[rahul05ranjan]

What

Prose analysis split a comma-grouped number at the comma when a hyphenated word followed: a 1,000-word essay became 1, + 000-word (two caption/speech words), and an inner Cue break a 1,||000-word essay was accepted. Plain 2020-word and standalone 1,000 were already correct.

Why

In LEXICAL_UNIT, the number alternative matched 1,000 but its trailing lookahead (?!\p{N}|-[\p{L}\p{M}]) failed before -word. The comma-group alternative then died entirely, so the bare \p{N}+ form matched just 1 and the word alternative re-matched 000-word.

Fix

packages/script/src/lexical.ts — split the number alternative into two:

• a comma-grouped number now keeps a trailing hyphenated word suffix in the same unit (mirroring what the word alternative already does for 2020-word), so 1,000-word, 1，000-word, and 123,456,789-point are one unit;
• bare numbers keep the existing lookahead, so 2020-word still falls through to the word alternative unchanged.

Standalone 1,000, 1,234.56, and list-like comma behavior are untouched.

Tests

Added to packages/script/test/script.test.ts (tokens + caption words for ASCII comma, full-width comma, and multi-group cases) and packages/script/test/prose-boundaries.test.ts (1,||000-word now throws SCRIPT_CAPTION_BREAK_TOKEN_BOUNDARY like 2020-||word). Both files pass via node --test (23/23); both new cases fail on unfixed main.

Fixes #391