master-skill 0.12.10 → 0.12.12

package/tools/compiled-teaching-sources.json

@@ -0,0 +1,89 @@
+{
+  "_comment": "CBETA 不收、但有免费全文可取的编集语录。周检 3i 用它把人设里当原话引的句子拿到原书里逐字找。3h 只能查 CBETA，查不到时对这些祖师一律记『未判定』——2026-09-15 修掉的那批拼接引文（虚云、印光）正是长在这个盲区里。coverage 决定这一步能不能判错：complete 表示该祖师声明的编集语录全都在下面取得到，找不到即伪造；partial 表示还有声明了却取不到全文的，找不到只记未判定，绝不报错。宁可少判，不可错判。",
+  "corpora": [
+    {
+      "master": "master-yinguang",
+      "corpus_title": "《印光法师文钞》",
+      "coverage": "complete",
+      "coverage_reason": "meta.json 声明的编集语录是正编、续编、三编，外加一个只作总名解析用的『Yinguang:Wenchao』（正续三编的合称，非另一部书）。这三编就是《文钞》的全部，且都在殆知阁语料库里。2026-09-16 核验：三个文件在 commit e9e11f1 下分别为 1285781 / 1410227 / 1973347 字节，且 voice.md 现有三条引文逐字命中（正编 2 条、续编 1 条）。故查不到即可判错。",
+      "verified_on": "2026-09-16",
+      "not_a_separate_book": [
+        "Yinguang:Wenchao"
+      ],
+      "not_a_separate_book_reason": "『Yinguang:Wenchao』是正编、续编、三编的合称，meta.json 里留着它只为解析只写总名的引用，并非第四部书。列在这里，coverage=complete 才是可核验的声明而不是一句断言——测试会要求每个声明过的编集来源要么有全文、要么在这里说明为什么不需要。",
+      "texts": [
+        {
+          "id": "Yinguang:WenchaoZhengbian",
+          "title": "印光法师文钞正编",
+          "url": "https://raw.githubusercontent.com/daizhige-org/daizhigev20/e9e11f19b7e6bd9c1284bbdab71d2a9cb94d637c/佛藏/藏外/印光法师文钞.md",
+          "encoding": "utf-8"
+        },
+        {
+          "id": "Yinguang:WenchaoXubian",
+          "title": "印光法师文钞续编",
+          "url": "https://raw.githubusercontent.com/daizhige-org/daizhigev20/e9e11f19b7e6bd9c1284bbdab71d2a9cb94d637c/佛藏/藏外/印光法师文钞续编.md",
+          "encoding": "utf-8"
+        },
+        {
+          "id": "Yinguang:WenchaoSanbian",
+          "title": "印光法师文钞三编",
+          "url": "https://raw.githubusercontent.com/daizhige-org/daizhigev20/e9e11f19b7e6bd9c1284bbdab71d2a9cb94d637c/佛藏/藏外/印光法师文钞三编.md",
+          "encoding": "utf-8"
+        }
+      ],
+      "source_note": "殆知阁古代文献 v2.0（github.com/daizhige-org/daizhigev20），默认分支是 data 不是 master——用 master 取会 404，2026-09-16 就这样踩过一次。地址钉在 commit e9e11f1（2026-09-12）上，分支移动不会悄悄换掉被核对的底本；哪天该 commit 取不到，3i 记未判定而不是判错。url 一律写字面字符、不要预先百分号编码：编码由 fetch_compiled_text 统一做，清单里再编一次就成了 %25 开头的双重编码，一取就 404 —— 2026-09-16 首次实跑正是这样让印光三部全部『取不到』的。"
+    },
+    {
+      "master": "master-xuyun",
+      "corpus_title": "《虚云和尚法汇》《虚云老和尚年谱》",
+      "coverage": "partial",
+      "coverage_reason": "meta.json 声明了开示录、法汇、年谱三种，BFNN 上只有岑学吕编的法汇六部与年谱，没有《虚云老和尚开示录》——后者是净慧所编，《虚云和尚全集》相对岑本新增约六十余万字，仅开示就多出 110 余则。也就是说虚云的真引文完全可能出自这里取不到的那部分，所以这部语料只能用来确认，不能用来判错。哪天开示录有了可取的免费全文，再改成 complete。",
+      "verified_on": "2026-09-16",
+      "texts": [
+        {
+          "id": "Xuyun:Nianpu",
+          "title": "虚云和尚年谱",
+          "url": "http://bookgb.bfnn.org/books2/1184.htm",
+          "encoding": "gb18030"
+        },
+        {
+          "id": "Xuyun:Fahui",
+          "title": "虚云和尚法汇—法语",
+          "url": "http://bookgb.bfnn.org/books2/1185.htm",
+          "encoding": "gb18030"
+        },
+        {
+          "id": "Xuyun:Fahui",
+          "title": "虚云和尚法汇—开示",
+          "url": "http://bookgb.bfnn.org/books2/1186.htm",
+          "encoding": "gb18030"
+        },
+        {
+          "id": "Xuyun:Fahui",
+          "title": "虚云和尚法汇—书问",
+          "url": "http://bookgb.bfnn.org/books2/1187.htm",
+          "encoding": "gb18030"
+        },
+        {
+          "id": "Xuyun:Fahui",
+          "title": "虚云和尚法汇—文记",
+          "url": "http://bookgb.bfnn.org/books2/1188.htm",
+          "encoding": "gb18030"
+        },
+        {
+          "id": "Xuyun:Fahui",
+          "title": "虚云和尚法汇—规约",
+          "url": "http://bookgb.bfnn.org/books2/1189.htm",
+          "encoding": "gb18030"
+        },
+        {
+          "id": "Xuyun:Fahui",
+          "title": "虚云和尚法汇—诗歌偈赞",
+          "url": "http://bookgb.bfnn.org/books2/1190.htm",
+          "encoding": "gb18030"
+        }
+      ],
+      "source_note": "BFNN（bookgb.bfnn.org）转录岑学吕编本，gb18030 编码，默认 Python-urllib 即可取（不像 FoJin 会挡 UA）。2026-09-16 核验：voice.md 现有三条引文逐字命中 1186《法汇—开示》。"
+    }
+  ]
+}

package/tools/master_builder.py

@@ -309,6 +309,76 @@ def build_from_spec(spec: dict, output_dir: str) -> dict:
     }
 
 
+_SKILL_NAME = re.compile(r"^name:[ \t]*(\S+)[ \t]*$", re.MULTILINE)
+
+
+def register_teacher(teacher_dir: str, skills_dir: str) -> dict:
+    """Make a generated persona invocable by linking it into a skills directory.
+
+    The generator writes to `${CLAUDE_SKILL_DIR}/masters/master-{slug}/`, so that
+    `master-skill update` can carry generated personas across runtime updates.
+    Claude Code does not look there: it loads `<skills dir>/<name>/SKILL.md` and
+    scans no deeper. Measured 2026-09-16 with Claude Code 2.1.273 in an isolated
+    config, `/skills` listed a probe skill at `~/.claude/skills/master-control/`
+    and not one at `~/.claude/skills/create-master/masters/master-probe/`; a
+    directory symlink `~/.claude/skills/master-probe` made it appear in the same
+    session, without a restart.
+
+    An existing entry of the same name is never replaced: a user regenerating a
+    prebuilt master (`master-huineng`) must not overwrite the installed one.
+    """
+    teacher = Path(teacher_dir).resolve()
+    skill_md = teacher / "SKILL.md"
+    if not skill_md.is_file():
+        raise ValueError(f"{teacher} has no SKILL.md")
+    text = skill_md.read_text(encoding="utf-8")
+    frontmatter = text.split("---", 2)[1] if text.startswith("---") else ""
+    match = _SKILL_NAME.search(frontmatter)
+    if not match or match.group(1) != teacher.name:
+        raise ValueError(
+            f"SKILL.md name must equal the directory name {teacher.name!r} — "
+            "Claude Code invokes the skill by that name"
+        )
+
+    skills = Path(skills_dir).expanduser()
+    created_skills_dir = not skills.is_dir()
+    skills.mkdir(parents=True, exist_ok=True)
+    link = skills / teacher.name
+
+    if link.exists() or link.is_symlink():
+        if link.exists() and link.resolve() == teacher:
+            return {
+                "registered": str(link),
+                "target": str(teacher),
+                "invoke": f"/{teacher.name}",
+                "already_registered": True,
+                "restart_required": False,
+            }
+        raise ValueError(
+            f"{link} already exists and is not this persona; not replacing it. "
+            f"Rename the generated persona, or remove {link} yourself."
+        )
+
+    try:
+        os.symlink(teacher, link, target_is_directory=True)
+    except OSError:
+        if os.name != "nt":
+            raise
+        import _winapi  # symlinks need a privilege on Windows; junctions do not
+
+        _winapi.CreateJunction(str(teacher), str(link))
+
+    return {
+        "registered": str(link),
+        "target": str(teacher),
+        "invoke": f"/{teacher.name}",
+        "already_registered": False,
+        # Claude Code watches skill directories that existed when the session
+        # started; a skills directory created now is seen after a restart.
+        "restart_required": created_skills_dir,
+    }
+
+
 def main(argv: list[str] | None = None) -> int:
     parser = argparse.ArgumentParser(
         description="Build a create-master persona from an explicit generation spec"
@@ -320,15 +390,29 @@ def main(argv: list[str] | None = None) -> int:
         action="store_true",
         help="run a deterministic no-network generation smoke",
     )
-    parser.add_argument("--output", required=True, help="master output directory")
+    modes.add_argument(
+        "--register",
+        metavar="TEACHER_DIR",
+        help="link a generated persona into --skills-dir so it can be invoked",
+    )
+    parser.add_argument("--output", help="master output directory (with --spec / --offline-smoke)")
+    parser.add_argument(
+        "--skills-dir",
+        default=os.path.join("~", ".claude", "skills"),
+        help="skills directory to register into (with --register; default ~/.claude/skills)",
+    )
     args = parser.parse_args(argv)
+    if not args.register and not args.output:
+        parser.error("--output is required with --spec / --offline-smoke")
 
     try:
-        if args.offline_smoke:
-            spec = _offline_smoke_spec()
+        if args.register:
+            summary = register_teacher(args.register, args.skills_dir)
+        elif args.offline_smoke:
+            summary = build_from_spec(_offline_smoke_spec(), args.output)
         else:
             spec = json.loads(Path(args.spec).read_text(encoding="utf-8"))
-        summary = build_from_spec(spec, args.output)
+            summary = build_from_spec(spec, args.output)
     except (OSError, json.JSONDecodeError, ValueError) as exc:
         print(f"ERROR: {exc}", file=sys.stderr)
         return 1

package/tools/verify_sources.py

@@ -673,6 +673,80 @@ def collect_excerpt_quotes() -> list[tuple[str, str, str, int | None]]:
     return quotes
 
 
+def collect_compiled_excerpt_blocks() -> list[tuple[str, str, str]]:
+    """(位置, 引文, 所引篇名)：「原典」块中引用格式指向 CBETA 之外编集语录的那些。
+
+    `collect_excerpt_quotes` 只收引用格式带 CBETA 经号的块，而《文钞》没有经号，
+    于是 master-yinguang 的五个「原典」块对 3f 不可见；它们的 `>` 行又是裸行文、
+    不带引号，`collect_persona_quotes` 同样收不到。2026-09-16 核出其中两块是用
+    真语拼接的改写，却一直以「原典」示人 —— 没有任何一步检查看得见它们。
+    """
+    base = Path(PREBUILT_DIR)
+    blocks: list[tuple[str, str, str]] = []
+    for path in sorted(base.glob("*/sources/*-excerpts.md")):
+        where = path.relative_to(base).as_posix()
+        label_line, lines = 0, []
+        for number, line in enumerate(path.read_text(encoding="utf-8").splitlines(), 1):
+            if line.startswith("原典"):
+                label_line, lines = number, []
+            elif not label_line:
+                continue
+            elif line.startswith(">"):
+                lines.append(line[1:].strip())
+            elif line.startswith("#"):
+                label_line, lines = 0, []
+            elif "引用格式" in line:
+                citation = _DOC_CITATION.search(line)
+                text = citation.group(1) if citation else ""
+                if text and not _DOC_CBETA_ID.search(text) and any(lines):
+                    blocks.append((f"{where}:{label_line}", "\n".join(lines), text))
+                label_line, lines = 0, []
+    return blocks
+
+
+def classify_compiled_excerpt_blocks(
+    blocks: list[tuple[str, str, str]],
+    corpora: dict[str, dict],
+    fetch,
+) -> tuple[list[tuple[str, str, str]], list[tuple[str, str]], list[tuple[str, str]]]:
+    """「原典」块是不是所引编集语录的原文；省略号分段，每段都要在原书里找得到。
+
+    与 3i 判引文行同理：原书取不到一律记未判定，接口不通不是证据。语料不全
+    （coverage=partial）也只能确认、不能定罪 —— 判「原书没有」需要读得到全部。
+    """
+    mismatched: list[tuple[str, str, str]] = []
+    verified: list[tuple[str, str]] = []
+    unknown: list[tuple[str, str]] = []
+    bodies: dict[str, str | None] = {}
+    for where, quote, citation in blocks:
+        master = where.split("/", 1)[0]
+        corpus = corpora.get(master)
+        if not corpus:
+            unknown.append((where, f"{master} declares no fetchable corpus"))
+            continue
+        texts = corpus.get("texts") or []
+        for text in texts:
+            url = str(text.get("url"))
+            if url not in bodies:
+                bodies[url] = fetch(url, text.get("encoding") or "utf-8")
+        readable = [bodies.get(str(t.get("url"))) for t in texts]
+        if not any(body for body in readable):
+            unknown.append((where, "could not read the declared full texts"))
+            continue
+        segments = [s for s in re.split(r"…+", quote) if len(_han_only(s)) >= EXCERPT_MIN_CLAUSE * 2]
+        if not segments:
+            unknown.append((where, "no segment long enough to search"))
+            continue
+        absent = [s for s in segments if not any(b and _han_only(s) in b for b in readable)]
+        if not absent:
+            verified.append((where, citation))
+        elif corpus.get("coverage") == "complete":
+            mismatched.append((where, absent[0].strip(), citation))
+        else:
+            unknown.append((where, f"{master}'s free full texts do not cover every declared compilation"))
+    return mismatched, verified, unknown
+
+
 def cbeta_juan_plain_text(html: str) -> str:
     """`/stable/juans` 返回的 HTML → 正文。
 
@@ -963,11 +1037,35 @@ CBETA_SEARCH_URL = "https://cbdata.dila.edu.tw/stable/search"
 _QUOTE_SAMPLE = re.compile(r'^\s*\d+\.\s*[“"「『]([^”"」』\n]{8,200})')
 _QUOTE_BLOCK = re.compile(r'^\s*>\s*[“"「『]([^”"」』\n]{8,200})')
 _QUOTE_SAID = re.compile(r'(?:云|曰|偈云|经云|论云)\s*[：:]?\s*[“"「『]([^”"」』\n]{8,200})')
+# 具名引出 + **冒号**：「佛说：""」「神秀偈：""」「慧能曰：""」「达摩祖师偈：""」。
+# 冒号是把「引原典」与「人设自己的话」分开的判别式 —— 后者写作「常说"看看那个想要
+# 解决问题的心"」「先问"为什么想读？"」，一律没有冒号。2026-09-16 量过：这一条能收进
+# 慧能的风幡、神秀与达摩的偈、阿姜查所引三段巴利经文，而不碰任何一句话术示例。
+_QUOTE_ATTRIBUTED = re.compile(
+    r"(?:佛|世尊|[㐀-鿿]{2,6}(?:祖师|大师|尊者|菩萨|长老|禅师|居士)?)"
+    r'\s*(?:偈曰|偈云|偈|曰|说)\s*[：:]\s*[“"「『]([^”"」』\n]{8,200})'
+)
+# 《书名》同行引文，不需要动词：「《金刚经》"一切有为法…"」「闻《金刚经》至"应无所住
+# 而生其心"」。原先只认「云/曰」，这类引文一条都进不来。
+_QUOTE_TITLED = re.compile(r'《[^》\n]{2,30}》[^“"「『\n]{0,10}[“"「『]([^”"」』\n]{8,200})')
 _QUOTE_BOILER = re.compile(
     r"具格上师|亲近善知识|不可由文字|网络传授|须依止|本平台|不得对个体|面对面访谈"
     r"|如需深入学习|SuttaCentral|BDRC|fojin"
+    # 书单与指引句：「汉译可参《菩提道灯论》（任杰译）」「《清净道论》汉译：叶均居士
+    # 译本」「…可在 ajahnchah.org 免费下载」。它们写在引号里，却不是谁说过的话，
+    # 送去全文检索只会变成查无此句。2026-09-16 量出 22 条这样的行。
+    r"|可参|可详参|可阅|可查|查阅|译本|出版社|下载|开示全集|不可不读|逐句观照"
 )
 _QUOTE_PARAPHRASE = re.compile(r"转述|非原文|主旨|整理|概括|要旨|讲解")
+# 讲「这句话该不该引、该怎么标」的行，本身不是引文：纠错说明（「常被当作龙树的话
+# 引用，但《大智度论》中没有」）、禁用示例（「不可用宗喀巴的精确分判作为阿底峡立场」）。
+# 收了它们，周检会对一条文档已经查明并改正的记录拉响假警报。
+# 标记必须是关于**引用行为**的成句短语：试过「勿」「不可用」这类泛词，会误伤《坛经》
+# 「勿使惹尘埃」、罗什「且勿急」、虚云「不可用意识思量卜度」这些真引文（2026-09-16 实测）。
+# 「中没有」同样要紧跟书名号，否则撞上「心中没有」之类的寻常行文。
+_QUOTE_META = re.compile(
+    r"常被当作|误传|讹传|应保守表述|不要把|不得加引号|引用规范|disclaimer|》中没有|》中查无"
+)
 _QUOTE_HAN = re.compile(r"[\u3400-\u9fff]")
 
 
@@ -987,9 +1085,20 @@ def collect_persona_quotes() -> list[tuple[str, str, str]]:
         ):
             where_base = path.relative_to(base).as_posix()
             for number, line in enumerate(path.read_text(encoding="utf-8").splitlines(), 1):
-                if "出处" in line or "引用格式" in line or _QUOTE_PARAPHRASE.search(line):
+                if (
+                    "出处" in line
+                    or "引用格式" in line
+                    or _QUOTE_PARAPHRASE.search(line)
+                    or _QUOTE_META.search(line)
+                ):
                     continue
-                for pattern, needs_title in ((_QUOTE_SAMPLE, False), (_QUOTE_BLOCK, False), (_QUOTE_SAID, True)):
+                for pattern, needs_title in (
+                    (_QUOTE_SAMPLE, False),
+                    (_QUOTE_BLOCK, False),
+                    (_QUOTE_SAID, True),
+                    (_QUOTE_ATTRIBUTED, False),
+                    (_QUOTE_TITLED, False),
+                ):
                     match = pattern.search(line)
                     if not match:
                         continue
@@ -1101,6 +1210,100 @@ def classify_persona_quotes(
     return mismatched, unknown
 
 
+COMPILED_SOURCES_FILE = Path(__file__).parent / "compiled-teaching-sources.json"
+
+
+def compiled_teaching_corpora() -> dict[str, dict]:
+    """{祖师目录: 该祖师在 CBETA 之外、有免费全文可取的编集语录}。
+
+    3h 只能查 CBETA，所以虚云的《法汇》、印光的《文钞》一律落进「未判定」——
+    2026-09-15 修掉的那批拼接引文正是长在这个盲区里。清单把原书地址登记下来，
+    3i 就能真的进原书逐字找。
+    """
+    if not COMPILED_SOURCES_FILE.exists():
+        return {}
+    data = json.loads(COMPILED_SOURCES_FILE.read_text(encoding="utf-8"))
+    return {str(entry["master"]): entry for entry in data.get("corpora") or []}
+
+
+def _han_only(text: str) -> str:
+    """只留汉字。两边都这么归一，标点和空白的写法差异就不会造成假的对不上。"""
+    return "".join(_QUOTE_HAN.findall(re.sub(r"<[^>]+>", "\n", text)))
+
+
+def fetch_compiled_text(url: str, encoding: str = "utf-8") -> str | None:
+    """取一部编集语录的全文，归一成纯汉字；取不到返回 None（未知，不是「原书没有」）。"""
+    import urllib.error
+    import urllib.parse
+    import urllib.request
+
+    # urllib 不像 curl 会自己处理非 ASCII 路径，原样传中文路径会抛 UnicodeEncodeError。
+    split = urllib.parse.urlsplit(url)
+    safe = urllib.parse.urlunsplit(split._replace(path=urllib.parse.quote(split.path)))
+    try:
+        with urllib.request.urlopen(safe, timeout=90) as response:
+            raw = response.read()
+    except (urllib.error.URLError, OSError, ValueError):
+        return None
+    return _han_only(raw.decode(encoding, errors="replace"))
+
+
+def classify_compiled_teaching_quotes(
+    quotes: list[tuple[str, str, str]],
+    corpora: dict[str, dict],
+    fetch,
+) -> tuple[list[tuple[str, str, str]], list[tuple[str, str]], list[tuple[str, str]], list[str]]:
+    """到编集语录原书里逐字找人设当原话引的句子。
+
+    判「原书没有这句」只对 coverage 标 `complete` 的语料成立：正编、续编、三编就是
+    《文钞》的全部，都取得到，找不到即伪造。虚云标 `partial` —— 净慧编的《开示录》
+    比岑学吕的《法汇》多出六十余万字，BFNN 上没有，找不到只说明这一步够不着。
+    任何一部取不到，也一律记未判定：接口不通不是证据。
+
+    第四个返回值是「整部语料一篇都没取到」的祖师 —— 那说明这一步对他什么也没检查。
+    不把它单独报出来，一个取数早就坏掉的 3i 会年复一年地绿着，跟没有这道门禁一样。
+    """
+    mismatched: list[tuple[str, str, str]] = []
+    verified: list[tuple[str, str]] = []
+    unknown: list[tuple[str, str]] = []
+    bodies: dict[str, str | None] = {}
+    touched: set[str] = set()
+    for where, master, quote in quotes:
+        corpus = corpora.get(master)
+        if not corpus:
+            continue
+        touched.add(master)
+        wanted = _han_only(quote)
+        if len(wanted) < 8:
+            unknown.append((where, "quote too short to search"))
+            continue
+        found_in, unreachable = None, []
+        for text in corpus.get("texts") or []:
+            url = str(text.get("url"))
+            if url not in bodies:
+                bodies[url] = fetch(url, text.get("encoding") or "utf-8")
+            body = bodies[url]
+            if body is None:
+                unreachable.append(str(text.get("title")))
+            elif wanted in body:
+                found_in = str(text.get("title"))
+                break
+        if found_in:
+            verified.append((where, found_in))
+        elif unreachable:
+            unknown.append((where, f"could not read {', '.join(sorted(set(unreachable)))}"))
+        elif corpus.get("coverage") == "complete":
+            mismatched.append((where, quote, str(corpus.get("corpus_title") or master)))
+        else:
+            unknown.append((where, f"{master}'s free full texts do not cover every declared compilation"))
+    unreadable = sorted(
+        str(corpora[master].get("corpus_title") or master)
+        for master in touched
+        if all(bodies.get(str(t.get("url"))) is None for t in corpora[master].get("texts") or [])
+    )
+    return mismatched, verified, unknown, unreadable
+
+
 def verify_ids(bridge, cbeta_map: dict[str, list[str]], titles: dict[str, str]) -> dict[str, dict]:
     """Verify all CBETA IDs and return {full_cbeta_id: {text_id, short_id, title, ...}}.
 
@@ -1435,6 +1638,43 @@ def _run_legacy_link_verification(*, fix: bool) -> int:
     if not quote_line_mismatched and not quote_line_unknown:
         print(f"  All {len(persona_quotes)} quoted lines are in CBETA, in a work the persona declares")
 
+    # Step 3i: 3h 够不着的那些 —— 祖师自己的语录本就在 CBETA 之外，进原书逐字找
+    # （见 classify_compiled_teaching_quotes）。
+    print("\n[3i/4] Checking quoted lines against compiled teachings CBETA does not hold...")
+    compiled_corpora = compiled_teaching_corpora()
+    compiled_mismatched, compiled_verified, compiled_unknown, compiled_unreadable = (
+        classify_compiled_teaching_quotes(persona_quotes, compiled_corpora, fetch_compiled_text)
+    )
+    for where, quote, corpus_title in compiled_mismatched:
+        print(f"    [WRONG] {where}: {corpus_title} has no 「{quote[:40]}」")
+    for title in compiled_unreadable:
+        print(f"    [BROKEN] {title}: not one declared full text loaded — this step checked nothing")
+    if compiled_verified:
+        print(f"  Verified {len(compiled_verified)} quoted line(s) in the compiled teachings:")
+        for where, title in compiled_verified:
+            print(f"    {where}: {title}")
+    if compiled_unknown:
+        print(f"  Could not check {len(compiled_unknown)} quoted line(s) — unknown, not wrong:")
+        for where, reason in compiled_unknown:
+            print(f"    {where}: {reason}")
+    if not compiled_corpora:
+        print("  No compiled-teaching corpora are declared")
+
+    # 同一步里的第二类：「原典」块本身。3f 只认带 CBETA 经号的引用格式，《文钞》
+    # 没有经号，这些块此前对每一步检查都不可见（见 collect_compiled_excerpt_blocks）。
+    compiled_blocks = collect_compiled_excerpt_blocks()
+    block_mismatched, block_verified, block_unknown = classify_compiled_excerpt_blocks(
+        compiled_blocks, compiled_corpora, fetch_compiled_text
+    )
+    for where, segment, citation in block_mismatched:
+        print(f"    [WRONG] {where}: {citation} has no 「{segment[:40]}」")
+    if block_verified:
+        print(f"  Verified {len(block_verified)} 「原典」 block(s) word for word in the compiled teachings")
+    if block_unknown:
+        print(f"  Could not check {len(block_unknown)} 「原典」 block(s) — unknown, not wrong:")
+        for where, reason in block_unknown:
+            print(f"    {where}: {reason}")
+
     # Step 4: Update URLs
     # Build replacement map: full_cbeta_id -> str(internal_text_id)
     id_replacement_map: dict[str, str] = {}
@@ -1483,6 +1723,9 @@ def _run_legacy_link_verification(*, fix: bool) -> int:
     print(f"  Excerpt quotes not in the cited text: {len(quote_mismatched)}")
     print(f"  BDRC records that do not match: {len(bdrc_mismatched)}")
     print(f"  Quoted lines CBETA does not have: {len(quote_line_mismatched)}")
+    print(f"  Quoted lines the compiled teachings do not have: {len(compiled_mismatched)}")
+    print(f"  Compiled teaching corpora that could not be read: {len(compiled_unreadable)}")
+    print(f"  Excerpt blocks the compiled teachings do not have: {len(block_mismatched)}")
     if unknown_to_cbeta:
         print(f"  CBETA unreachable for:     {len(unknown_to_cbeta)} (not counted as wrong)")
     if dry_run and all_changes:

package/tools/version_manager.py

@@ -67,7 +67,14 @@ def rollback(teacher_dir: str, target_version: str) -> bool:
 
 
 def cleanup_old_versions(teacher_dir: str) -> int:
-    """Remove old versions beyond MAX_VERSIONS limit."""
+    """Remove archived versions beyond MAX_VERSIONS, oldest first.
+
+    Prunes only what `list_versions` reports — directories named `v…` — and never a
+    `_before_rollback` copy. The first version selected *every* directory under
+    `versions/`, so anything else a maintainer kept there was deleted once the
+    archive passed the limit, and the backup `rollback` writes so that a bad
+    rollback can be undone was evicted by age like an ordinary version.
+    """
     versions_dir = os.path.join(teacher_dir, "versions")
     if not os.path.exists(versions_dir):
         return 0
@@ -75,8 +82,11 @@ def cleanup_old_versions(teacher_dir: str) -> int:
     entries = []
     for entry in os.listdir(versions_dir):
         entry_path = os.path.join(versions_dir, entry)
-        if os.path.isdir(entry_path):
-            entries.append((entry_path, os.path.getmtime(entry_path)))
+        if not os.path.isdir(entry_path):
+            continue
+        if not entry.startswith("v") or entry.endswith("_before_rollback"):
+            continue
+        entries.append((entry_path, os.path.getmtime(entry_path)))
 
     entries.sort(key=lambda x: x[1], reverse=True)