ja: handle furigana — two kinds, opposite treatment (closes #704)

[christopherthompson81]

Closes #704.

Before: a parenthesised reading was phonemized as ordinary text after the kanji it annotates, so the reading came out twice — 漢字（かんじ） read känd͡ʑi känd͡ʑi, where a reader says it once.

The issue proposed one rule; implementing it turned up a distinction worth splitting on

True ruby positioning is not plain text. What reaches a phonemizer is one of several flattenings of markup — and they don't all mean the same thing.

DECLARED — Unicode interlinear annotation (U+FFF9 base U+FFFA ruby U+FFFB, defined for exactly this and discouraged for interchange, so rare) and the aozora ｜base《ruby》. These say they are ruby, so the author has stated the reading: the ruby wins and the base is dropped.

｜日本《にっぽん》  →  nippo̞ꜜɴ      (overriding the default にほん)
日本にっぽん   →  nippo̞ꜜɴ

That's the entire reason for writing it.

PARENTHESISED — base（ruby） / base(ruby). A convention, not a declaration; the same shape is an ordinary parenthetical. So the guard is equality with the computed reading, and a mismatch means "keep both" rather than "the author is overriding".

The must-not-change cases all survive

None equals the computed reading, so none is touched:

input	output	why
日本（にほんじん）	niho̞ꜜɴ niho̞nd͡ʑiɴ	a gloss
日本（にっぽん）	niho̞ꜜɴ nippo̞ꜜɴ	an alternate reading in the ambiguous form
会議（ミーティング）	käꜜiɡi miːtiŋɡɯᵝ	a katakana gloss of a loanword

⚠ Note 日本（にっぽん） and ｜日本《にっぽん》 now differ, and that is the point: the parenthesised one is a parenthetical the author may have meant as a contrast; the declared one is an instruction.

The test asserts the property that actually matters

Not "the output looks right" but: an annotated sentence is byte-identical to the same sentence written without the annotation — including at the token boundary the rewrite creates.

That caught the one thing worth checking. Dropping the annotation joins 東京 directly to に, so the result has to equal 東京に行く:

東京（とうきょう）に行く  →  to̞ːkʲo̞ːni ikɯᵝ
東京に行く              →  to̞ːkʲo̞ːni ikɯᵝ   ✓

Checks

• npm run typecheck clean
• 3114 → 3118 tests passing

🤖 Generated with Claude Code

https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr