diff --git a/tests/test_cjk_typography.py b/tests/test_cjk_typography.py
new file mode 100644
index 0000000..5a2bcc2
--- /dev/null
+++ b/tests/test_cjk_typography.py
@@ -0,0 +1,204 @@
+#!/usr/bin/env python3
+"""Guard the book's two Chinese-typography conventions.
+
+Neither rule is enforced anywhere else: ``check_project_rules.py`` looks at
+fences and links, ``check_emphasis.py`` at bold markers, and the workspace
+``format_checker.py`` at quotes, fence languages and heading spacing. Intra-line
+typography had no gate at all, and it drifted exactly as far as you would
+expect -- one sibling book reached 45% unspaced CJK/Latin boundaries and another
+reached a 50/50 split on parentheses, both mixed *within single files*. A
+one-off sweep normalised all fourteen books; this file is what stops the next
+paragraph from drifting back.
+
+RULE 1 -- a CJK ideograph may not sit directly against ``[A-Za-z0-9]``.
+         ``Harness的定义`` / ``第4章`` are violations; ``Harness 的定义`` is not.
+
+RULE 2 -- a parenthesis whose content contains any CJK must be full-width.
+         ``(用户命名空间映射)`` is a violation; ``（用户命名空间映射）`` is not.
+         Latin-only parentheses are NOT touched: ``新建一个层 (Layer)`` and the
+         ``项目(Agent)的经验`` hugging style both stay legal, because their
+         content has no CJK.
+
+Deliberately NOT checked, because the whole cluster agrees on the opposite:
+Chinese punctuation against Latin (``（如 UI、监控服务）``) needs no space --
+full-width punctuation already carries its own visual padding.
+
+Everything a reader never sees as prose is excluded before the scan: fenced code
+blocks and display-math blocks are skipped whole, and inline code, math, link
+and image targets, HTML and bare URLs are replaced with Private Use Area
+placeholders -- characters that can never themselves be mistaken for a letter, a
+digit or an ideograph. Every mask pattern excludes that range in turn, because a
+greedy ``\\S+`` will otherwise swallow the placeholder its predecessor just left
+behind.
+"""
+
+from __future__ import annotations
+
+import re
+import unittest
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+SKIP_DIRS = {".git", "_book", "node_modules", ".obsidian", "output", ".agent", "_site"}
+
+PUA_FIRST = 0xE000
+PUA = "\\ue000-\\uf8ff"
+CJK = r"一-鿿㐀-䶿"
+LATIN = r"A-Za-z0-9"
+
+FENCE = re.compile(r"^\s*(?:```|~~~)")
+MATH_BLOCK = re.compile(r"^\s*\$\$\s*$")
+LINK_DEF = re.compile(r"^\s*\[[^\]]+\]:\s*\S+")
+
+MASKS = (
+    re.compile(r"<!--.*?-->"),
+    re.compile(r"``[^`]+``"),
+    re.compile(r"`[^`]*`"),
+    re.compile(r"\$\$[^$]*\$\$"),
+    # only treat $...$ as math when it actually contains math syntax, so that a
+    # pair of prices ("$25/$125 per M tokens") is not read as one formula
+    re.compile(r"\$(?=[^$\n]*[\\^_{])[^$\n]*\$"),
+    re.compile(r"!?\]\([^)\s]*(?:\s+\"[^\"]*\")?\)"),
+    # a real tag, not "| <1% | - | <5 | 各类GPU | >400ms |"
+    re.compile(r"</?[A-Za-z][A-Za-z0-9-]*(?:\s[^<>]*)?/?>"),
+    re.compile("(?<![" + PUA + r"\w/])(?:https?|ftp)://[^\s" + PUA + "]+"),
+)
+
+ADJACENT = re.compile(f"[{CJK}][{LATIN}]|[{LATIN}][{CJK}]")
+HALF_PAREN = re.compile(r"\([^()]*\)")
+
+
+def mask(line: str) -> str:
+    counter = [0]
+
+    def repl(_match: "re.Match[str]") -> str:
+        counter[0] += 1
+        return chr(PUA_FIRST + counter[0] - 1)
+
+    for pattern in MASKS:
+        line = pattern.sub(repl, line)
+    return line
+
+
+def prose_lines(text: str):
+    lines = text.split("\n")
+    in_front = bool(lines) and lines[0].strip() == "---"
+    in_fence = False
+    in_math = False
+    for number, line in enumerate(lines, start=1):
+        if in_front:
+            if number > 1 and line.strip() == "---":
+                in_front = False
+            continue
+        if FENCE.match(line):
+            in_fence = not in_fence
+            continue
+        if in_fence:
+            continue
+        if MATH_BLOCK.match(line):
+            in_math = not in_math
+            continue
+        if in_math or LINK_DEF.match(line):
+            continue
+        yield number, line
+
+
+def markdown_files(root: Path = ROOT):
+    for path in sorted(root.rglob("*.md")):
+        if SKIP_DIRS.isdisjoint(path.relative_to(root).parts):
+            yield path
+
+
+def unspaced(text: str):
+    """RULE 1 violations."""
+    for number, line in prose_lines(text):
+        for match in ADJACENT.finditer(mask(line)):
+            yield number, match.group(0), line.strip()[:100]
+
+
+def half_width_parens(text: str):
+    """RULE 2 violations."""
+    for number, line in prose_lines(text):
+        for match in HALF_PAREN.finditer(mask(line)):
+            if re.search(f"[{CJK}]", match.group(0)):
+                yield number, match.group(0)[:40], line.strip()[:100]
+
+
+class CjkTypographyTests(unittest.TestCase):
+    def test_cjk_and_latin_stay_apart(self) -> None:
+        found = [
+            f"{path.relative_to(ROOT)}:{number}: {hit!r} in {snippet}"
+            for path in markdown_files()
+            for number, hit, snippet in unspaced(path.read_text(encoding="utf-8"))
+        ]
+        self.assertEqual(
+            [], found, "缺少中文与拉丁字符之间的半角空格：\n" + "\n".join(found[:40])
+        )
+
+    def test_parens_holding_chinese_are_full_width(self) -> None:
+        found = [
+            f"{path.relative_to(ROOT)}:{number}: {hit!r} in {snippet}"
+            for path in markdown_files()
+            for number, hit, snippet in half_width_parens(path.read_text(encoding="utf-8"))
+        ]
+        self.assertEqual(
+            [], found, "含中文的括号应使用全角：\n" + "\n".join(found[:40])
+        )
+
+    def test_both_rules_can_actually_fail(self) -> None:
+        """A check that cannot fail is not evidence."""
+        self.assertEqual([], list(unspaced("使用 Harness 的记忆子系统。\n")))
+        self.assertEqual(
+            [(1, "s的", "使用 Harness的记忆子系统。")],
+            list(unspaced("使用 Harness的记忆子系统。\n")),
+        )
+        self.assertEqual([(1, "第4", "第4 章")], list(unspaced("第4 章\n")))
+        self.assertEqual([], list(half_width_parens("启用（用户命名空间映射）。\n")))
+        self.assertEqual(
+            [(1, "(用户命名空间映射)", "启用 (用户命名空间映射)。")],
+            list(half_width_parens("启用 (用户命名空间映射)。\n")),
+        )
+
+    def test_latin_only_parens_are_left_alone(self) -> None:
+        for sample in (
+            "在基础镜像之上新建一个层 (Layer)。",
+            "多个生产级智能体项目(Agent)的经验表明。",
+            "定义一组相关联的应用容器为一个项目 (project)。",
+        ):
+            self.assertEqual([], list(half_width_parens(sample + "\n")), sample)
+
+    def test_non_prose_regions_are_excluded(self) -> None:
+        for sample in (
+            "见 `mini_harness/核心.py` 说明。",
+            "参考 [第 4 章的运行时](04_runtime/README.md)。",
+            "地址是 https://example.com/中文path 这一条。",
+            '<img src="a.png" alt="第4章示意图"> 的属性不算正文。',
+            "见 [文档](a.md) https://e.com/x 结束。",
+            "公式 $$L = (预测 - 真实)^2$$ 属于数学。",
+            "记 $a^*$ 与 $O_t$ 为最优解。",
+        ):
+            self.assertEqual([], list(unspaced(sample + "\n")), sample)
+            self.assertEqual([], list(half_width_parens(sample + "\n")), sample)
+
+    def test_prices_are_not_mistaken_for_math(self) -> None:
+        """A pair of '$' around prose must not mask the prose between them."""
+        line = "定价 $25/$125 per M tokens，另有 $1 亿额度给 Glasswing计划。\n"
+        self.assertEqual([(1, "g计", line.strip())], list(unspaced(line)))
+
+    def test_html_text_content_is_still_prose(self) -> None:
+        self.assertEqual(
+            [(1, "文a", "<span>中文abc</span> 属于 HTML。")],
+            list(unspaced("<span>中文abc</span> 属于 HTML。\n")),
+        )
+
+    def test_a_table_row_is_not_an_html_tag(self) -> None:
+        line = "| 33-64 | 其他矿工 | <1% | - | <5 | 各类GPU | >400ms |\n"
+        self.assertEqual([(1, "类G", line.strip())], list(unspaced(line)))
+
+    def test_display_math_blocks_are_skipped_whole(self) -> None:
+        text = "正文一行。\n\n$$\n\\text{原始Adam的权重衰减} = \\eta \\lambda \\theta\n$$\n\n正文二行。\n"
+        self.assertEqual([], list(unspaced(text)))
+
+
+if __name__ == "__main__":
+    unittest.main()