ci: enforce ascii-only pr body and widen the self-check grep

.github/workflows/commits.yml

@@ -25,6 +25,11 @@ jobs:
           PR_TITLE: ${{ github.event.pull_request.title }}
         run: PYTHONPATH=. python3 scripts/check_pr_title.py
 
+      - name: Validate PR body
+        env:
+          PR_BODY: ${{ github.event.pull_request.body }}
+        run: PYTHONPATH=. python3 scripts/check_pr_body.py
+
   commit-messages:
     name: commit messages
     runs-on: ubuntu-latest

AGENTS.md

@@ -151,7 +151,7 @@ No other languages anywhere in the message — not just the subject; body and fo
 |---|---|
 | subject | English, lowercase start, ≤ 72 chars, no period |
 | body | **All English**; when citing a non-English plan / discussion, **translate** it, don't paste |
-| punctuation | No full-width punctuation (`：`,`，`,`。`,`「」`,`""` …), no `§`-numbering, no non-English path names; the latin part of a §N.M anchor is fine |
+| punctuation | **ASCII-only** — not just no full-width punctuation (`：`,`，`,`。`,`「」`,`""` …) but also no em-dash `—`, curly quotes, or ellipsis `…` (all non-ASCII, all rejected by CI); no `§`-numbering, no non-English path names; the latin part of a §N.M anchor is fine |
 | trailer | `Co-authored-by: ...` is ASCII by format |
 
 **Why:** Conventional-Commits tooling (commitlint / semantic-release / changelog generators) parses ASCII grammar and mis-lints on non-English text; cross-language reviewers and a public commit history both need English.
@@ -285,10 +285,11 @@ Filling rules:
 - internal branch names / commit-hash references (no reviewer context).
 
 **Preview-verification (required):**
-1. After drafting, **grep for full-width / non-English chars first**:
+1. After drafting, **grep for any non-ASCII char first** (the CI lints the whole message as ASCII-only via `commit_lint._is_ascii`, and the squash commit body is this PR description — so it must be ASCII too):
    ```bash
-   grep -cP "[\x{4E00}-\x{9FFF}]|[\x{3000}-\x{303F}]|[\x{FF00}-\x{FFEF}]" /tmp/pr_description.md
-   # must be 0
+   grep -nP "[^\x00-\x7F]" /tmp/pr_description.md
+   # must print nothing (0 matches). Catches em-dash/curly-quotes/ellipsis,
+   # not just CJK + full-width — a CJK-only pattern gives a false pass.
    ```
 2. show the full text for preview;
 3. only after the user edits/confirms, run `gh pr create --title "..." --body "$(cat /tmp/pr_description.md)"`;

scripts/check_pr_body.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+import os
+import sys
+
+from scripts.commit_lint import check_pr_body
+
+
+def main() -> int:
+    body = os.environ.get("PR_BODY", "")
+    result = check_pr_body(body)
+    if result.ok:
+        return 0
+
+    for lineno, line in enumerate(body.splitlines(), start=1):
+        if any(ord(ch) > 0x7F for ch in line):
+            print(f"Invalid PR body line {lineno}: {line}", file=sys.stderr)
+    for error in result.errors:
+        print(f"- {error}", file=sys.stderr)
+    return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/commit_lint.py

@@ -42,6 +42,18 @@ def check_pr_title(title: str, config: CommitLintConfig | None = None) -> LintRe
     return _check_message(title, subject_limit=(config or CommitLintConfig()).pr_title_subject_limit)
 
 
+def check_pr_body(body: str) -> LintResult:
+    """The PR body becomes the squash commit body, so it must be ASCII-only.
+
+    Only the ASCII rule applies (the body is free-form Markdown, not a
+    Conventional-Commits header), so the header/subject checks are skipped.
+    """
+    errors: list[str] = []
+    if not _is_ascii(body):
+        errors.append("must be ASCII-only English")
+    return LintResult(errors)
+
+
 def _check_message(message: str, *, subject_limit: int) -> LintResult:
     errors: list[str] = []
     text = message.strip()

tests/test_commit_lint.py

@@ -7,6 +7,7 @@
 from scripts.commit_lint import (
     CommitLintConfig,
     check_commit_message,
+    check_pr_body,
     check_pr_title,
 )
 
@@ -73,6 +74,25 @@ def test_pr_title_rejects_subject_over_pr_limit() -> None:
     assert "subject must be 90 characters or fewer" in result.errors
 
 
+def test_pr_body_accepts_ascii_markdown() -> None:
+    result = check_pr_body("## Summary\n\nPlain ASCII body - fine.\n")
+
+    assert result.ok
+
+
+def test_pr_body_rejects_em_dash() -> None:
+    # The em-dash (U+2014) is the char that slipped past the CJK-only grep and
+    # failed the post-merge commit lint once the PR body became the squash body.
+    result = check_pr_body("drops the hint only — empty submits still ignored")
+
+    assert not result.ok
+    assert "must be ASCII-only English" in result.errors
+
+
+def test_pr_body_allows_empty() -> None:
+    assert check_pr_body("").ok
+
+
 def test_commit_msg_hook_runs_commitlint_before_python_checker(monkeypatch, tmp_path: Path) -> None:
     message_file = tmp_path / "COMMIT_EDITMSG"
     message_file.write_text("docs: add launch notes\n")