CJK: full-width punctuation renders as annotations, and half-width numbers/words get truncated to "…" #128

Summary

In Chinese scripts, full-width punctuation (，。：「」！？（）…) is rendered wrong, and half-width numbers or Latin words on the same line can be squeezed into … and disappear. Both come from the same root cause in MarqueeTextView.swift.

Environment

• Textream 1.7.1 (Homebrew cask), macOS 26.6.2, Apple Silicon
• Font: Sans, size LG, overlay: Pinned (notch width 410), mode: Word Tracking, speech locale: zh-TW

Steps to reproduce

1. Paste this script: 那一年他 36 歲，搬到新的城市，而且換過三份工作。
2. Start the prompter.

Expected: 那一年他 36 歲，搬到新的城市，而且換過三份工作。
Actual: 那一年他 ...歲，搬到新的城市，而且換過三份工作。 The 36 is replaced by an ellipsis, and every full-width punctuation mark is shown italic, dimmed, and preceded by an extra space.

Whether a number disappears depends on how full the line is, so it comes and goes when the notch width or font size changes.

Root cause

1. Full-width punctuation becomes a standalone "annotation" word.
Unicode.Scalar.isCJK does not cover CJK Symbols and Punctuation (U+3000–303F) or Halfwidth and Fullwidth Forms (U+FF00–FFEF). So splitTextIntoWords emits each punctuation mark as its own word, and isAnnotationWord flags it because it has no letters or numbers. It is then rendered with the annotation style (italic, cueUnreadOpacity).

Tokenizer output for the sample (〔〕 = flagged as annotation):

那|一|年|他|36|歲|〔，〕|搬|到|新|的|城|市|〔，〕|而|且|換|過|三|份|工|作|〔。〕

2. buildLines underestimates line width, so SwiftUI truncates a word.
buildLines measures each word with NSString.size(withAttributes:), but each word is rendered as a separate SwiftUI Text(word + " "). For a standalone full-width punctuation mark, the two disagree a lot. Measured on this machine with NSFont.systemFont(ofSize: 20, weight: .semibold):

When the real width exceeds containerWidth, the HStack compresses its children. Single CJK characters can't shrink, so the only multi-character word on the line (36, or a Latin word like Claude) is truncated to …. I reproduced this outside the app by rendering the same HStack(spacing: 0) of Text(word + " ") views, with .lineLimit(1), at the width buildLines computed. The result matches what the prompter shows: 那一年他 ...歲，…

Suggested fix

• In splitTextIntoWords, attach punctuation to the preceding CJK character (e.g. 歲，) instead of emitting it as its own word. At minimum, add U+3000–303F and U+FF00–FFEF to the CJK handling, so these marks are never treated as annotations.
• In buildLines, measure words the same way they are rendered, or round each word width up (ceil) and keep a small safety margin, so accumulated per-word rounding can't overflow the line.
• Optionally, add .fixedSize() to the word Text, so an overflow wraps to the next line instead of silently hiding script text.

Related: #60, #95 (other Chinese support issues).