专名本土化（信达雅）：冻结译名表 + 渲染层，并修掉校验器的正则盲点

- 新增 dfannals/glossary.py（译名表）与 dfannals/localize.py（渲染层）
  人名＝音译＋绰号意译（Ral Fastenhatchets→拉尔·缚斧、朗古德·商熊「剖胎之寒」），
  地名整体意译（Fordedwinds→涉风渡、The Volcano Of Ravens→渡鸦火山）
- 生成链路本土化：素材渲染、人物表、扩写白名单全部走译名；已有产物做确定性替换
- 修复 factcheck 的 \b 盲点：Unicode 下汉字也算词字符，紧贴汉字的英文名不被检查
  （曾因此漏掉编造专名）。修好后抓出并定点修掉长版第 1 集的 2 处编造
- 修正 title_form 大小写（YETI→Yeti），此前 Yeti/Goblin 一直漏译
- 新增 tests/test_localize.py（13 项回归）

dfannals/cast.py

@@ -15,7 +15,7 @@ import re
 from dataclasses import dataclass
 from pathlib import Path
 
-from dfannals import factcheck
+from dfannals import factcheck, localize
 from dfannals.config import MAX_TOKENS_STRUCTURED, Gateway
 from dfannals.episodes import Episode
 from dfannals.legends import Event, World
@@ -221,7 +221,8 @@ def build_cast(
             "",
         ]
 
-    return "\n".join(lines) + "\n", rows, suspicions
+    # 专名本土化：表格与小传都输出中文译名
+    return localize.for_world(world).render("\n".join(lines)) + "\n", rows, suspicions
 
 
 def write_cast(path: Path, text: str) -> Path:

dfannals/expand.py

@@ -18,7 +18,7 @@ from collections import Counter
 from dataclasses import replace
 from pathlib import Path
 
-from dfannals import deslop, volume
+from dfannals import deslop, localize, volume
 from dfannals import episodes as ep_mod
 from dfannals.chronicle import cjk_length
 from dfannals.config import (
@@ -100,14 +100,16 @@ def allowed_names(world: World, thread: Thread, episode: ep_mod.Episode) -> list
             names.add(name)
         figure = world.figures.get(fid)
         if figure and figure.race:
-            names.add(figure.race)
-            names.add(_title_form(figure.race))
+            # 种族值也要给中文：模型引到雪人、鹰人这类词时不能写成新角色
+            names.add(localize.translate_race(figure.race))
     for event in episode.events:
         for ref in world.render_refs(event):
             value = ref.partition("=")[2].strip()
             if value:
                 names.add(value)
-    return sorted(n for n in names if n)
+    # 全部过一遍译名表：白名单里必须是正文里真正会出现的写法
+    translator = localize.for_world(world)
+    return sorted({tr for tr in (translator.render(n).strip() for n in names) if tr})
 
 
 def chunk_messages(

dfannals/factcheck.py

@@ -35,7 +35,10 @@ COMMON_WORDS = {
 }
 
 # 拉丁专名序列：首字母大写的词，可含 ' - 与数字
-NAME_RE = re.compile(r"\b[A-Z][A-Za-z'\-]*(?:\s+[A-Z][A-Za-z'\-]*)*\b")
+# 英文专名的边界不能用 \b：Unicode 下汉字也算「词字符」，于是 `在Mistbreath的`
+# 这种紧贴汉字的英文名根本不会被匹配——实测正是这个盲点让一个编造的地名混过了校验。
+# 改成前后看「是不是英文字母」，与汉字相邻也能抓到。
+NAME_RE = re.compile(r"(?<![A-Za-z])[A-Z][A-Za-z'\-]*(?:\s+[A-Z][A-Za-z'\-]*)*(?![A-Za-z])")
 _CJK = re.compile(r"[\u4e00-\u9fff]")
 
 

dfannals/glossary.py

@@ -0,0 +1,208 @@
+"""专名译名表（冻结）。
+
+为什么要有这张表：正文里满屏 `Ral Fastenhatchets`、`Fordedwinds` 对中文读者很难读。
+但如果交给模型每次现译，同一个角色会在第 3 集变成另一个名字，所以译名必须冻结在这里。
+
+翻译策略（信达雅）：
+
+1. **人名 = 名字音译 ＋ 姓氏／绰号意译**，用「·」连接。
+   这与《魔戒》把 Thorin Oakenshield 译作「索林·橡木盾」是同一套惯例，
+   读起来像中国人名，又保住了原名的意思。
+2. **地名整体意译**。DF 地名同样是词表拼接（Fordedwinds = forded + winds），
+   直译成「涉风渡」比音译成「福德温兹」更像地名。
+3. **称号（the X of Y）单独译**，DF 给传奇人物的长称号自带画面感，直译最好。
+
+DF 的名字由词表拼接而成，所以同一批词根会在不同世界反复出现：
+这张表按「整词」而非「词根」收词，避免把 Fateddie 硬拆成"命定+死"这类生硬堆叠。
+遇到表里没有的名字，localize 会原样保留并登记进待译清单，而不是硬造一个。
+"""
+
+# ------------------------------------------------------------------ 名字词干（音译）
+STEMS: dict[str, str] = {
+    "alath": "阿拉斯",
+    "amxu": "阿姆苏",
+    "ana": "安娜",
+    "artuk": "阿图克",
+    "asmel": "阿斯梅尔",
+    "astri": "阿斯特里",
+    "domas": "多马斯",
+    "dostngosp": "多斯特恩戈斯普",
+    "ducim": "杜西姆",
+    "espu": "埃斯普",
+    "evala": "埃瓦拉",
+    "guspu": "古斯普",
+    "id": "伊德",
+    "ineth": "伊内斯",
+    "ingish": "英吉什",
+    "langgud": "朗古德",
+    "lolor": "洛洛尔",
+    "meng": "蒙",
+    "moldath": "莫尔达斯",
+    "momuz": "莫穆兹",
+    "mosus": "莫苏斯",
+    "mudi": "穆迪",
+    "ngokang": "恩戈康",
+    "ngom": "恩戈姆",
+    "niri": "尼里",
+    "nithe": "尼瑟",
+    "ongu": "翁古",
+    "onul": "奥努尔",
+    "pari": "帕里",
+    "perom": "佩罗姆",
+    "ral": "拉尔",
+    "rigoth": "里戈斯",
+    "sathrel": "萨瑟雷尔",
+    "sayrsnis": "赛尔斯尼斯",
+    "shosa": "绍萨",
+    "sina": "西娜",
+    "someg": "索梅格",
+    "song": "松",
+    "stinthad": "斯廷哈德",
+    "tekkud": "泰库德",
+    "thrun": "斯伦",
+    "tirist": "提里斯特",
+    "tulon": "图隆",
+    "uknol": "乌克诺尔",
+    "usel": "乌塞尔",
+    "utes": "乌特斯",
+    "zuntir": "尊提尔",
+    "zutthan": "祖坦",
+}
+
+# ------------------------------------------------------------------ 姓氏／绰号（意译）
+# 键是姓名后半段的拼接结果（小写、无空格）
+SURNAMES: dict[str, str] = {
+    "baldedattic": "秃阁",
+    "basementrampart": "地垒",
+    "blazedbrushes": "焰刷",
+    "blottedmine": "污矿",
+    "bonestands": "骨立",
+    "bronzescorched": "铜焦",
+    "churchbrass": "堂铜",
+    "clappedbrass": "拍铜",
+    "crazeowned": "狂据",
+    "decentpage": "端页",
+    "dyesong": "染歌",
+    "fastenhatchets": "缚斧",
+    "fateddie": "命终",
+    "fencefinders": "篱寻",
+    "fisherfair": "渔市",
+    "flayerline": "剥血",
+    "frillyknots": "褶结",
+    "glimmeredhelps": "光援",
+    "glowroofs": "光顶",
+    "gleamcircled": "光环",
+    "growllie": "吼卧",
+    "hammershake": "锤摇",
+    "huggedfastened": "抱固",
+    "incenselyrics": "香诗",
+    "jackalbrims": "豺帽",
+    "matchedmenaces": "匹凶",
+    "menacedslinks": "危潜",
+    "mirroredfloors": "镜层",
+    "oilshove": "油推",
+    "palmsaved": "掌救",
+    "playfills": "戏填",
+    "ruinedhorrors": "毁怖",
+    "searcheddated": "寻期",
+    "squirtlocks": "喷锁",
+    "steelgrouped": "钢聚",
+    "stilledearthen": "静土",
+    "stolendipped": "窃浸",
+    "strikepaints": "击彩",
+    "talonhollows": "爪穴",
+    "theatercrested": "剧冠",
+    "thornticks": "棘蜱",
+    "tongswall": "钳墙",
+    "tradebears": "商熊",
+    "walkclasped": "行扣",
+    "watchedleads": "守铅",
+    "wheeleddrinks": "轮饮",
+    "wisdomskins": "智皮",
+    "yellplank": "啸板",
+}
+
+# ------------------------------------------------------------------ 地名（整体意译）
+PLACES: dict[str, str] = {
+    "clampokes": "钳戳",
+    "fordedwinds": "涉风渡",
+    "fuchsiarumor": "紫谣",
+    "gloryholes": "荣耀穴",
+    "lashheroes": "鞭雄",
+    "leapdrinks": "跃饮",
+    "liecats": "卧猫",
+    "mortalpears": "殇梨",
+    "paperhushed": "纸寂",
+    "rainseduced": "雨诱",
+    "tomessoldiers": "典兵",
+    "toolfainted": "器昏",
+    "walkedbirds": "行鸟",
+}
+
+# ------------------------------------------------------------------ 称号（the X of Y）
+EPITHETS: dict[str, str] = {
+    "the eviscerated chill of pregnancies": "剖胎之寒",
+    "the trickery of mornings": "晨之诡计",
+}
+
+# ------------------------------------------------------------------ 势力／组织
+GROUPS: dict[str, str] = {
+    "the volcano of ravens": "渡鸦火山",
+}
+
+# ------------------------------------------------------------------ 种族
+RACES: dict[str, str] = {
+    "AARDVARK_MAN": "土豚人",
+    "ALLIGATOR": "鳄鱼",
+    "ALLIGATOR_MAN": "鳄人",
+    "BAT_GIANT": "巨蝠",
+    "BEAR_BLACK": "黑熊",
+    "BEAR_GRIZZLY": "灰熊",
+    "BEAR_GRIZZLY_MAN": "灰熊人",
+    "BEAR_POLAR": "白熊",
+    "BIRD_OWL_GREAT_HORNED": "大雕鸮",
+    "BIRD_ROC": "大鹏",
+    "BIRD_WREN": "鹩鹪",
+    "BLIND_CAVE_BEAR": "盲穴熊",
+    "BLIND_CAVE_OGRE": "盲穴食人魔",
+    "BOBCAT": "山猫",
+    "CAMEL_1_HUMP_MAN": "单峰驼人",
+    "CARDINAL_MAN": "红衣主教雀人",
+    "CAVE_DRAGON": "穴龙",
+    "CHEETAH": "猎豹",
+    "CHINCHILLA_MAN": "绒鼠人",
+    "COLOSSUS_BRONZE": "青铜巨像",
+    "COUGAR": "美洲狮",
+    "CYCLOPS": "独眼巨人",
+    "DOG": "犬",
+    "DWARF": "矮人",
+    "EAGLE_MAN": "鹰人",
+    "ELEPHANT": "象",
+    "ELF": "精灵",
+    "FALCON": "隼族",
+    "GOAT_MOUNTAIN_MAN": "山羊山人",
+    "GOBLIN": "哥布林",
+    "HUMAN": "人类",
+    "HYDRA": "九头蛇",
+    "JAGUAR": "美洲豹",
+    "KOBOLD": "狗头人",
+    "LEOPARD": "豹",
+    "LION": "狮",
+    "TIGER": "虎",
+    "TROLL": "巨魔",
+    "WOLF": "狼",
+    "YETI": "雪人",
+}
+
+# ------------------------------------------------------------------ 地点类型
+SITE_TYPES: dict[str, str] = {
+    "cave": "洞穴",
+    "dark fortress": "黑堡",
+    "dark pits": "幽坑",
+    "forest retreat": "林居",
+    "fortress": "要塞",
+    "hamlet": "村落",
+    "hillocks": "丘屯",
+    "tomb": "陵墓",
+    "town": "城镇",
+}

dfannals/localize.py

@@ -0,0 +1,240 @@
+"""专名本土化：把史料里的英文专名翻成中文（只作用于显示层）。
+
+铁律：**只改显示，不改史料**。World 里的英文原名、事件字段、事实校验依据全部保持原样，
+译文只在这一层生成，所以：
+
+- 事实校验仍以英文原名为准（译文通过 `Translator.accepted()` 参与校验，不会被当成编造）；
+- 译名不写回 data/export，不改变 legends 的解析结果，去掉这层也不影响任何已产出的史料。
+
+表里查不到的名字**原样保留**并登记进 `pending`，绝不硬造一个音译——
+一个难读的英文名远好过一个每次都不一样的乱译名。
+"""
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+
+from dfannals import glossary
+from dfannals.legends import World
+
+_WORD_SPLIT = re.compile(r"\s+")
+_CJK = r"\u4e00-\u9fff\u3000-\u303f\uff00-\uffef"
+
+
+def split_epithet(name: str) -> tuple[str, str]:
+    """拆出 DF 的长称号：`Xxx Yyy the Zzz Of Www` → ('Xxx Yyy', 'the zzz of www')。"""
+    lowered = name.strip().lower()
+    marker = " the "
+    if marker in lowered:
+        main, epithet = lowered.split(marker, 1)
+        return main, "the " + epithet
+    return lowered, ""
+
+
+def translate_person(name: str) -> str | None:
+    """人名：名字音译 ＋ 姓氏意译。查不全就返回 None（宁可保留英文）。"""
+    if not name:
+        return None
+    main, epithet = split_epithet(name)
+    parts = _WORD_SPLIT.split(main)
+    stem = glossary.STEMS.get(parts[0])
+    if stem is None:
+        return None
+    surname = ""
+    if len(parts) > 1:
+        surname = glossary.SURNAMES.get("".join(parts[1:]), "")
+        if not surname:
+            return None
+    out = stem if not surname else f"{stem}·{surname}"
+    if epithet:
+        out += "「" + glossary.EPITHETS.get(epithet, "") + "」"
+        if not glossary.EPITHETS.get(epithet):
+            return None
+    return out
+
+
+def translate_place(name: str) -> str | None:
+    """地名：整体意译；带称号的写成「主名（称号）」。"""
+    if not name:
+        return None
+    main, epithet = split_epithet(name)
+    base = glossary.PLACES.get(main.replace(" ", ""))
+    if base is None:
+        return None
+    if epithet:
+        extra = glossary.EPITHETS.get(epithet)
+        return f"{base}（{extra}）" if extra else None
+    return base
+
+
+def translate_group(name: str) -> str | None:
+    if not name:
+        return None
+    return glossary.GROUPS.get(name.strip().lower())
+
+
+def translate_race(value: str) -> str:
+    """种族值（DWARF、EAGLE_MAN）→ 中文；查不到原样返回。"""
+    if not value:
+        return value
+    return glossary.RACES.get(value.strip().upper(), value)
+
+
+def translate_site_type(value: str) -> str:
+    if not value:
+        return value
+    return glossary.SITE_TYPES.get(value.strip().lower(), value)
+
+
+@dataclass
+class Translator:
+    """把一整个世界的专名译好，供渲染与校验共用。"""
+
+    people: dict[str, str] = field(default_factory=dict)   # 英文原名 → 中文
+    places: dict[str, str] = field(default_factory=dict)
+    groups: dict[str, str] = field(default_factory=dict)
+    stems: dict[str, str] = field(default_factory=dict)     # 单用名（正文里绝大多数是简称）
+    surnames: dict[str, str] = field(default_factory=dict)  # 单用姓（对白里互相叫姓）
+    races: dict[str, str] = field(default_factory=dict)      # 含全大写与首字母大写两种写法
+    pending: list[str] = field(default_factory=list)       # 表里没有、原样保留的名字
+
+    def person(self, name: str) -> str:
+        return self.people.get(name, name)
+
+    def place(self, name: str) -> str:
+        return self.places.get(name, name)
+
+    def group(self, name: str) -> str:
+        return self.groups.get(name, name)
+
+    def _pairs(self) -> list[tuple[str, str]]:
+        """按英文原名长度降序：先替换长的，避免 `Ral` 抢在 `Ral Fastenhatchets` 前面。
+
+        顺序很重要：全名 → 单独的姓 → 单独的名。换掉全名后正文里已无拉丁字母，
+        后面两步不会再动它。
+        """
+        full = list(self.people.items()) + list(self.places.items()) + list(self.groups.items())
+        full.sort(key=lambda kv: -len(kv[0]))
+        tail = list(self.surnames.items()) + list(self.stems.items())
+        tail.sort(key=lambda kv: -len(kv[0]))
+        return full + tail
+
+    def render(self, text: str) -> str:
+        """把已有正文里的英文专名换成中文，并收拾替换后多余的空格。"""
+        for source, target in self._pairs():
+            text = re.sub(rf"(?<![A-Za-z]){re.escape(source)}(?![A-Za-z])", target, text)
+        # 种族：既要处理枚举值 EAGLE_MAN，也要处理正文里的 Yeti/Eagle_man
+        for source, target in sorted(self.races.items(), key=lambda kv: -len(kv[0])):
+            text = re.sub(rf"(?<![A-Za-z_]){re.escape(source)}(?![A-Za-z_])", target, text)
+        # 地点类型是小写短语（forest retreat、dark pits）
+        for source, target in sorted(glossary.SITE_TYPES.items(), key=lambda kv: -len(kv[0])):
+            text = re.sub(rf"(?<![A-Za-z]){re.escape(source)}(?![A-Za-z])", target, text,
+                          flags=re.IGNORECASE)
+        return tidy_spaces(text)
+
+    def accepted(self) -> set[str]:
+        """译名集合：事实校验要把它们当已知名字，否则会把译文当成凭空编造。"""
+        out: set[str] = set()
+        for target in [*self.people.values(), *self.places.values(), *self.groups.values()]:
+            out.add(target)
+            out.update(re.split(r"[·（）「」，、]", target))
+        return {name for name in out if name}
+
+    def coverage(self) -> str:
+        rows = [
+            (
+                f"译名覆盖：人物 {len(self.people)} 个｜地点 {len(self.places)} 个｜"
+                f"势力 {len(self.groups)} 个"
+            ),
+        ]
+        if self.pending:
+            rows.append(f"待译（原样保留）{len(self.pending)} 个：" + "、".join(self.pending[:12]))
+        return "\n".join(rows)
+
+
+def tidy_spaces(text: str) -> str:
+    """去掉中日韩字符之间被替换后留下的空格：`拉尔 说` → `拉尔说`。
+
+    英文专名之间、以及英文与汉字之间的空格保留（`Ral Fastenhatchets` 用得上）。
+    """
+    text = re.sub(rf"(?<=[{_CJK}])[ \t]+(?=[{_CJK}])", "", text)
+    text = re.sub(rf"(?<=[{_CJK}])[ \t]+(?=[，。、；：！？）】」])", "", text)
+    text = re.sub(rf"(?<=[（【「])[ \t]+(?=[{_CJK}])", "", text)
+    return text
+
+
+def title_form(value: str) -> str:
+    """YETI → Yeti、EAGLE_MAN → Eagle_man、fastenhatchets → Fastenhatchets。
+
+    两个坑都踩过：
+    - 只改首字母不转小写，`YETI` 会得到 `YETI`，永远匹配不上正文里的 `Yeti`；
+    - 按下划线逐段首字母大写，`EAGLE_MAN` 会得到 `Eagle_Man`，
+      而史料渲染惯例是 `Eagle_man`（见 legends.pretty）。
+    """
+    lowered = value.lower()
+    return lowered[:1].upper() + lowered[1:]
+
+
+def build(world: World) -> Translator:
+    """按世界里的实际名字建翻译器。"""
+    translator = Translator()
+    # 名与姓都单独收一份：正文里大量使用简称（"Ral"）和对白里只呼姓（"Fastenhatchets"）
+    translator.stems = {title_form(en): zh for en, zh in glossary.STEMS.items()}
+    translator.surnames = {title_form(en): zh for en, zh in glossary.SURNAMES.items()}
+    translator.races = {}
+    for en, zh in glossary.RACES.items():
+        translator.races[en] = zh
+        translator.races[title_form(en)] = zh
+        # 逐段大写的写法（Eagle_Man）也登记：模型两种都可能写
+        translator.races["_".join(p[:1].upper() + p[1:].lower() for p in en.split("_"))] = zh
+    for figure in world.figures.values():
+        raw = figure.name
+        if not raw:
+            continue
+        translated = translate_person(raw)
+        shown = _display(raw)
+        if translated:
+            translator.people[shown] = translated
+            translator.people[raw] = translated
+        elif _WORD_SPLIT.split(shown.lower())[0] in glossary.STEMS:
+            # 名字能读、姓氏没词条：这是真正需要补的缺口，登记下来
+            translator.pending.append(shown)
+    for site in world.sites.values():
+        raw = site.name
+        if not raw:
+            continue
+        translated = translate_place(raw)
+        shown = _display(raw)
+        if translated:
+            translator.places[shown] = translated
+            translator.places[raw] = translated
+    for entity in world.entities.values():
+        raw = entity.name
+        if not raw:
+            continue
+        translated = translate_group(raw)
+        shown = _display(raw)
+        if translated:
+            translator.groups[shown] = translated
+            translator.groups[raw] = translated
+    return translator
+
+
+_CACHE: dict[int, tuple[World, Translator]] = {}
+
+
+def for_world(world: World) -> Translator:
+    """带缓存的建立器：一次运行里反复渲染素材，不能每次都重建 4 万个人名。"""
+    hit = _CACHE.get(id(world))
+    if hit is not None and hit[0] is world:
+        return hit[1]
+    translator = build(world)
+    _CACHE[id(world)] = (world, translator)
+    return translator
+
+
+def _display(raw: str) -> str:
+    """史料的程序化名全小写，渲染层会做首字母大写；译名要按展示形式建索引。"""
+    from dfannals.legends import pretty
+
+    return pretty(raw)

dfannals/volume.py

@@ -12,7 +12,7 @@ from dataclasses import dataclass, replace
 from pathlib import Path
 
 from dfannals import episodes as ep_mod
-from dfannals import gate
+from dfannals import gate, localize
 from dfannals.cast import collect_rows
 from dfannals.chronicle import cjk_length
 from dfannals.config import (
@@ -194,12 +194,13 @@ def render_material(
         "",
         "### 写作要求",
         "",
-        f"- 本集 {CHAPTER_MIN_CHARS}–{CHAPTER_MAX_CHARS} 字，中文正文，专名保留英文。",
+        f"- 本集 {CHAPTER_MIN_CHARS}–{CHAPTER_MAX_CHARS} 字，中文正文，人名地名用上面给出的中文译名。",
         "- 贴着主角写：他们的目标、算计、得失做主语；对手要是个具体的人。",
         "- 性别按上面卡片写，不要默认“他”；史料未载性别时用名字或身份称呼。",
         "- 史料之外的世界大事不要写；只有影响到主角时才提一句。",
     ]
-    return "\n".join(lines)
+    # 专名本土化：素材里的英文原名换成中文译名，让模型直接用中文写
+    return localize.for_world(world).render("\n".join(lines))
 
 
 def build_messages(