title = base_title or default_title
        image_item = {
            **image,
            "id": f"IMG{image_serial:04d}-{image['index']:02d}",
        }
        return {
            "id": image_item["id"],
            "kind": "image",
            "title": title,
            "file": entry_name,
            "file_label": Path(entry_name).name,
            "images": [image_item],
            "row_count": 0,
            "touched_count": 0,
            "marked_count": 0,
            "parts": [],
            "first_image_id": image_item["id"],
            "last_image_id": image_item["id"],
        }

    def image_chapter_from_item(chapter_id: str, item: dict[str, Any]) -> dict[str, Any]:
        return {
            **item,
            "id": chapter_id,
            "items": [item],
            "_sort": ordered_index.get(item.get("file", ""), 10**9),
        }

    def chapter_from_items(chapter_id: str, title: str, items: list[dict[str, Any]]) -> dict[str, Any]:
        parts = [item for item in items if item.get("kind") == "text"]
        images = [image for item in items if item.get("kind") == "image" for image in item.get("images", [])]
        return {
            "id": chapter_id,
            "kind": "text",
            "title": title,
            "row_count": sum(part["row_count"] for part in parts),
            "touched_count": sum(part["touched_count"] for part in parts),
            "marked_count": sum(part["marked_count"] for part in parts),
            "parts": parts,
            "images": images,
            "items": items,
            "_sort": min((ordered_index.get(item.get("file", ""), 10**9) for item in items), default=10**9),
        }

    def chapter_from_parts(chapter_id: str, title: str, parts: list[dict[str, Any]]) -> dict[str, Any]:
        return chapter_from_items(chapter_id, title, parts)

    def fold_image_chapters_into_text(input_chapters: list[dict[str, Any]]) -> list[dict[str, Any]]:
        folded: list[dict[str, Any]] = []
        last_text_chapter: dict[str, Any] | None = None
        for chapter in input_chapters:
            if chapter.get("kind") != "image":
                folded.append(chapter)
                last_text_chapter = chapter if chapter.get("kind") == "text" else last_text_chapter
                continue
            if last_text_chapter is None:
                folded.append(chapter)
                continue
            image_items = chapter.get("items") or [chapter]
            last_text_chapter.setdefault("items", []).extend(image_items)
            last_text_chapter.setdefault("images", []).extend(
                image
                for item in image_items
                for image in item.get("images", [])
            )
        return folded

    def flatten_node_files(node: dict[str, Any]) -> list[dict[str, str]]:
        items: list[dict[str, str]] = []
        href = node.get("href", "")
        if href in rows_by_file or href in image_by_file:
            items.append({"href": href, "title": node.get("title") or ""})
        for child in node.get("children", []):
            items.extend(flatten_node_files(child))
        return items

    chapters: list[dict[str, Any]] = []
    assigned_files: set[str] = set()
    chapter_serial = 0

    def append_file_as_chapter(entry_name: str, title_override: str = "") -> None:
        nonlocal chapter_serial
        if entry_name in image_by_file:
            images = image_by_file.get(entry_name, [])
            for image_index, image in enumerate(images, start=1):
                chapter_serial += 1
                image_item = make_image_item(entry_name, image, title_override, image_index, len(images))
                chapters.append(image_chapter_from_item(f"C{chapter_serial:04d}", image_item))
            return
        if entry_name in rows_by_file:
            chapter_serial += 1
            part = make_part(entry_name, title_override)
            chapters.append(chapter_from_parts(f"C{chapter_serial:04d}", title_override or part["source_title"] or part["title"], [part]))

    def append_grouped_files(title: str, files: list[str]) -> None:
        nonlocal chapter_serial
        chapter_items: list[dict[str, Any]] = []
        text_title = title

        def flush_chapter_items() -> None:
            nonlocal chapter_serial, chapter_items, text_title
            if not chapter_items:
                return
            text_parts = [item for item in chapter_items if item.get("kind") == "text"]
            if not text_parts:
                for item in chapter_items:
                    if item.get("kind") == "image":
                        chapter_serial += 1
                        chapters.append(image_chapter_from_item(f"C{chapter_serial:04d}", item))
                chapter_items = []
                text_title = ""
                return
            chapter_serial += 1
            chapters.append(chapter_from_items(f"C{chapter_serial:04d}", text_title or text_parts[0]["source_title"] or text_parts[0]["title"], chapter_items))
            chapter_items = []
            text_title = ""

        for file_index, entry_name in enumerate(files):
            if entry_name in image_by_file:
                images = image_by_file.get(entry_name, [])
                for image_index, image in enumerate(images, start=1):
                    image_title = title if file_index == 0 and not chapter_items else ""
                    chapter_items.append(make_image_item(entry_name, image, image_title, image_index, len(images)))
            elif entry_name in rows_by_file:
                chapter_items.append(make_part(entry_name))
        flush_chapter_items()

    for spec in chapter_specs:
        node_files = [entry_name for entry_name in spec.get("files", []) if entry_name not in assigned_files]
        if not node_files:
            continue
        if len(node_files) == 1 and node_files[0] in image_by_file:
            append_file_as_chapter(node_files[0], spec.get("title") or "")
        else:
            append_grouped_files(spec.get("title") or "", node_files)
        assigned_files.update(node_files)

    chapter_nodes = nav_tree
    while (
        len(chapter_nodes) == 1
        and chapter_nodes[0].get("href") not in rows_by_file
        and chapter_nodes[0].get("children")
    ):
        chapter_nodes = chapter_nodes[0]["children"]

    if not chapters:
        for node in chapter_nodes:
            node_files = []
            seen_in_node: set[str] = set()
            for item in flatten_node_files(node):
                href = item["href"]
                if href in assigned_files or href in seen_in_node:
                    continue
                seen_in_node.add(href)
                node_files.append(item)
            if not node_files:
                continue
            append_grouped_files(node.get("title") or "", [item["href"] for item in node_files])
            assigned_files.update(item["href"] for item in node_files)

    for entry_name in content_files:
        if entry_name in assigned_files:
            continue
        append_file_as_chapter(entry_name)

    chapters.sort(key=lambda chapter: chapter.get("_sort", 10**9))
    chapters = fold_image_chapters_into_text(chapters)
    for index, chapter in enumerate(chapters, start=1):
        chapter["id"] = f"C{index:04d}"
        chapter.pop("_sort", None)
        if chapter.get("kind") == "text":
            for item_index, item in enumerate(chapter.get("items", []), start=1):
                item["parent_chapter_id"] = chapter["id"]
                item["display_level"] = 1
                item["toc_title"] = item.get("title") or item.get("file_label") or item.get("id") or f"item{item_index:04d}"
        else:
            chapter.setdefault("display_level", 0)

    return {
        "chapters": chapters,
        "chapter_count": len(chapters),
        "text_chapter_count": sum(1 for chapter in chapters if chapter.get("kind") != "image"),
        "image_chapter_count": image_item_count(chapters) or sum(1 for chapter in chapters if chapter.get("kind") == "image"),
        "top_level_image_chapter_count": sum(1 for chapter in chapters if chapter.get("kind") == "image"),
        "part_count": sum(len(chapter.get("parts", [])) for chapter in chapters),
        "image_count": sum(
            len(chapter.get("images", [])) if chapter.get("kind") == "image" else sum(
                len(item.get("images", []))
                for item in chapter.get("items", [])
                if item.get("kind") == "image"
            )
            for chapter in chapters
        ),
        "asset_count": len(collect_asset_paths_from_structure({"chapters": chapters})),
        "row_count": len(rows),
    }


def build_rows(session_root: Path) -> list[dict[str, Any]]:
    html_entries = html_entries_in_reading_order(session_root)
    rows: list[dict[str, Any]] = []
    serial = 0
    for file_order, entry_name in enumerate(html_entries, start=1):
        path = extracted_path(session_root, entry_name)
        if not path.exists():
            continue
        document_title = read_document_title(session_root, entry_name) or Path(entry_name).stem
        text = path.read_text(encoding="utf-8", errors="replace")
        matches = list(P_RE.finditer(text))
        i = 0
        file_serial = 0
        while i < len(matches) - 1:
            open_tag = matches[i].group(1)
            attrs = open_tag[2:-1]
            jp_inner = matches[i].group(2)
            cn_inner = matches[i + 1].group(2)
            next_open = matches[i + 1].group(1)
            next_attrs = next_open[2:-1]
            if is_gray_source(attrs, jp_inner) and not is_gray_source(next_attrs, cn_inner):
                serial += 1
                file_serial += 1
                rows.append(
                    {
                        "id": f"R{serial:05d}",
                        "file": entry_name,
                        "file_order": file_order,
                        "file_row_index": file_serial,
                        "file_label": Path(entry_name).name,
                        "document_title": document_title,
                        "ja_p_index": i,
                        "cn_p_index": i + 1,
                        "jp_html": jp_inner,
                        "jp_text": strip_tags(jp_inner),
                        "cn_html": cn_inner,
                        "cn_text": strip_tags(cn_inner),
                    }
                )
                i += 2
            else:
                i += 1
    write_json(session_root / "review_state" / "rows.json", rows)
    return rows


def ensure_state(session_root: Path, epub_path: Path, reset: bool = False) -> dict[str, Any]:
    extract_epub(epub_path, session_root, reset=reset)
    rows_path = session_root / "review_state" / "rows.json"
    rows = read_json(rows_path, None)
    if rows is None:
        rows = build_rows(session_root)
    state_path = session_root / "review_state" / "state.json"
    state = read_json(
        state_path,
        {
            "version": STATE_VERSION,
            "created_at": now_iso(),
            "updated_at": now_iso(),
            "source_epub": str(epub_path),
            "edits": {},
        },
    )
    state.setdefault("edits", {})
    state["source_epub"] = str(epub_path)
    write_json(state_path, state)
    write_session_manifest(session_root, epub_path)
    return state


def merge_rows(session_root: Path) -> list[dict[str, Any]]:
    rows = read_json(session_root / "review_state" / "rows.json", [])
    state = read_json(session_root / "review_state" / "state.json", {"edits": {}})
    edits = state.get("edits", {})
    merged = []
    for row in rows:
        item = dict(row)
        edit = edits.get(row["id"], {})
        item["current_html"] = edit.get("current_html", row["cn_html"])
        item["marked"] = bool(edit.get("marked", False))
        item["issue_type"] = edit.get("issue_type", "")
        item["severity"] = edit.get("severity", "")
        item["comment"] = edit.get("comment", "")
        item["learn_note"] = edit.get("learn_note", "")
        item["tags"] = edit.get("tags", "")
        item["edited"] = item["current_html"] != row["cn_html"]
        item["updated_at"] = edit.get("updated_at", "")
        merged.append(item)
    return merged


def write_feedback_summary(session_root: Path) -> Path:
    rows = merge_rows(session_root)
    marked = [
        row
        for row in rows
        if row["edited"]
        or row["marked"]
        or row["comment"]
        or row["learn_note"]
        or row["issue_type"]
        or row["severity"]
        or row["tags"]
    ]
    out = session_root / "feedback" / "feedback_for_codex.md"
    lines = [
        "# 双语 EPUB 审校反馈",
        "",
        f"- 更新时间：{now_iso()}",
        f"- 有记录段落：{len(marked)}",
        "",
        "这些记录用于之后优化翻译、润色、术语和角色口吻。`修改前` 来自初始双语 EPUB，`修改后` 来自网页编辑器当前保存内容。",
        "",
    ]
    for row in marked:
        lines.extend(
            [
                f"## {row['id']} · {row['file_label']} · P{row['cn_p_index']}",
                "",
                f"- 文件：`{row['file']}`",
                f"- 问题类型：{row['issue_type'] or '未填写'}",
                f"- 严重程度：{row['severity'] or '未填写'}",
                f"- 标签：{row['tags'] or '无'}",
                f"- 备注：{row['comment'] or '无'}",
                f"- 可沉淀规则：{row['learn_note'] or '无'}",
                "",
                "日文原文：",
                "",
                row["jp_text"],
                "",
                "修改前：",
                "",
                strip_tags(row["cn_html"]),
                "",
                "修改后：",
                "",
                strip_tags(row["current_html"]),
                "",
            ]
        )
    out.parent.mkdir(parents=True, exist_ok=True)
    out.write_text("\n".join(lines), encoding="utf-8")
    return out


def apply_edits_to_xhtml(session_root: Path) -> list[str]:
    rows = merge_rows(session_root)
    edited_rows = [row for row in rows if row["edited"]]
    by_file: dict[str, list[dict[str, Any]]] = {}
    for row in edited_rows:
        by_file.setdefault(row["file"], []).append(row)
    modified = []
    for entry_name, items in by_file.items():
        path = extracted_path(session_root, entry_name)
        text = path.read_text(encoding="utf-8", errors="replace")
        matches = list(P_RE.finditer(text))
        replace_by_index = {int(row["cn_p_index"]): row["current_html"] for row in items}
        parts: list[str] = []
        last = 0
        for index, match in enumerate(matches):
            parts.append(text[last:match.start()])
            if index in replace_by_index:
                parts.append(match.group(1))
                parts.append(replace_by_index[index])
                parts.append(match.group(3))
            else:
                parts.append(match.group(0))
            last = match.end()
        parts.append(text[last:])
        path.write_text("".join(parts), encoding="utf-8")
        modified.append(entry_name)
    write_feedback_summary(session_root)
    return modified


def export_epub(session_root: Path) -> Path:
    apply_edits_to_xhtml(session_root)
    entries = read_json(session_root / "review_state" / "epub_entries.json", [])
    manifest = read_json(session_manifest_path(session_root), {})
    state_source = read_json(session_root / "review_state" / "state.json", {}).get("source_epub", "reviewed.epub")
    source_name = manifest.get("source_name") or Path(state_source).name
    ts = dt.datetime.now().strftime("%Y%m%d_%H%M%S")
    out_dir = session_root / "exports"
    out_dir.mkdir(parents=True, exist_ok=True)
    out_path = out_dir / f"{safe_slug(Path(source_name).stem)}_网页审校版_{ts}.epub"
    with zipfile.ZipFile(out_path, "w") as zf:
        names_written: set[str] = set()
        mimetype = extracted_path(session_root, "mimetype")
        if mimetype.exists():
            zf.write(mimetype, "mimetype", compress_type=zipfile.ZIP_STORED)
            names_written.add("mimetype")
        for entry in entries:
            name = entry["filename"]
            if name in names_written:
                continue
            path = extracted_path(session_root, name)
            if not path.exists():
                continue
            compress = zipfile.ZIP_STORED if name == "mimetype" else zipfile.ZIP_DEFLATED
            if entry.get("is_dir"):
                info = zipfile.ZipInfo(name)
                zf.writestr(info, b"")
            else:
                zf.write(path, name, compress_type=compress)
            names_written.add(name)
    append_jsonl(
        session_root / "feedback" / "translation_feedback.jsonl",
        {"ts": now_iso(), "event": "export_epub", "output": str(out_path)},
    )
    return out_path


def translation_jobs_root(review_root: Path) -> Path:
    return review_root / TRANSLATION_JOB_DIR_NAME


def translation_job_root(review_root: Path, job_id: str) -> Path:
    safe_id = safe_slug(job_id, max_len=160)
    if safe_id != job_id:
        raise ValueError("invalid translation job id")
    root = (translation_jobs_root(review_root) / job_id).resolve()
    try:
        root.relative_to(translation_jobs_root(review_root).resolve())
    except ValueError as exc:
        raise ValueError("invalid translation job id") from exc
    return root


def translation_job_path(review_root: Path, job_id: str) -> Path:
    return translation_job_root(review_root, job_id) / "job.json"


def make_translation_job_id(session_id: str) -> str:
    seed = f"{session_id}-{now_iso()}-{time.time_ns()}"
    digest = hashlib.sha1(seed.encode("utf-8")).hexdigest()[:8]
    return f"tr_{dt.datetime.now().strftime('%Y%m%d_%H%M%S')}_{digest}"


def read_translation_job(review_root: Path, job_id: str) -> dict[str, Any]:
    return read_json(translation_job_path(review_root, job_id), {})


def write_translation_job(review_root: Path, job: dict[str, Any]) -> None:
    job["updated_at"] = now_iso()
    write_json_atomic(translation_job_path(review_root, str(job["id"])), job)


def add_translation_job_log(job: dict[str, Any], message: str, level: str = "info") -> None:
    logs = job.setdefault("logs", [])
    logs.append({"ts": now_iso(), "level": level, "message": str(message)[:1200]})
    if len(logs) > TRANSLATION_LOG_LIMIT:
        del logs[: len(logs) - TRANSLATION_LOG_LIMIT]


def public_translation_job(job: dict[str, Any]) -> dict[str, Any]:
    settings = job.get("settings") if isinstance(job.get("settings"), dict) else {}
    return {
        "id": job.get("id", ""),
        "status": job.get("status", ""),
        "source_session_id": job.get("source_session_id", ""),
        "source_name": job.get("source_name", ""),
        "created_at": job.get("created_at", ""),
        "updated_at": job.get("updated_at", ""),
        "started_at": job.get("started_at", ""),
        "finished_at": job.get("finished_at", ""),
        "progress": job.get("progress", {}),
        "error": job.get("error", ""),
        "logs": job.get("logs", []),
        "output_epub": job.get("output_epub", ""),
        "output_session_id": job.get("output_session_id", ""),
        "settings_summary": {
            "output_mode": settings.get("output_mode", "bilingual"),
            "range_mode": settings.get("range_mode", "limit"),
            "limit": settings.get("limit", DEFAULT_TRANSLATION_LIMIT),
            "glossary_path": settings.get("glossary_path", ""),
            "temperature": settings.get("temperature", 0.2),
        },
    }


def normalize_translation_job_settings(data: dict[str, Any], config: dict[str, Any]) -> dict[str, Any]:
    output_mode = str(data.get("output_mode") or "bilingual").strip().lower()
    if output_mode not in {"bilingual", "translated"}:
        raise ValueError("输出模式必须是 bilingual 或 translated")
    range_mode = str(data.get("range_mode") or "limit").strip().lower()
    if range_mode not in {"all", "limit"}:
        raise ValueError("翻译范围必须是 all 或 limit")
    try:
        limit = int(data.get("limit") or DEFAULT_TRANSLATION_LIMIT)
    except (TypeError, ValueError) as exc:
        raise ValueError("试译段数必须是数字") from exc
    limit = max(1, min(limit, TRANSLATION_RANGE_LIMIT_MAX))
    try:
        temperature = float(data.get("temperature", 0.2))
    except (TypeError, ValueError) as exc:
        raise ValueError("temperature 必须是数字") from exc
    temperature = max(0.0, min(1.0, temperature))
    glossary_value = str(data.get("glossary_path") or config.get("glossary_path") or "").strip()
    return {
        "output_mode": output_mode,
        "range_mode": range_mode,
        "limit": limit,
        "temperature": temperature,
        "glossary_path": normalize_glossary_path_value(glossary_value),
        "translation_prompt": normalize_gpt_prompt("translation_prompt", data.get("translation_prompt") or config.get("translation_prompt")),
        "format_prompt": normalize_gpt_prompt("format_prompt", data.get("format_prompt") or config.get("format_prompt")),
        "character_prompt": normalize_gpt_prompt("character_prompt", data.get("character_prompt") or config.get("character_prompt")),
    }


def html_looks_like_navigation(entry_name: str, text: str) -> bool:
    basename = posixpath.basename(entry_name).lower()
    if basename in {"nav.xhtml", "toc.xhtml", "toc.html", "toc.htm"} or "toc" in basename:
        return True
    heading = extract_heading_from_html(text)
    if re.search(r"(目录|目次|contents?)", heading, flags=re.I):
        return True
    matches = list(P_RE.finditer(text))
    if len(matches) < 4:
        return False
    link_count = sum(1 for match in matches if re.search(r"<a\b[^>]*href=", match.group(2), flags=re.I))
    return link_count / max(1, len(matches)) >= 0.45


def extract_translation_source_items(session_root: Path) -> list[dict[str, Any]]:
    items: list[dict[str, Any]] = []
    serial = 0
    for file_order, entry_name in enumerate(html_entries_in_reading_order(session_root), start=1):
        path = extracted_path(session_root, entry_name)
        if not path.exists():
            continue
        text = path.read_text(encoding="utf-8", errors="replace")
        if html_looks_like_navigation(entry_name, text):
            continue
        document_title = read_document_title(session_root, entry_name) or Path(entry_name).stem
        file_serial = 0
        for p_index, match in enumerate(P_RE.finditer(text)):
            inner = match.group(2)
            source_text = strip_tags(inner)
            if not source_text:
                continue
            if IMG_TAG_RE.search(inner) and not source_text:
                continue
            if not KANA_RE.search(strip_html_keep_ruby_text(inner)):
                continue
            serial += 1
            file_serial += 1
            items.append(
                {
                    "id": f"T{serial:05d}",
                    "file": entry_name,
                    "file_order": file_order,
                    "file_row_index": file_serial,
                    "file_label": Path(entry_name).name,
                    "document_title": document_title,
                    "p_index": p_index,
                    "p_open": match.group(1),
                    "p_close": match.group(3),
                    "source_html": inner,
                    "source_text": source_text,
                }
            )
    return items


def translation_context_lines(items: list[dict[str, Any]], index: int, radius: int = TRANSLATION_CONTEXT_RADIUS) -> list[str]:
    start = max(0, index - radius)
    end = min(len(items), index + radius + 1)
    context: list[str] = []
    for idx in range(start, end):
        if idx == index:
            continue
        item = items[idx]
        direction = "前文" if idx < index else "后文"
        context.append(f"{direction} {item.get('id', '')}: {truncate_prompt_text(item.get('source_text'), 260)}")
    return context


def build_translation_messages(
    item: dict[str, Any],
    items: list[dict[str, Any]],
    index: int,
    settings: dict[str, Any],
    glossary_lines: list[str],
) -> list[dict[str, str]]:
    glossary_text = "\n".join(f"- {line}" for line in glossary_lines)
    context_lines = translation_context_lines(items, index)
    system_prompt = "\n\n".join(
        [
            "你是日语轻小说到简体中文的整书翻译助手。每次只处理一个源文段落。",
            "只输出该段落的简体中文译文 HTML 片段，不解释，不输出 Markdown。",
            "不得合并、拆分、补写或删除源文信息；意义优先，再改写成自然中文轻小说文风。",
        ]
    )
    user_parts = [
        "【翻译内容提示词】",
        str(settings.get("translation_prompt") or DEFAULT_TRANSLATION_PROMPT).strip(),
        "",
        "【目标段落：日文原文 HTML】",
        f"<p>{sanitize_source_html_for_prompt(item.get('source_html', ''))}</p>",
    ]
    if glossary_text:
        user_parts.extend(["", "【本段命中术语】", glossary_text])
    if context_lines:
        user_parts.extend(["", "【少量上下文】", "\n".join(f"- {line}" for line in context_lines)])
    user_parts.extend(
        [
            "",
            "【角色状态与口吻提示词】",
            str(settings.get("character_prompt") or DEFAULT_CHARACTER_PROMPT).strip(),
            "",
            "【格式与标点提示词】",
            str(settings.get("format_prompt") or DEFAULT_FORMAT_PROMPT).strip(),
            "",
            "请只输出这个段落的中文译文 HTML 片段，不要包含外层 <p>。",
        ]
    )
    return [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": "\n".join(user_parts)},
    ]


def add_gray_style_to_p(open_tag: str) -> str:
    def replace_style(match: re.Match[str]) -> str:
        quote = match.group(1)
        value = match.group(2).strip()
        if re.search(r"\bcolor\s*:", value, flags=re.I):
            value = re.sub(r"\bcolor\s*:\s*[^;]+;?", "color: #808080;", value, count=1, flags=re.I)
            return f"style={quote}{value}{quote}"
        suffix = "" if not value or value.endswith(";") else ";"
        return f'style={quote}{value}{suffix} color: #808080;{quote}'

    if re.search(r"\sstyle\s*=", open_tag, flags=re.I):
        return re.sub(r"style\s*=\s*(['\"])(.*?)\1", replace_style, open_tag, count=1, flags=re.I | re.S)
    return re.sub(r">\s*$", ' style="color: #808080;">', open_tag)


def replace_translated_paragraphs(
    text: str,
    replacements: dict[int, str],
    output_mode: str,
) -> str:
    parts: list[str] = []
    last = 0
    for index, match in enumerate(P_RE.finditer(text)):
        parts.append(text[last:match.start()])
        if index not in replacements:
            parts.append(match.group(0))
        elif output_mode == "bilingual":
            source_open = add_gray_style_to_p(match.group(1))
            parts.append(source_open)
            parts.append(match.group(2))
            parts.append(match.group(3))
            parts.append("\n<p>")
            parts.append(replacements[index])
            parts.append("</p>")
        else:
            parts.append(match.group(1))
            parts.append(replacements[index])
            parts.append(match.group(3))
        last = match.end()
    parts.append(text[last:])
    return "".join(parts)


def build_translation_output_epub(
    session_root: Path,
    source_epub: Path,
    source_name: str,
    items: list[dict[str, Any]],
    translations: dict[str, str],
    output_mode: str,
    out_path: Path,
) -> Path:
    entries = read_json(session_root / "review_state" / "epub_entries.json", [])
    by_file: dict[str, dict[int, str]] = {}
    for item in items:
        translated = translations.get(item["id"])
        if translated is None:
            continue
        by_file.setdefault(item["file"], {})[int(item["p_index"])] = translated

    out_path.parent.mkdir(parents=True, exist_ok=True)
    with zipfile.ZipFile(out_path, "w") as zf:
        names_written: set[str] = set()
        mimetype = extracted_path(session_root, "mimetype")
        if mimetype.exists():
            zf.write(mimetype, "mimetype", compress_type=zipfile.ZIP_STORED)
            names_written.add("mimetype")
        for entry in entries:
            name = entry["filename"]
            if name in names_written:
                continue
            path = extracted_path(session_root, name)
            if not path.exists():
                continue
            compress = zipfile.ZIP_STORED if name == "mimetype" else zipfile.ZIP_DEFLATED
            if entry.get("is_dir"):
                info = zipfile.ZipInfo(name)
                zf.writestr(info, b"")
            elif name in by_file and name.lower().endswith((".xhtml", ".html", ".htm")):
                text = path.read_text(encoding="utf-8", errors="replace")
                updated = replace_translated_paragraphs(text, by_file[name], output_mode)
                zf.writestr(name, updated.encode("utf-8"), compress_type=compress)
            else:
                zf.write(path, name, compress_type=compress)
            names_written.add(name)
    return out_path


def translation_output_path(review_root: Path, source_name: str, output_mode: str) -> Path:
    suffix = "中日双语" if output_mode == "bilingual" else "中文"
    ts = dt.datetime.now().strftime("%Y%m%d_%H%M%S")
    stem = safe_slug(Path(source_name).stem or "translated_epub", max_len=100)
    return review_root / TRANSLATION_OUTPUT_DIR_NAME / f"{stem}_AI翻译_{suffix}_{ts}.epub"


def register_output_session(review_root: Path, out_path: Path) -> str:
    session_root = session_root_for(review_root, out_path)
    ensure_state(session_root, out_path, reset=True)
    write_session_manifest(session_root, out_path, source_name=out_path.name)
    return session_public_id(session_root)


def run_translation_job(review_root: Path, job_id: str) -> None:
    job = read_translation_job(review_root, job_id)
    if not job:
        return
    translations: dict[str, str] = {}
    selected_items: list[dict[str, Any]] = []
    try:
        job["status"] = "running"
        job["started_at"] = now_iso()
        job["error"] = ""
        job["progress"] = {
            "current": 0,
            "total": 0,
            "percent": 0,
            "failed": 0,
            "message": "正在准备源 EPUB……",
            "current_file": "",
        }
        add_translation_job_log(job, "翻译任务已启动。")
        write_translation_job(review_root, job)

        source_session = session_root_from_id(review_root, str(job.get("source_session_id") or ""))
        if not (source_session / "review_state" / "state.json").exists():
            raise ValueError("源书籍会话不存在或不完整")
        source_epub = session_epub_path(source_session)
        if not source_epub.exists():
            raise ValueError("源 EPUB 文件不存在")
        settings = job.get("settings") if isinstance(job.get("settings"), dict) else {}
        config = read_gpt_config(review_root)
        if not config.get("configured"):
            raise ValueError("尚未配置 GPT API Key")
        config = {
            **config,
            "base_url": str(settings.get("base_url") or config.get("base_url") or DEFAULT_GPT_BASE_URL),
            "model": str(settings.get("model") or config.get("model") or DEFAULT_GPT_MODEL),
        }
        glossary_file = Path(str(settings.get("glossary_path") or config.get("glossary_path") or ""))
        glossary = read_glossary_mapping_from_path(glossary_file)

        all_items = extract_translation_source_items(source_session)
        source_total = len(all_items)
        if settings.get("range_mode") == "all":
            selected_items = all_items
        else:
            selected_items = all_items[: int(settings.get("limit") or DEFAULT_TRANSLATION_LIMIT)]
        if not selected_items:
            raise ValueError("没有找到可翻译的日文正文段落")

        job["progress"].update(
            {
                "total": len(selected_items),
                "source_total": source_total,
                "message": f"已找到 {source_total} 个日文段落，开始翻译 {len(selected_items)} 段。",
            }
        )
        add_translation_job_log(job, job["progress"]["message"])
        write_translation_job(review_root, job)

        failed_count = 0
        errors: list[dict[str, Any]] = []
        for index, item in enumerate(selected_items):
            job["progress"].update(
                {
                    "current": index,
                    "percent": round(index / max(1, len(selected_items)) * 100, 1),
                    "message": f"正在翻译 {item['id']}（{index + 1}/{len(selected_items)}）",
                    "current_file": item.get("file", ""),
                    "current_item_id": item.get("id", ""),
                }
            )
            write_translation_job(review_root, job)
            glossary_lines = glossary_matches_for_source(glossary, item.get("source_html", ""), item.get("source_text", ""))
            try:
                raw = call_openai_compatible_chat(
                    config,
                    build_translation_messages(item, selected_items, index, settings, glossary_lines),
                    temperature=float(settings.get("temperature", 0.2)),
                    timeout=120,
                )
                candidate_html = sanitize_cn_html(
                    restore_glossary_ruby_candidates(normalize_candidate_punctuation(raw), glossary_lines)
                )
                if not candidate_html or not strip_tags(candidate_html):
                    raise RuntimeError("GPT 返回内容为空或无法作为译文保存")
                translations[item["id"]] = candidate_html
            except Exception as exc:
                failed_count += 1
                error_text = str(exc)[:900]
                errors.append({"item_id": item["id"], "file": item.get("file", ""), "p_index": item.get("p_index"), "error": error_text})
                translations[item["id"]] = f"【本段 AI 翻译失败：{html.escape(error_text)}】"
                add_translation_job_log(job, f"{item['id']} 翻译失败：{error_text}", level="warning")
            if index == 0 or (index + 1) % 5 == 0 or failed_count:
                write_json_atomic(
                    translation_job_root(review_root, job_id) / "translations.json",
                    {"items": selected_items, "translations": translations, "errors": errors},
                )

        job["progress"].update(
            {
                "current": len(selected_items),
                "percent": 100,
                "failed": failed_count,
                "message": "正在生成 EPUB……",
                "current_item_id": "",
            }
        )
        write_translation_job(review_root, job)
        source_name = str(job.get("source_name") or source_epub.name)
        out_path = translation_output_path(review_root, source_name, str(settings.get("output_mode") or "bilingual"))
        build_translation_output_epub(
            source_session,
            source_epub,
            source_name,
            selected_items,
            translations,
            str(settings.get("output_mode") or "bilingual"),
            out_path,
        )
        output_session_id = register_output_session(review_root, out_path)
        job["status"] = "completed"
        job["finished_at"] = now_iso()
        job["output_epub"] = str(out_path)
        job["output_session_id"] = output_session_id
        job["errors"] = errors
        job["progress"].update(
            {
                "message": f"翻译完成，已生成 EPUB 并加入书架。失败段落 {failed_count} 个。",
                "failed": failed_count,
            }
        )
        add_translation_job_log(job, job["progress"]["message"])
        write_translation_job(review_root, job)
        append_jsonl(
            source_session / "feedback" / "translation_feedback.jsonl",
            {
                "ts": now_iso(),
                "event": "translation_job_completed",
                "job_id": job_id,
                "output": str(out_path),
                "output_session_id": output_session_id,
                "selected_count": len(selected_items),
                "failed_count": failed_count,
                "glossary_path": str(settings.get("glossary_path") or ""),
            },
        )
    except Exception as exc:
        job = read_translation_job(review_root, job_id) or job
        job["status"] = "failed"
        job["finished_at"] = now_iso()
        job["error"] = str(exc)[:1200]
        job.setdefault("progress", {}).update({"message": f"翻译失败：{job['error']}"})
        add_translation_job_log(job, job["error"], level="error")
        write_translation_job(review_root, job)


def create_app(review_root: Path, initial_epub_path: Path | None = None, reset: bool = False) -> Flask:
    app = Flask(
        __name__,
        root_path=str(Path(__file__).resolve().parent),
        static_folder="static",
        static_url_path="/static",
    )
    review_root.mkdir(parents=True, exist_ok=True)
    translation_jobs_root(review_root).mkdir(parents=True, exist_ok=True)
    app.config["REVIEW_ROOT"] = review_root
    app.config["APP_VERSION"] = version
    app.config["SESSION_ROOT"] = None
    app.config["EPUB_PATH"] = None
    app.config["LOCK"] = threading.Lock()
    app.config["TRANSLATION_THREADS"] = {}

    def set_active_session(session_root: Path, epub_path: Path | None = None) -> None:
        app.config["SESSION_ROOT"] = session_root
        app.config["EPUB_PATH"] = epub_path or session_epub_path(session_root)
        update_app_manifest(review_root, session_root)

    def active_session() -> tuple[Path | None, Path | None]:
        session_root = app.config.get("SESSION_ROOT")
        epub_path = app.config.get("EPUB_PATH")
        if isinstance(session_root, Path):
            return session_root, epub_path if isinstance(epub_path, Path) else session_epub_path(session_root)
        active_id = read_json(app_manifest_path(review_root), {}).get("active_session_id", "")
        if active_id:
            try:
                manifest_root = session_root_from_id(review_root, active_id)
            except ValueError:
                return None, None
            if (manifest_root / "review_state" / "state.json").exists():
                set_active_session(manifest_root)
                return app.config["SESSION_ROOT"], app.config["EPUB_PATH"]
        return None, None

    def no_active_response():
        return jsonify(
            {
                "has_session": False,
                "version": version,
                "version_rules": VERSION_RULES,
                "error": "尚未打开 EPUB。请先上传中日双语 EPUB，或从已有会话中选择一个。",
                "review_root": str(review_root),
            }
        ), 409

    if initial_epub_path is not None:
        session_root = session_root_for(review_root, initial_epub_path)
        ensure_state(session_root, initial_epub_path, reset=reset)
        set_active_session(session_root, initial_epub_path)

    @app.route("/")
    def index():
        return send_from_directory(app.static_folder, "index.html")

    @app.route("/api/session")
    def api_session():
        session_root, epub_path = active_session()
        if session_root is None or epub_path is None:
            return jsonify(
                {
                    "has_session": False,
                    "version": version,
                    "version_rules": VERSION_RULES,
                    "review_root": str(review_root),
                    "sessions": list_review_sessions(review_root),
                }
            )
        rows = merge_rows(session_root)
        touched = [row for row in rows if rowTouched_py(row)]
        manifest = read_json(session_manifest_path(session_root), {})
        return jsonify(
            {
                "has_session": True,
                "version": version,
                "version_rules": VERSION_RULES,
                "id": session_public_id(session_root),
                "source_name": manifest.get("source_name") or epub_path.name,
                "source_epub": str(epub_path),
                "session_root": str(session_root),
                "row_count": len(rows),
                "touched_count": len(touched),
                "marked_count": sum(1 for row in rows if row.get("marked")),
                "feedback_md": str(session_root / "feedback" / "feedback_for_codex.md"),
                "feedback_jsonl": str(session_root / "feedback" / "translation_feedback.jsonl"),
                "reading_position": read_reading_position(session_root),
            }
        )

    @app.route("/api/version")
    def api_version():
        return jsonify({"version": version, "version_rules": VERSION_RULES})

    @app.route("/api/sessions")
    def api_sessions():
        session_root, _epub_path = active_session()
        return jsonify(
            {
                "review_root": str(review_root),
                "version": version,
                "version_rules": VERSION_RULES,
                "active_session_id": session_public_id(session_root) if session_root else "",
                "sessions": list_review_sessions(review_root),
            }
        )

    @app.route("/api/upload", methods=["POST"])
    def api_upload():
        file_storage = request.files.get("epub")
        if file_storage is None or not file_storage.filename:
            return jsonify({"error": "请选择一个 .epub 文件"}), 400
        reset_upload = str(request.form.get("reset", "")).lower() in {"1", "true", "yes", "on"}
        try:
            epub_path = copy_upload_to_review_root(file_storage, review_root)
            session_root = session_root_for(review_root, epub_path)
            with app.config["LOCK"]:
                if reset_upload and session_root.exists():
                    shutil.rmtree(session_root)
                ensure_state(session_root, epub_path, reset=reset_upload)
                write_session_manifest(session_root, epub_path, source_name=file_storage.filename)
                set_active_session(session_root, epub_path)
            summary = summarize_session(session_root)
            summary["has_session"] = True
            return jsonify(summary)
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        except zipfile.BadZipFile:
            return jsonify({"error": "文件不是有效的 EPUB/ZIP 包"}), 400
        except Exception as exc:
            return jsonify({"error": f"EPUB 解析失败：{exc}"}), 500

    @app.route("/api/session/open", methods=["POST"])
    def api_open_session():
        data = request.get_json(silent=True) or {}
        session_id = str(data.get("session_id", ""))
        if not session_id:
            return jsonify({"error": "缺少 session_id"}), 400
        try:
            session_root = session_root_from_id(review_root, session_id)
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        if not (session_root / "review_state" / "state.json").exists():
            return jsonify({"error": "会话不存在或不完整"}), 404
        with app.config["LOCK"]:
            set_active_session(session_root)
        summary = summarize_session(session_root)
        summary["has_session"] = True
        return jsonify(summary)

    @app.route("/api/rows")
    def api_rows():
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        return jsonify({"rows": merge_rows(session_root)})

    @app.route("/api/structure")
    def api_structure():
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        return jsonify(build_structure(session_root))

    @app.route("/api/reading-position", methods=["GET", "POST"])
    def api_reading_position():
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        if request.method == "GET":
            return jsonify({"reading_position": read_reading_position(session_root)})
        data = request.get_json(silent=True) or {}
        if not data and request.data:
            try:
                data = json.loads(request.data.decode("utf-8", errors="replace"))
            except json.JSONDecodeError:
                data = {}
        position = normalize_reading_position(data)
        with app.config["LOCK"]:
            state_path = session_root / "review_state" / "state.json"
            state = read_json(state_path, {"edits": {}})
            state["reading_position"] = position
            state["updated_at"] = now_iso()
            write_json(state_path, state)
        return jsonify({"status": "ok", "reading_position": position})

    @app.route("/api/asset/<path:entry_name>")
    def api_asset(entry_name: str):
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        entry_name = urllib.parse.unquote(entry_name)
        structure = build_structure(session_root)
        if entry_name not in collect_asset_paths_from_structure(structure):
            return jsonify({"error": "asset not allowed"}), 404
        if not is_raster_image_asset(entry_name):
            return jsonify({"error": "asset type not allowed"}), 415
        try:
            path = extracted_path(session_root, entry_name)
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        if not path.exists() or not path.is_file():
            return jsonify({"error": "asset not found"}), 404
        response = make_response(send_from_directory(path.parent, path.name))
        response.headers["X-Content-Type-Options"] = "nosniff"
        response.headers["Content-Security-Policy"] = "default-src 'none'; img-src 'self' data:; style-src 'none'; script-src 'none'; object-src 'none'"
        return response

    @app.route("/api/gpt/config", methods=["GET", "POST"])
    def api_gpt_config():
        if request.method == "GET":
            return jsonify(public_gpt_config(review_root))
        data = request.get_json(silent=True) or {}
        try:
            if data.get("reset_prompts"):
                current = read_gpt_config(review_root)
                data = {
                    **data,
                    "base_url": data.get("base_url") or current["base_url"],
                    "model": data.get("model") or current["model"],
                    "keep_existing": data.get("keep_existing", True),
                    **gpt_prompt_defaults(),
                }
            return jsonify(save_gpt_config(review_root, data))
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400

    @app.route("/api/glossary", methods=["GET", "POST"])
    def api_glossary():
        if request.method == "GET":
            try:
                return jsonify(public_glossary(review_root))
            except ValueError as exc:
                return jsonify({"error": str(exc)}), 400
        data = request.get_json(silent=True) or {}
        try:
            requested_path = str(data.get("path") or "").strip()
            if requested_path:
                update_gpt_config_fields(review_root, {"glossary_path": normalize_glossary_path_value(requested_path)})
            if "entries" not in data:
                return jsonify(public_glossary(review_root))
            entries = data.get("entries", [])
            if not isinstance(entries, list):
                return jsonify({"error": "entries 必须是数组"}), 400
            return jsonify(save_glossary(review_root, entries))
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400

    @app.route("/api/glossary/entry", methods=["POST", "DELETE"])
    def api_glossary_entry():
        data = request.get_json(silent=True) or {}
        source = str(data.get("source") or "").strip()
        if not source:
            return jsonify({"error": "缺少术语原文"}), 400
        try:
            glossary = read_glossary(review_root)
            if request.method == "DELETE":
                if source not in glossary:
                    return jsonify({"error": "术语不存在"}), 404
                del glossary[source]
            else:
                target = str(data.get("target") or "").strip()
                if not target:
                    return jsonify({"error": "缺少术语译名"}), 400
                old_source = str(data.get("old_source") or source).strip()
                if old_source and old_source != source:
                    glossary.pop(old_source, None)
                glossary[source] = target
            return jsonify(save_glossary(review_root, [{"source": key, "target": value} for key, value in glossary.items()]))
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400

    @app.route("/api/glossary/search")
    def api_glossary_search():
        term = str(request.args.get("q") or "")
        try:
            matches = search_glossary_entries(read_glossary(review_root), term)
            return jsonify({"query": term, "count": len(matches), "matches": matches})
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400

    @app.route("/api/translation/defaults")
    def api_translation_defaults():
        config = public_gpt_config(review_root)
        return jsonify(
            {
                "version": version,
                "gpt": config,
                "defaults": {
                    "output_mode": "bilingual",
                    "range_mode": "limit",
                    "limit": DEFAULT_TRANSLATION_LIMIT,
                    "temperature": 0.2,
                    "translation_prompt": config.get("translation_prompt") or DEFAULT_TRANSLATION_PROMPT,
                    "format_prompt": config.get("format_prompt") or DEFAULT_FORMAT_PROMPT,
                    "character_prompt": config.get("character_prompt") or DEFAULT_CHARACTER_PROMPT,
                    "glossary_path": config.get("glossary_path", ""),
                },
                "prompt_defaults": gpt_prompt_defaults(),
            }
        )

    @app.route("/api/session/<session_id>/translation-source")
    def api_translation_source(session_id: str):
        try:
            source_root = session_root_from_id(review_root, session_id)
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        if not (source_root / "review_state" / "state.json").exists():
            return jsonify({"error": "会话不存在或不完整"}), 404
        try:
            items = extract_translation_source_items(source_root)
            files: dict[str, int] = {}
            for item in items:
                files[item["file"]] = files.get(item["file"], 0) + 1
            return jsonify(
                {
                    "session": summarize_session(source_root),
                    "source_count": len(items),
                    "files": [{"file": key, "count": value} for key, value in files.items()],
                    "sample": [
                        {
                            "id": item["id"],
                            "file": item["file"],
                            "p_index": item["p_index"],
                            "text": truncate_prompt_text(item["source_text"], 140),
                        }
                        for item in items[:5]
                    ],
                }
            )
        except Exception as exc:
            return jsonify({"error": f"读取翻译源失败：{exc}"}), 500

    @app.route("/api/translation/start", methods=["POST"])
    def api_translation_start():
        data = request.get_json(silent=True) or {}
        session_id = str(data.get("session_id") or "").strip()
        if not session_id:
            return jsonify({"error": "缺少 session_id"}), 400
        try:
            source_root = session_root_from_id(review_root, session_id)
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        if not (source_root / "review_state" / "state.json").exists():
            return jsonify({"error": "会话不存在或不完整"}), 404
        try:
            requested_base_url = str(data.get("base_url") or "").strip()
            requested_model = str(data.get("model") or "").strip()
            config = read_gpt_config(review_root)
            if requested_base_url and normalize_gpt_base_url(requested_base_url) != config["base_url"]:
                return jsonify({"error": "整书翻译不能临时覆盖 Base URL。请先保存 API 设置。"}), 400
            if requested_model and requested_model != config["model"]:
                return jsonify({"error": "整书翻译不能临时覆盖模型。请先保存 API 设置。"}), 400
            if not config.get("configured"):
                return jsonify({"error": "尚未配置 GPT API Key。请先保存 API 设置。"}), 400
            settings = normalize_translation_job_settings(data, config)
            manifest = read_json(session_manifest_path(source_root), {})
            source_epub = session_epub_path(source_root)
            job_id = make_translation_job_id(session_id)
            job = {
                "id": job_id,
                "status": "pending",
                "source_session_id": session_id,
                "source_epub": str(source_epub),
                "source_name": manifest.get("source_name") or source_epub.name,
                "created_at": now_iso(),
                "updated_at": now_iso(),
                "started_at": "",
                "finished_at": "",
                "progress": {
                    "current": 0,
                    "total": 0,
                    "percent": 0,
                    "failed": 0,
                    "message": "任务已创建，等待后台启动。",
                    "current_file": "",
                    "current_item_id": "",
                },
                "settings": {
                    **settings,
                    "model": config["model"],
                    "base_url": config["base_url"],
                },
                "output_epub": "",
                "output_session_id": "",
                "error": "",
                "logs": [],
            }
            add_translation_job_log(job, "任务已创建。")
            write_translation_job(review_root, job)
            thread = threading.Thread(target=run_translation_job, args=(review_root, job_id), daemon=True)
            app.config["TRANSLATION_THREADS"][job_id] = thread
            thread.start()
            return jsonify({"status": "ok", "job_id": job_id, "job": public_translation_job(job)})
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        except Exception as exc:
            return jsonify({"error": f"创建翻译任务失败：{exc}"}), 500

    @app.route("/api/translation/jobs/<job_id>")
    def api_translation_job(job_id: str):
        try:
            job = read_translation_job(review_root, job_id)
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        if not job:
            return jsonify({"error": "翻译任务不存在"}), 404
        return jsonify({"job": public_translation_job(job)})

    @app.route("/api/row/<row_id>/retranslate", methods=["POST"])
    def api_retranslate_row(row_id: str):
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        rows = merge_rows(session_root)
        row = next((item for item in rows if item["id"] == row_id), None)
        if row is None:
            return jsonify({"error": "row not found"}), 404
        data = request.get_json(silent=True) or {}
        try:
            config = read_gpt_config(review_root)
            requested_base_url = str(data.get("base_url") or "").strip()
            requested_model = str(data.get("model") or "").strip()
            if requested_base_url and normalize_gpt_base_url(requested_base_url) != config["base_url"]:
                return jsonify({"error": "重翻请求不能临时覆盖 Base URL。请先保存 API 设置，避免把已保存密钥发往未确认端点。"}), 400
            if requested_model and requested_model != config["model"]:
                return jsonify({"error": "重翻请求不能临时覆盖模型。请先保存 API 设置后再重翻。"}), 400
            instruction = str(data.get("instruction") or "")
            glossary_lines = glossary_matches_for_prompt(review_root, row)
            raw = call_openai_compatible_chat(config, build_retranslate_messages(row, rows, config, glossary_lines, instruction))
            candidate_html = sanitize_cn_html(restore_glossary_ruby_candidates(normalize_candidate_punctuation(raw), glossary_lines))
            if not candidate_html:
                return jsonify({"error": "GPT 返回内容为空或无法作为译文保存"}), 502
            candidate_text = strip_tags(candidate_html)
            record = {
                "ts": now_iso(),
                "event": "row_retranslated",
                "row_id": row_id,
                "file": row["file"],
                "model": config["model"],
                "base_url": config["base_url"],
                "prompt_version": version,
                "translation_prompt": config.get("translation_prompt", ""),
                "format_prompt": config.get("format_prompt", ""),
                "character_prompt": config.get("character_prompt", ""),
                "glossary_path": str(glossary_path(review_root)),
                "glossary_matches": glossary_lines,
                "instruction": instruction[:1200],
                "jp_text": row["jp_text"],
                "before_cn_text": strip_tags(row.get("current_html") or row.get("cn_html") or ""),
                "candidate_cn_text": candidate_text,
            }
            append_jsonl(session_root / "feedback" / "translation_feedback.jsonl", record)
            return jsonify(
                {
                    "status": "ok",
                    "row_id": row_id,
                    "candidate_html": candidate_html,
                    "candidate_text": candidate_text,
                    "model": config["model"],
                    "glossary_path": str(glossary_path(review_root)),
                    "glossary_matches": glossary_lines,
                    "updated_at": record["ts"],
                }
            )
        except ValueError as exc:
            return jsonify({"error": str(exc)}), 400
        except RuntimeError as exc:
            return jsonify({"error": str(exc)}), 502

    @app.route("/api/row/<row_id>", methods=["POST"])
    def api_save_row(row_id: str):
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        data = request.get_json(silent=True) or {}
        rows = read_json(session_root / "review_state" / "rows.json", [])
        base = next((row for row in rows if row["id"] == row_id), None)
        if base is None:
            return jsonify({"error": "row not found"}), 404
        current_html = sanitize_cn_html(str(data.get("current_html", "")))
        edit = {
            "current_html": current_html,
            "marked": bool(data.get("marked", False)),
            "issue_type": str(data.get("issue_type", ""))[:80],
            "severity": str(data.get("severity", ""))[:40],
            "comment": str(data.get("comment", ""))[:4000],
            "learn_note": str(data.get("learn_note", ""))[:4000],
            "tags": str(data.get("tags", ""))[:500],
            "updated_at": now_iso(),
        }
        with app.config["LOCK"]:
            state_path = session_root / "review_state" / "state.json"
            state = read_json(state_path, {"edits": {}})
            state.setdefault("edits", {})[row_id] = edit
            state["updated_at"] = now_iso()
            write_json(state_path, state)
            append_jsonl(
                session_root / "feedback" / "translation_feedback.jsonl",
                {
                    "ts": edit["updated_at"],
                    "event": "row_saved",
                    "row_id": row_id,
                    "file": base["file"],
                    "cn_p_index": base["cn_p_index"],
                    "edited": current_html != base["cn_html"],
                    "marked": edit["marked"],
                    "issue_type": edit["issue_type"],
                    "severity": edit["severity"],
                    "comment": edit["comment"],
                    "learn_note": edit["learn_note"],
                    "tags": edit["tags"],
                    "jp_text": base["jp_text"],
                    "before_cn_text": strip_tags(base["cn_html"]),
                    "after_cn_text": strip_tags(current_html),
                    "before_cn_html": base["cn_html"],
                    "after_cn_html": current_html,
                },
            )
        return jsonify({"status": "ok", "row_id": row_id, "updated_at": edit["updated_at"]})

    @app.route("/api/apply", methods=["POST"])
    def api_apply():
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        with app.config["LOCK"]:
            modified = apply_edits_to_xhtml(session_root)
        return jsonify(
            {
                "status": "ok",
                "modified_files": modified,
                "feedback_md": str(session_root / "feedback" / "feedback_for_codex.md"),
            }
        )

    @app.route("/api/export", methods=["POST"])
    def api_export():
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        with app.config["LOCK"]:
            out_path = export_epub(session_root)
        return jsonify(
            {
                "status": "ok",
                "output": str(out_path),
                "download_url": f"/downloads/export/{urllib.parse.quote(out_path.name)}",
            }
        )

    @app.route("/api/feedback")
    def api_feedback():
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        path = write_feedback_summary(session_root)
        return jsonify(
            {
                "status": "ok",
                "feedback_md": str(path),
                "feedback_jsonl": str(session_root / "feedback" / "translation_feedback.jsonl"),
                "feedback_md_url": "/downloads/feedback/feedback_for_codex.md",
                "feedback_jsonl_url": "/downloads/feedback/translation_feedback.jsonl",
            }
        )

    @app.route("/downloads/export/<path:filename>")
    def download_export(filename: str):
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        return send_from_directory(session_root / "exports", filename, as_attachment=True)

    @app.route("/downloads/feedback/<path:filename>")
    def download_feedback(filename: str):
        session_root, _epub_path = active_session()
        if session_root is None:
            return no_active_response()
        if filename not in {"feedback_for_codex.md", "translation_feedback.jsonl"}:
            return jsonify({"error": "unknown feedback file"}), 404
        return send_from_directory(session_root / "feedback", filename, as_attachment=True)

    @app.route("/api/shutdown", methods=["POST"])
    def api_shutdown():
        def stop() -> None:
            time.sleep(0.3)
            os._exit(0)

        threading.Thread(target=stop, daemon=True).start()
        return jsonify({"status": "ok"})

    return app


def wait_until_ready(url: str, proc: subprocess.Popen, timeout: int = 15) -> bool:
    deadline = time.time() + timeout
    while time.time() < deadline:
        if proc.poll() is not None:
            return False
        try:
            with urllib.request.urlopen(f"{url}/api/session", timeout=1) as response:
                if response.status == 200:
                    return True
        except OSError:
            time.sleep(0.25)
    return False


def run_daemon(args: argparse.Namespace) -> int:
    port = find_free_port(args.port, args.host)
    review_root = Path(args.review_root).resolve()
    review_root.mkdir(parents=True, exist_ok=True)
    if args.epub:
        epub_path = Path(args.epub).resolve()
        if not epub_path.exists():
            print(f"EPUB not found: {epub_path}", file=sys.stderr)
            return 1
        try:
            validate_epub_file(epub_path)
        except ValueError as exc:
            print(str(exc), file=sys.stderr)
            return 1
        session_root = session_root_for(review_root, epub_path)
        if args.reset and session_root.exists():
            shutil.rmtree(session_root)
        session_root.mkdir(parents=True, exist_ok=True)
        log_path = session_root / "server.log"
    else:
        epub_path = None
        log_path = review_root / "server.log"
    cmd = [
        sys.executable,
        str(Path(__file__).resolve()),
        "--review-root",
        str(review_root),
        "--host",
        args.host,
        "--port",
        str(port),
    ]
    if epub_path is not None:
        cmd.insert(2, str(epub_path))
    cmd.append("--no-browser")
    if args.reset:
        cmd.append("--reset")
    creationflags = 0
    popen_kwargs: dict[str, Any] = {}
    if os.name == "nt":
        creationflags = subprocess.CREATE_NEW_PROCESS_GROUP | subprocess.DETACHED_PROCESS
    else:
        popen_kwargs["start_new_session"] = True
    with log_path.open("a", encoding="utf-8") as log:
        proc = subprocess.Popen(
            cmd,
            stdin=subprocess.DEVNULL,
            stdout=log,
            stderr=subprocess.STDOUT,
            creationflags=creationflags,
            **popen_kwargs,
        )
    url = f"http://{display_host(args.host)}:{port}"
    if not wait_until_ready(url, proc):
        print(f"review editor failed to start; see {log_path}", file=sys.stderr)
        return 1
    if not args.no_browser:
        try:
            webbrowser.open(url)
        except Exception:
            pass
    print(url)
    if epub_path is not None:
        print(f"session: {session_root}")
    else:
        print(f"review root: {review_root}")
    print(f"log: {log_path}")
    return 0


def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(description="Open a bilingual EPUB in a browser review/editor.")
    parser.add_argument("epub", nargs="?", help="Optional path to a bilingual EPUB. Omit it to upload EPUBs in the browser.")
    parser.add_argument("--review-root", default=DEFAULT_REVIEW_ROOT, help="Directory for review sessions.")
    parser.add_argument("--host", default="127.0.0.1", help="Host to bind. Use 0.0.0.0 for trusted LAN/server access.")
    parser.add_argument("--port", type=int, default=5177, help="Port, auto-advances if busy.")
    parser.add_argument("--daemon", action="store_true", help="Start in background and print URL.")
    parser.add_argument("--no-browser", action="store_true", help="Do not auto-open the browser.")
    parser.add_argument("--reset", action="store_true", help="Re-extract the EPUB and discard prior review state.")
    return parser


def main(argv: list[str] | None = None) -> int:
    args = build_parser().parse_args(argv)
    epub_path = Path(args.epub).resolve() if args.epub else None
    if epub_path is not None:
        if not epub_path.exists():
            print(f"EPUB not found: {epub_path}", file=sys.stderr)
            return 1
        try:
            validate_epub_file(epub_path)
        except ValueError as exc:
            print(str(exc), file=sys.stderr)
            return 1
    if args.daemon:
        return run_daemon(args)

    review_root = Path(args.review_root).resolve()
    app = create_app(review_root, epub_path, reset=args.reset)
    port = find_free_port(args.port, args.host)
    url = f"http://{display_host(args.host)}:{port}"
    if not args.no_browser:
        try:
            webbrowser.open(url)
        except Exception:
            pass

    def on_sigterm(_signum: int, _frame: Any) -> None:
        sys.exit(0)

    try:
        signal.signal(signal.SIGTERM, on_sigterm)
    except Exception:
        pass
    print(f"EPUB review editor: {url}")
    print(f"Review root: {review_root}")
    session_root = app.config.get("SESSION_ROOT")
    if isinstance(session_root, Path):
        print(f"Session: {session_root}")
        print(f"Rows: {len(read_json(session_root / 'review_state' / 'rows.json', []))}")
    else:
        print("No EPUB loaded yet. Upload one in the browser.")
    app.run(host=args.host, port=port, debug=False, threaded=True)
    return 0


if __name__ == "__main__":
    raise SystemExit(main())