From 5fdd7bcf74c75c503f56118dad54a77362530d18 Mon Sep 17 00:00:00 2001
From: ihsieh31 <410110@email.wlsh.tyc.edu.tw>
Date: Sun, 25 Oct 2026 16:41:25 +0800
Subject: [PATCH] fix(script): keep a comma-grouped number and its hyphenated
 suffix whole (#391)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The number alternative of LEXICAL_UNIT only took numeric hyphenated
suffixes, and its lookahead rejects stopping before -letters, so a
comma group followed by a word fell back to the bare prefix and split
at the comma: 1,000-word became 1, and 000-word, which also let a Cue
break between them through (SCRIPT_CAPTION_BREAK_TOKEN_BOUNDARY never
fired).

After its numeric suffixes the number alternative now takes one
hyphenated word suffix, with the ' ’ . - joins the plain-word
alternative allows. The joins matter: the number alternative comes
first, so it now reaches compounds the word alternative used to own,
and without them 90-year-old's or 2020-word.com would be cut at the
apostrophe or dot. 1,000-word, 123,456,789-point and 1,000.5-word each
stay one unit, every un-grouped compound tokenizes as before, and the
regression test pins both.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 packages/script/src/lexical.ts                |  2 +-
 packages/script/test/prose-boundaries.test.ts | 19 +++++++++++++++++++
 2 files changed, 20 insertions(+), 1 deletion(-)

diff --git a/packages/script/src/lexical.ts b/packages/script/src/lexical.ts
index b54a8d26c..19481df9e 100644
--- a/packages/script/src/lexical.ts
+++ b/packages/script/src/lexical.ts
@@ -7,7 +7,7 @@ const CHARACTER_UNIT = String.raw`[\p{Script=Han}\p{Script=Hiragana}\p{Script=Ka
 // A Latin name beside Han prose ends at the script boundary, even without an authored space.
 const WORD_CHARACTER = String.raw`(?:(?!${CHARACTER_UNIT})[\p{L}\p{M}\p{N}])`;
 const LEXICAL_UNIT = new RegExp([
-  String.raw`(?:\p{N}{1,3}(?:[,，]\p{N}{3})+|\p{N}+)(?:[.．]\p{N}+)?(?:-\p{N}+(?:[.．]\p{N}+)?)*(?!\p{N}|-[\p{L}\p{M}])`,
+  String.raw`(?:\p{N}{1,3}(?:[,，]\p{N}{3})+|\p{N}+)(?:[.．]\p{N}+)?(?:-\p{N}+(?:[.．]\p{N}+)?)*(?:-${WORD_CHARACTER}+(?:['’.-]${WORD_CHARACTER}+)*)?(?!\p{N}|-[\p{L}\p{M}])`,
   String.raw`${CHARACTER_UNIT}\p{M}*`,
   String.raw`(?:(?!${CHARACTER_UNIT})[\p{L}\p{N}])${WORD_CHARACTER}*(?:['’.-]${WORD_CHARACTER}+)*`,
 ].join("|"), "gu");
diff --git a/packages/script/test/prose-boundaries.test.ts b/packages/script/test/prose-boundaries.test.ts
index 01e15e07b..47f42c370 100644
--- a/packages/script/test/prose-boundaries.test.ts
+++ b/packages/script/test/prose-boundaries.test.ts
@@ -17,6 +17,25 @@ test("annotations cannot create speech boundaries inside words or attached punct
   assert.equal(caption('hello,@{beat!} world'), 'hello, world');
 });
 
+test("comma-grouped numbers keep a hyphenated suffix as one word", () => {
+  // A comma group is unreachable for the plain-word alternative, so the compound used to
+  // fall back to the bare prefix and accept a Cue break between `1,` and `000-word`.
+  const parsed = parse('a 1,000-word essay works');
+  assert.deepEqual(parsed.tokens.map(token => token.text), ['a', '1,000-word', 'essay', 'works']);
+  assert.equal(parse('123,456,789-point').tokens[0]!.text, '123,456,789-point');
+  assert.equal(parse('1,000.5-word').tokens[0]!.text, '1,000.5-word');
+  assert.equal(parse('a 1，000-word essay').tokens[1]!.text, '1，000-word');
+  // The number alternative now reaches hyphenated words first, so it must keep the word
+  // alternative's apostrophe and dot joins rather than cutting `'s` off a compound.
+  for (const word of ["90-year-old's", "2020-word's", "20-something’s", '3-in-1\'s', '2020-word.com', "1,000-dollar's"]) {
+    assert.deepEqual(parse(`the ${word} here`).tokens.map(token => token.text), ['the', word, 'here'], word);
+  }
+  assert.equal(parse('a 1,000-word.').tokens[1]!.text, '1,000-word');
+  for (const body of ['a 1,||000-word essay works', 'a 123,||456,789-point here', 'a 2020-||word essay works']) {
+    assert.throws(() => parse(body), /SCRIPT_CAPTION_BREAK_TOKEN_BOUNDARY/u, body);
+  }
+});
+
 test("comments do not divide words, and source edits retain the comment", () => {
   const source = '<line>hel<!-- note -->lo 한<!-- 설명 -->글. @{beat!}</line>';
   const parsed = parseScript("boundary", source);