Closes #145

What broke

The engine produces different line breaks than the browser when CJK text is followed by opening brackets:

Input: "서울(Seoul)과 부산(Busan)" at 180px

Browser:   서울           ✓  ( goes to next line with its content
           (Seoul)과

Pretext:   서울(          ✗  ( stuck to preceding CJK
           Seoul)과

Same bug for all CJK languages: 東京(Tokyo), 北京(Beijing), 인공지능(AI).

Why this only breaks CJK, not English

English and CJK go through different segmentation paths after the analysis merge:

• English: AB(CD) → [AB(] [CD)] — even though ( is merged backward, the line breaker can still break within AB( at character boundaries. No visible problem.
• CJK: 서울( enters buildBaseCjkUnits() which splits by character → [서] [울] [(] — the ( becomes an isolated unit, disconnected from Seoul). The meaningful group (Seoul) is broken apart, and the line breaker can't reconnect them.

Why it broke

The first-pass merge in src/analysis.ts (~line 1006) has a rule: "non-word-like punctuation sticks to the previous text segment." This correctly handles closing punctuation (), ., ,) but also catches opening brackets ((, [, {) via isEscapedQuoteClusterSegment() matching all kinsokuEnd characters.

This merge runs before the forward-sticky pass, so ( gets consumed backward before it can be attached forward to Seoul):

Word segmenter:     서울 | ( | Seoul | ) | 과
leftSticky merge:   서울 | ( | Seoul) | 과       ← ) sticks to Seoul (correct)
first-pass merge:   서울( | Seoul) | 과           ← BUG: ( sticks backward
forward-sticky:     (nothing — ( already gone)

How it was fixed

Added !mergedContainsCJK[prevIndex] guard at two merge points in src/analysis.ts:

1. First-pass merge (~line 1008) — the actual fix
2. Escaped-quote merge (~line 1057) — belt-and-suspenders on the same path

When the previous segment contains CJK, skip the backward merge. The bracket reaches the forward-sticky pass which correctly attaches it forward:

first-pass merge:   SKIPPED (prev is CJK)
forward-sticky:     서울 | (Seoul) | 과           ✓

Why this approach

• 2 lines of guard conditions — not a restructure, minimal risk
• Non-CJK unchanged — English AB(CD) still works as before
• Root cause fix — the backward merge was the wrong operation for brackets after CJK; the forward-sticky pass already does the right thing when brackets reach it

Tests

• bun test — 88/88 pass
• bun run scripts/cjk-bracket-check.ts --browser=chrome — 10/10 pass

  ✓ A1: Korean parenthesized English       ✓ A6: Chinese abbreviation bracket
  ✓ A2: Japanese parenthesized English     ✓ A7: Korean square brackets
  ✓ A3: Chinese parenthesized English      ✓ A8: Japanese square brackets
  ✓ A4: Korean abbreviation bracket        ✓ A9: Korean curly braces
  ✓ A5: Japanese abbreviation bracket      ✓ A10: Mixed CJK + nested brackets