@kensio/pinyinjs 1.6.0 → 1.7.1

package/CHANGELOG.md

@@ -1,634 +0,0 @@
-# Changelog
-
-Notable changes to `@kensio/pinyinjs`. Versions follow
-[semantic versioning](https://semver.org) from 1.0.0 onward.
-
-Two things are outside the version contract, for the reasons
-[the API page](docs/api/#stability) gives: the readings a conversion produces,
-which a dictionary rebuild can change, and the artifact format under `data/`.
-
-## Unreleased
-
-## 1.6.0
-
-### Added
-
-- **Hanzi and pinyin together, rather than one instead of the other.**
-  `convertToAnnotatedHtml` writes the characters with their reading above them,
-  as `<ruby>` markup — the output every other mode could not give, and what a
-  learner's text, a subtitle and a dictionary entry all want. `convert` and
-  `convertToHtml` write the reading _instead of_ the characters.
-  `toAnnotatedHtml(pieces)` renders pieces you already have, and the command
-  line has `pinyinjs annotate`.
-
-  Each `<rt>` holds exactly what `convertToHtml` would have written, so tone
-  classes and uncertainty marking work inside an annotation as they do outside
-  one. `<rp>` gives the parenthesised fallback, the `<ruby>` carries `lang="zh"`
-  and the syllables inside it keep `zh-Latn-CN-pinyin`, since on an English page
-  a screen reader needs to be told both.
-
-  **A base is not always one character**, which is what makes this harder than
-  it looks and what every naive per-character annotation gets wrong. 玩儿 is two
-  characters and the one syllable `wánr`; 95% is `bǎifēnzhījiǔshíwǔ` over three
-  written characters in the other order; 1988 is `yī jiǔ bā bā` over one. All
-  are annotated whole. `ConvertedPiece` gains a `source` — the characters a
-  piece reads, or undefined where it reads on into the ones named before it —
-  which is what makes that expressible.
-
-  Pinyin orthography stays out of the hanzi it annotates: no word spaces, since
-  Chinese is not written with them; the hyphen of `gāngān-jìngjìng` sits in the
-  reading, being a boundary inside one word rather than a gap between two; and
-  the base keeps 。 where the conversion writes a full stop.
-
-- **`readings`: a conversion takes readings the caller asserts.** No rule
-  settles every polyphone, and some texts are genuinely ambiguous —
-  孩子越长越漂亮 grows where 头发越长越漂亮 lengthens, and nothing in the
-  characters says which. An application that knows its own content can now say
-  so, rather than the library guessing on its behalf.
-
-  ```ts
-  convert(dictionary, "这篇文章不太长。", { readings: { 太长: "tài cháng" } });
-  ```
-
-  A **word** hint asserts something about the text it names and rewrites exactly
-  those characters, so a bare `长` leaves 校长 as `xiàozhǎng` — the dictionary
-  knowing that word is better evidence about the stretch than a remark about one
-  of its characters, and that is what makes a corrections table safe to
-  accumulate. Naming the whole word does reach it, and keeps it whole:
-  `{ 银行: "yín xíng" }` gives `yínxíng`.
-
-  A **positional** hint asserts something about one character of one text, and
-  nothing outranks it, the enclosing word included. Positions count code points
-  from the start of the text, across any non-Han runs in it. The list form takes
-  both kinds together.
-
-  Hints enter as lattice edges rather than as rules, because the rules layer
-  cannot invent a reading no source attests — which is exactly what a hint is,
-  the difference being that a caller attests it. Spacing is untouched, an
-  unmarked syllable is 轻声, and a hint that cannot be parsed throws rather than
-  being skipped: a correction that silently does nothing is worse than one that
-  fails.
-
-### Fixed
-
-- **Polyphones defaulted to whichever reading a dictionary happened to print
-  first.** Unihan's `kTGHZ2013` ranked above `kMandarin`, and that field
-  prefixes every reading with its entry number in 《現代漢語規範詞典》 — 勒 is
-  `212.050:lè 212.100:lēi` — so its order is a page order, and taking the first
-  reading meant taking whichever one sorted earlier by pinyin. 殷 read `yān` and
-  楷 `jiē`. The dictionary-indexed fields now rank last: they are a good
-  _source_ of readings and a meaningless _ranking_ of them.
-
-  `kHanyuPinlu` still leads where it lists more than one reading, because there
-  it has counted them against each other, but where it lists one it counted
-  nothing — it is saying "this occurred", not "this leads" — and a lone reading
-  now ranks last too. Only ten characters have one that disagrees with
-  `kMandarin`, and eight of those are simply pre-1985: the corpus behind the
-  field, 《現代漢語頻率詞典》, predates the 普通话异读词审音表, and the 132
-  occurrences carrying 绩 as `jī` are the old standard rather than a case for
-  it.
-
-  354 readings fixed against 94 broken over 25 characters, the largest being 贾
-  `gǔ`→`jiǎ`, 苔 `tāi`→`tái`, 町 `dīng`→`tǐng`, 殷 `yān`→`yīn` and 芥
-  `gài`→`jiè`. CPP's polyphone score is 90.34%, up from 89.05%, and the gold
-  corpus does not move. What broke is mostly `kMandarin` being wrong with
-  nothing behind it — 媛 `yuán`→`yuàn`, 桔 `jú`→`jié`, 耶 `yē`→`yé`, 缪
-  `miào`→`móu` — which needs a third opinion with corpus evidence behind it
-  rather than an override, since here the sources disagree. 䘚 is now paired
-  with 卒, so `core` and `standard` each gained a key.
-
-- **`full` read 特徵 as `tèzhǐ` where `standard` read it right.**
-  `large_pinyin.txt` is a 简体 corpus, and where it carries a 繁體 headword
-  anyway it has read it as though the characters were 简体, taking each at face
-  value instead of as the 简体 character it stands for: 徵 as itself rather than
-  as 征, 沈 as the surname rather than as 沉, so 沈溺 came out `shěnnì`. Of the
-  2,854 corpus headwords carrying a 繁體-only character, 72 are ones CC-CEDICT
-  pairs with a 简体 form and 13 of those read wrongly.
-
-  The reading was only half of it. A corpus headword becomes an entry of its
-  own, and an entry outranks the 繁體 key that the traditional derivation hangs
-  on the 简体 entry — so the word was in the dictionary twice, and which copy
-  answered depended on the tier. `standard` excludes the phrase tail and was
-  right; `full`, the default, was wrong. Loading more dictionary made the answer
-  worse, which is also the upgrade path the docs describe, of converting with
-  `standard` and reloading as `full`.
-
-  A headword CC-CEDICT knows _only_ as another word's 繁體 spelling now
-  contributes no entry, and the word keeps the reading CC-CEDICT pairs it with.
-  68 entries drop out and their keys are re-claimed from the 简体 entry, so
-  `full` loses 2 keys; 简→繁→简 round trips gain 38 against 5 lost on 繁→简→繁,
-  both unchanged as percentages. The 2,782 headwords CC-CEDICT does not pair are
-  rare or mixed-script words whose characters do read them, and are left alone.
-  Three of the words are build assertions now, this being the kind of defect
-  that comes back when a rolling source is refreshed and nothing is watching.
-
-- **长 read `zhǎng` wherever no word covered it**, so 这篇文章不太长 came out
-  `bú tài zhǎng` and 要多长时间 `duō zhǎng shíjiān`. The stored default is not
-  the defect — Unihan ranks the readings `zhǎng(1879) cháng(1179)` and names
-  `zhǎng` in `kMandarin`, and that ranking is what reads 署长, 团长, 公安局长 and
-  总会长 correctly off a bare 长 at the end of a title. What was missing is that
-  nothing could prefer the alternate at a position no word covers.
-
-  A degree adverb settles it, and only from the left: a growing 长 is a verb and
-  no 很, 太, 最 or 多 can modify one, while what follows an adjectival 长 is a
-  noun, a particle or the end of the sentence — which is what follows half the
-  verbs too. 得, 着 and the 越…越 correlative are the exceptions, and are guarded;
-  了 and 的 are deliberately not, since after an adverb they are the sentence
-  particle and the attributive, making 时间太长了 and 很长的道路 both `cháng`.
-
-  Measured over the same 88,866 lines of Tatoeba and zh.wikipedia the other
-  edge rules were sized against, 260 长 decode as a word of their own and this
-  moves 75 to `cháng`, all 75 correctly. On CPP's 40 hand-labelled 长 the
-  character goes 85.00% to 87.50%; overall CPP is unmoved at 90.34%, as expected
-  of one character in 20,139.
-
-- **越长越高 read `yuè cháng yuè gāo`.** 越长 is a key read `yuè cháng` and the
-  only one of its shape — 越大, 越高, 越好 and 越快 are all absent, so 越高越好
-  decodes as two words while 越长越高 reached for a word nothing else in the
-  paradigm has. It carries no part of speech, which is how a reading somebody
-  asserted is held rather than a word anybody counted, and only
-  `large_pinyin.txt` has it. Where the far half of the correlative names
-  something growing produces — 高, 大, 胖, 壮, 结实 — the edge is dropped and the
-  character's own `zhǎng` stands.
-
-  Unlike everything else here this is a heuristic with no corpus behind it: 越长
-  occurs three times in the 88,866 lines and all three are 越来越长 or 说的越长.
-  It forbids rather than forces, so `cháng` stays on as a rival one bucket
-  dearer and the syllable still reports itself uncertain — 越长越X is genuinely
-  ambiguous, 孩子越长越漂亮 growing where 头发越长越漂亮 lengthens, which is why
-  漂亮 is not in the list. 时间越长越好 and 头发越长 are untouched.
-
-## 1.5.3
-
-### Fixed
-
-- **The changelog inside the package described five releases as unreleased.**
-  Every tarball from 1.1.0 to 1.5.2 shipped a `## Unreleased` section holding
-  everything since 1.0.0, so a reader who opened the copy they installed —
-  which [the API page](docs/api/#stability) points them at — had no way to tell
-  which version any of it arrived in. The entries are filed under the versions
-  that shipped them now. No code changes; this release exists to put the
-  corrected file on the registry.
-
-## 1.5.2
-
-### Fixed
-
-- **Third-tone sandhi ran over the syllables rather than over the feet.** The
-  rule is usually stated as "a third tone before another third tone is said as
-  a second", and a left-to-right scan of a reading applying exactly that is
-  wrong about as often as it is right: sandhi's domain is the prosodic foot,
-  which is built out of structure ([Shih 1986][1]). 這家銀行的行長很喜歡旅行
-  came out `hángzháng hén xǐhuan`, running the subject into the predicate, and
-  紙老虎 came out `zhíláohǔ`.
-
-  Three passes now, innermost first, each reading what the one before it left.
-  Inside a constituent, every third tone but the last lowers. Between the
-  constituents of a word, so that 展覽館 is 展覽 + 館 and stays `zhánlánguǎn`
-  while 紙老虎 is 紙 + 老虎 and becomes `zhǐláohǔ` — the dictionary is asked
-  where a word divides, and only proposes one where both halves are words. And
-  between a monosyllabic word and the word after it, which it leans on: 我也很好
-  is `wó yé hén hǎo`, but 行長 and 老闆 keep their second syllable.
-
-  `applySandhi` takes a `SandhiGrouping` for this, and without one still treats
-  the whole reading as a single word. `convert` passes what the decoder found;
-  the `sandhi` command splits its argument on whitespace.
-
-  What is given up is the monosyllable leaning backwards — 保管好 is
-  `báoguán hǎo` and comes out `báoguǎn hǎo` — which needs to know which way it
-  attaches, a question about syntax rather than about the words.
-
-  [1]: https://www.researchgate.net/publication/36071823_The_Prosodic_Domain_of_Tone_Sandhi_in_Chinese
-
-- **14 characters read wrongly under `zh-TW`.** CC-CEDICT states a `Taiwan pr.`
-  note either as a definition of its own, where it qualifies the headword's
-  reading, or parenthesised inside one sense, where it qualifies that sense —
-  and the merge took both as facts about the headword. 從's `zòng` is the 侍從
-  and 從兄弟 senses, all of them bound forms, so `我從北京來` read
-  `wǒ zòng Běijīng lái`; 會 took `huǐ` from 一會兒 and `很會說話` followed it.
-  A note now has to cover the sense the entry leads with, which is what a
-  character means with nothing to narrow it — 和's `hàn` is the conjunction, and
-  it stays. Also fixed for 勞, 燥, 行, 勝, 匹, 多, 抵, 枕, 比, 玩, 署 and 聽.
-  Compounds keep their own notes, so 肉燥麵 is still `ròusào miàn`.
-
-  Only a character is tested this way; a word is reached only where it is
-  written, so 相親 keeps the `xiàngqīn` its dominant sense carries. CC-CEDICT
-  now supplies 335 deltas rather than 349.
-
-## 1.5.1
-
-Tooling and test-suite work only; nothing a consumer installs changed.
-
-## 1.5.0
-
-### Added
-
-- **`convertToHtml` declares what language it is emitting.** Every syllable
-  element now carries `lang="zh-Latn-CN-pinyin"`, or `zh-Latn-TW-pinyin` where
-  the conversion read `zh-TW`. Nothing about `yín` on its own says it is
-  Mandarin in the Latin alphabet, so without the tag a screen reader pronounces
-  it as whatever the surrounding page claims to be — English, usually — and the
-  browser hyphenates it that way too. The subtags are all registered, and the
-  `pinyin` variant's prefix in the IANA registry is `zh-Latn`.
-
-  **This changes the markup existing callers get**, which is why it is worth
-  saying plainly: the classes are unchanged, and `lang` is a new attribute
-  between `class` and `data-alternatives`. `lang: false` — `--no-lang` at the
-  command line — leaves it off, for a caller who would rather declare the
-  language once on a wrapper of their own than repeat it on every syllable.
-
-## 1.4.0
-
-### Added
-
-- **`pinyinjs script`,** script conversion at the command line, with `--to` for
-  the target and `--from-script` where the text is too short for detection to
-  tell. The lines carry the converted text and nothing else so it pipes;
-  `--json` carries the per-character evidence and what was uncertain. Also a
-  [script conversion](docs/script-conversion/) page, which the README and the
-  docs index now point at — `toScript` shipped with one row in the API table
-  and nothing a reader would find.
-
-- **`toScript`,** 简体 ↔ 繁體 conversion. 我们后来发现了头发问题 is
-  我們後來發現了頭髮問題. The reading is the evidence, in both directions, and it
-  is what nothing else in this space has: 发 is 發 or 髮 and only `fà` tells them
-  apart; 干 is 幹, 乾 or 干; 乾 stays 乾 in 乾隆 and simplifies to 干 in 乾燥.
-  Every other converter resolves these from a phrase table alone. This one
-  segments with the lattice and reads the result, so it generalises to words no
-  phrase table lists.
-
-  Evidence is taken strongest first — a word some source wrote in both scripts,
-  then the character at the reading it was decoded with, then the character's
-  commonest form — and `toScriptPieces` reports which of those settled every
-  character. `isUncertainChoice` marks the last one, the only case that is a
-  guess: 下面 is a surface or a bowl of noodles and nothing here can tell, so it
-  says so rather than picking silently. Over the gold corpus 97.7% of characters
-  are locked and 1.5% are guesses.
-
-  Targets are BCP-47: `zh-Hans`, `zh-Hant`, `zh-Hant-TW`, `zh-Hant-HK`, with a
-  bare `zh-Hant` taking Taiwan. The input script is detected unless `from` names
-  it, which matters more than it sounds — plenty of characters are current in
-  both scripts, so running 繁體 through the 简→繁 tables would turn 准將 into
-  準將.
-
-  **This is orthography, not translation.** 軟體 and 软件 are different words
-  rather than two spellings of one. OpenCC folds that into script conversion
-  with `s2twp`; pinyinjs does not.
-
-- **`data/script.map`,** the conversion tables, as their own fetchable file:
-  282 KB, 96 KB brotli, loaded by `loadScriptTables` only when something asks
-  to convert. A caller converting hanzi to pinyin pays nothing for it. The
-  character tables alone are right for 99.13% of the 267,639 entries whose two
-  scripts differ, and the word exception lists cover the rest.
-
-- **`hansOnly` and `hantOnly`,** the character sets `detectScript` has always
-  taken and nothing shipped. They ride in the same file.
-
-- **Round-trip scoring in `pnpm accuracy`.** 繁→简 is near-deterministic, so
-  简→繁→简 has to be the identity for essentially every key — half a million
-  cases with no hand-labelling, and every failure a real defect. It reports
-  99.6% over the words in use and 99.3% over every key. The reverse trip is
-  lossy by design and is reported rather than targeted.
-
-- **Hong Kong 繁體 glyph forms are recognised on the lookup path.** Taiwan and
-  Hong Kong write the same characters with the same readings and disagree about
-  the standard form of 58 of them — 裡 against 裏, 群 against 羣, 麵 against 麪.
-  Hong Kong spellings were missing dictionary keys and falling back to
-  converting character by character, so 羣眾 and 麪包 read worse than 群眾 and
-  麵包 did. The lookup path now normalises 繁體 glyph forms to the Taiwan
-  standard the keys are already built from: measured over the 9,833 keys with a
-  distinct Hong Kong spelling, 2,358 were already keys and normalisation reaches
-  a further 5,439. Only the reading is taken from the normalised form — the
-  characters a caller gets back are the ones they wrote. `toCanonicalGlyphs`,
-  `toRegionalGlyphs` and the `Region` type are exported for it, and
-  SCRIPTS-AND-LOCALES.md carries the reasoning.
-
-  Nineteen of OpenCC's mappings are deliberately not applied, because the
-  variant is a live spelling rather than a regional one. Six are current 繁體 —
-  台 appears 164 times in CC-CEDICT against 臺's 208 — and thirteen are current
-  **简体**, because the PRC simplification adopted the same 新字形 conventions
-  Hong Kong did: 着, 温, 脱, 户, 税, 卧, 悦, 兑, 葱, 幺, 棱, 檐, 痹. Normalising
-  those would have rewritten 简体 text, and 走着 `zǒuzhe` is the case that
-  caught it.
-
-- **OpenCC as a data source.** Its 简体 ↔ 繁體 character tables and the Taiwan
-  and Hong Kong variant tables, Apache-2.0, attributed in `NOTICE`. The glyph
-  tables are held in `src/script/glyphs.ts` as code rather than shipped as an
-  artifact, since 120 mappings do not justify a download and they have to work
-  before any dictionary has loaded; `pnpm build:data` re-derives them from
-  OpenCC and fails the build if the two have drifted.
-
-### Fixed
-
-- **240 dictionary entries could not be looked up at all.** Recognising Hong
-  Kong glyph forms on the lookup path means a key written in a form the path
-  normalises _away from_ can never be found, and the merge derived 繁體 forms
-  per word from Unihan's variant lists with no single standard behind them — so
-  it wrote 裏 where the corpus overwhelmingly writes 裡, and 上樑, 中峯 and 义藴
-  became keys nothing could reach. The merge now derives canonical glyph forms,
-  and keys a headword's canonical spelling alongside it where the phrase corpus
-  wrote one with a 繁體 variant. **No key is now unreachable**, and Hong Kong
-  spellings resolve slightly better than before: 5,499 of the 9,978 keys with a
-  distinct Hong Kong form, against 5,439.
-
-  Two things had to be got right alongside it. The evidence for "简体 writes
-  this character" comes from CC-CEDICT entries whose two columns differ, not
-  from the phrase corpus, which is nominally 简体 but carries 峯, 藴 and 枱 of
-  its own — counting it cut the table from 50 mappings to 32 and stopped Hong
-  Kong text converting at all. And a 繁體 form differing from its 简体 one only
-  by glyph normalisation is not evidence about script, or the alias key would
-  put that entry's 简体 characters into the 繁體-only set, which is how 发
-  briefly stopped being 简体-only.
-
-  `toScript` drops the pass that was papering over this. Key counts move:
-  723,149 for `full`, 97,997 for `standard`, 16,975 for `core`.
-
-## 1.3.0
-
-### Added
-
-- **`slug`,** hanzi to a URL-safe slug: 我想学中文。 is
-  `wo3-xiang3-xue2-zhong1wen2`. Built on the decode rather than on a finished
-  string, which is the whole point of it — 银行 is `yin2hang2` and not
-  `yin2xing2`, the separators fall on the word boundaries 分词连写 found, and
-  西安 keeps its 隔音符号 as a break so a toneless slug is `xi-an` rather than
-  `xian`, which is a different word. Tones are written by default because
-  dropping them collides, and `hash` puts a short hash of the _hanzi_ on the end
-  for the homophones tones cannot separate: 权利 and 权力 are both `quan2li4`.
-  The options generalise it past URLs to search keys and identifiers, which the
-  [slug page](docs/slug/) tabulates. `pinyinjs slug` at the command line.
-
-## 1.2.2
-
-### Added
-
-- **`TEACHING_JIAO` and `COUNTED_MEASURE`,** two more lattice rules, and
-  `QUANTITY_CHARACTERS`, the 汉字 a quantity is written with that the second of
-  them asks about. See the fixes below for what they do and what they measure.
-
-- **`counts` on the numeral options.** Whether the number stands immediately in
-  front of something it counts, which is what makes a lone 2 两:
-  `numeralHanzi(2, { counts: true })` is `两` and `numeralHanzi(2)` is `二`.
-  Unlike `liang` it is not a variable choice — 二个 is simply wrong — but
-  nothing about the number says whether it is counting, so the caller says.
-  `convert` sets it from the character after the digits.
-
-### Fixed
-
-- **教 was `jiào` even where it was teaching.** 他在北京大学教了三年书 came out
-  `jiàole sān nián shū`, and 我教英语, 谁教你法语 and 她教我如何游泳 with it.
-  The dictionary stores 教 as `jiào` with `jiāo` as an alternate, and `jiào` is
-  right for the compounds — 教育, 教师, 宗教, 主教 — which are words and reach
-  their reading through the word; what is left is the 教 standing as a word of
-  its own, and that one is the verb. The object is what says so: a pronoun, a
-  noun or a name after it, or an aspect particle, since only a verb takes 了,
-  过 or 得. Forcing the character's own reading was not enough on its own,
-  because a two-character reading carries its 教 into the position from
-  outside — 王老师教我们汉语 read `jiào` off 师教 and 来教我 off 来教 — so an
-  untagged pair ending in 教 goes too, while every tagged word keeps its
-  reading: 任教, 宗教, 主教, 佛教, 传教, 执教, 请教, 家教. Over 88,866 lines
-  181 教 decode as a word of their own and every one read `jiào`; 158 are now
-  `jiāo`, three of them wrongly — 统一教创始人, 方法教深思 and 做到了教政分离,
-  where a nominal compound takes an object's shape. CPP's polyphone score is
-  89.05%, up from 89.04%.
-
-- **A 量词 was swallowed by the word behind it.** 三个人 was `sān gèrén`, three
-  _personals_, because 个人 is a common noun and nothing weighed it against the
-  个 belonging to the 三 in front of it. It is now `sān gè rén`, and so are
-  两家人, 三杯水, 四匹马, 十八层楼 and 一段长时间. What makes it decidable is
-  that the dictionary tags the 量词: 个, 次, 天 and 杯 are `q`, while the
-  characters that only look like measure words in this position are not — 分 is
-  a verb, 部 and 成 are nouns, 年 and 点 are numerals — so 五分钟, 三部分,
-  五成分, 五年级 and 三点钟 are untouched, and those are exactly the words a
-  blanket rule would break. Nor does an ordinal count anything (第三集团军), nor
-  a numeral inside a longer word (唯一道路), and a number and its measure
-  written solid survive whole where the dictionary has them: 一辈子, 一会儿,
-  一口气 and 两口子 are unchanged. Over 88,866 lines the rule forbids 561 edges
-  and moves 53 decodes, three of them wrongly — 一批评, 这一名词 and 六七股灾,
-  where the 一 and the 六七 count nothing. Exactly one reading changes in the
-  whole corpus and it is a fix: 已经下了两天雨了 read 天雨 as `tiān yù`.
-
-- **The Han after a number was decoded without it.** 2个人 was `liǎng gèrén`
-  while 两个人 written out was already `liǎng gè rén`, because a Han run was
-  decoded on its own and had no idea a number preceded it. A run is now decoded
-  with the 汉字 the digits stand for in front of it, and reported without them,
-  so digits and 汉字 reach the same decode. `decodeRun` and `decodeRunScored`
-  take that context as a fourth argument; where a reading would straddle the
-  join — 1点儿事, whose 一点儿 is one `yìdiǎnr` over both sides of it — the run
-  is decoded alone, as before.
-
-- **A lone 2 in front of a 量词 was read 二.** 我们买了2个西瓜 came out as
-  `wǒmen mǎile èr gè xīguā`, and a 2 standing immediately in front of what it
-  counts is 两: it is now `liǎng gè`, and 2岁 is `liǎng suì`, 2点 `liǎng diǎn`
-  and 2万人 `liǎng wàn rén`, the same 两 that 2,000 and 2:00 already got. The
-  measure words a 2 can count with are an open list, so 两 is the default and
-  the exceptions are named — 月, 日, 号, 楼, 路, 班 and 期, where the digit
-  labels a position rather than counting (2月 is `èr yuè`), an ordinal, which
-  the 第 in front of it marks (第2次 is `dì èr cì`), and 十 and 百, which follow
-  二 as 20 and 200 already do. A 2 inside a larger number is untouched: 12个 is
-  still `shí'èr gè`.
-
-## 1.2.1
-
-### Fixed
-
-- **A decimal lost the grouping of its counted part.** 一共75.5元 was
-  `yígòng qī shí wǔ diǎn wǔ yuán`, with the 75 in loose syllables, and is now
-  `yígòng qīshíwǔ diǎn wǔ yuán`, the same word 一共75元 already got. Only what
-  follows the 点 is read a digit at a time, and that part is unchanged: 3.14 is
-  still `sān diǎn yī sì`, and sandhi still stops at the point.
-
-- **A decimal point was read as a full stop.** 一共75.5元 came back as
-  `Yígòng …`, capitalised as though the text were a sentence, and a conversion
-  keeping its digits capitalised again after the point: 我有75.5个。was
-  `Wǒ yǒu75.5Gè.` A stop between two digits now ends no sentence.
-
-- **A compound lost the zh-TW reading its own parts carry.** 垃圾分類 read
-  `lājīfēnlèi` under `zh-TW`, because CC-CEDICT marks `Taiwan pr.` on 垃圾 and
-  垃圾桶 and on no other compound, and the decoder prefers the longer word, so
-  the entry that had the delta was never consulted. 101 compounds now compose
-  one from their constituents, 垃圾 accounting for 14 of them and 比肩, 玳瑁,
-  从容, 骨头, 蜗牛 and 说服 for most of the rest. Three guards keep it honest:
-  the constituent has to survive segmentation (運行狀況 contains 行狀 but reads
-  運行 + 狀況), it has to be read in the compound exactly as its own entry reads
-  it, and it has to be a whole word rather than a bare character, since 會's `huǐ`
-  would rewrite 三合會 as `sānhéhuǐ` and 3,743 entries would compose that way
-  against the 101 that whole words compose.
-
-- **A character's zh-TW delta was often a choice between its 普通话 senses.**
-  从容地 read `cōngróng dì` and 他都知道 read `tā dū zhīdào` under `zh-TW`,
-  because 地 was stored with a Taiwan reading of `dì` and 都 with `dū`. Neither
-  is a locale shift: CC-CEDICT lists 地[de5] and 地[di4] as separate entries, so
-  `dì` is what 地 reads in 普通话 when it means the ground, and the adverbial
-  particle is `de` in Taipei too. A delta the word already has as a 普通话 sense
-  is now refused, so 71 characters and 3 words lose one, 都, 着, 应, 差, 称, 斗,
-  舍, 薄 and 万 among them, while 和 `hàn`, 期 `qí`, 垃 `lè` and 髮 `fǎ` stay,
-  no 普通话 sense reading any of those that way. A `Taiwan pr.` note is also
-  only read off a sense matching the reading the entry settled on, which drops
-  16 more: the note on 著's chess-move sense, which reads `zhāo`, had been
-  giving the aspect particle 着 a 國語 reading of `zhuó`.
-
-- **A locale delta overwrote a reading it was not measured against.** The delta
-  describes the entry's own 普通话 reading, so a polyphone the decode read some
-  other way, such as 得 forced to the modal `děi` rather than the particle `de`
-  its entry stores, now keeps what the decode worked out instead of being
-  rewritten from the entry.
-
-## 1.2.0
-
-### Fixed
-
-- **`--system ipa` capitalised the IPA.** 我去银行。他姓王。 came out as
-  `Uo˨˩˦ ... Tʰa˥ ɕiŋ˥˩ Uaŋ˧˥`, with the sentence capital and the proper noun
-  carried over from the pinyin as though a transcription spelled words. IPA
-  letters are symbols rather than an alphabet, since `[T]` is not `[t]` in a
-  larger size but a symbol the IPA does not have, so the capitals are now dropped for
-  IPA and bopomofo and kept for the three romanisations. `toTranscription`
-  takes `{ capitals: false }` for callers writing their own.
-
-## 1.1.0
-
-### Added
-
-- **`SYLLABLE_TONES` and `isAttestedTone`.** Which tones each of the 424
-  syllables is actually written in, extracted from the merged dictionary and
-  held to it by a build assertion. 424 syllables in five tones would be 2,120
-  combinations and only 1,708 of them occur, so a fifth of that grid is empty.
-
-- **`Syllable.originalTone`,** the tone a neutral syllable had before it was
-  reduced. Gwoyeu Romatzyh writes the neutral tone as a dot in front of the
-  syllable _in its original tonal spelling_, so 没有 méiyou is `mei.yeou`, and
-  nothing else in a pinyin syllable records what that tone was. Optional, never
-  inferred, and used by GR alone.
-
-- **Two more outside sources for the transcription tables.** 148 words from
-  en.wiktionary's Mandarin pronunciation blocks, in bopomofo, Wade-Giles, Yale
-  and GR with their tones, and all 50 rows of Wikipedia's _Help:IPA/Mandarin_
-  key. The 417-row syllabary already checked was toneless, so it could say
-  nothing about tone marks, the neutral tone, 儿化 or how a word is joined,
-  which is where the four fixes below were found.
-
-### Fixed
-
-- **Bopomofo wrote the tone mark after the 儿化 ㄦ.** 哪儿 nǎr was `ㄋㄚㄦˇ` and
-  is now `ㄋㄚˇㄦ`: the mark belongs to the nucleus and the suffix is not part of
-  what it marks. A mark written after the ㄦ is still read.
-
-- **Wade-Giles put the tone digit on the 儿化 suffix,** so 花儿 huār was
-  `hua-êrh¹`, a first-tone 兒. It is now `hua¹-'rh`, with the digit on the
-  syllable the tone belongs to and the suffix written as the reduced `'rh` that
-  keeps it apart from 女儿 nǚ'ér's `nü³-êrh²`.
-
-- **Gwoyeu Romatzyh wrote 儿化 as a plain `-l` suffix.** GR fuses it into the
-  rime instead, and the rules are now implemented as _Spelling in Gwoyeu
-  Romatzyh_ gives them: 玩儿 wánr is `wal` rather than `wanl`, 事儿 shìr is
-  `shell`, 今儿 jīnr is `jiel`. The fusion is many-to-one, so `jiel` is 今儿 and
-  鸡儿 both and 128 forms no longer round-trip, every one of them a form the
-  language does not write.
-
-- **Gwoyeu Romatzyh wrote `.mhe` for a neutral 么.** The `-h-` is the first tone
-  of a sonorant initial, and a neutral syllable is in no tone at all; the basic
-  form goes behind the dot, so 什么 shénme is `shern.me`.
-
-- **The IPA module said it followed _Help:IPA/Mandarin_ and did not.** It
-  follows the broader IPA column of _Comparison of Standard Chinese
-  transcription systems_, which is where its ground truth comes from; the two
-  pages differ at the medials, -ang, the empty rhyme and the diphthongs. No
-  output changed, only the claim.
-
-### Changed
-
-- **Reading a romanisation narrows on the tone that was written.**
-  `readWadeGiles("lo²")` was `[luó, ló]` and is now `[luó]`: 咯 is a
-  sentence-final particle and is only ever neutral, so ló is not a syllable
-  Mandarin has. The same goes for `readWadeGilesLoosely`, where `pan²` is pán
-  alone, and for `readYale`, `readIpa` and `readGwoyeu`, where it settles the 儿化
-  collisions the same way (`ell` is 二 èr, there being no first-tone 兒).
-  Narrowing never empties a list: a tone no candidate is written in leaves the
-  candidates alone, since that says the tone is wrong rather than the spelling.
-  Over the phrase corpus, taking the first candidate recovers 82.66% of
-  syllables when the tone digit is written, against 79.05% without it.
-
-- Eight Wade-Giles forms no longer round-trip, and one to three in each of the
-  other systems, because they are forms the language does not write: `lo` in the
-  four contour tones, a first-tone 兒, and a neutral 誒. Every attested
-  syllable-and-tone combination still comes back.
-
-## 1.0.0
-
-The first stable release. Everything the
-[API page](docs/api/) lists is now covered by semantic versioning, including the
-decoder internals and the build pipeline.
-
-### Added since 0.1.0-beta.0
-
-- **Transcription systems, all five, in both directions.** Bopomofo, Wade-Giles,
-  Yale, Gwoyeu Romatzyh and IPA at the syllable level, plus hanzi → any of them
-  end to end via `convert --system` and `convertToWadeGiles`. Checked against
-  417 rows of an outside syllabary in every system, which is 3,336 cells.
-- **Reading ambiguous Wade-Giles.** `readWadeGilesLoosely` allows for the
-  apostrophes and diacritics real texts drop, and `splitWadeGiles` puts back the
-  hyphens they lose. Measured: 52.07% of written syllables lose their identity
-  without their marks, and taking the first candidate recovers 79.05%.
-- **Numbers read aloud** (`src/numerals/`, the `number` command): counting and
-  spelling out, 两 against 二, decimals, percentages, fractions, `yāo`, and
-  times. `convert` reads the digits it meets and takes the style from what
-  follows them.
-- **Orthography rules.** The 重叠 hyphen for AABB and repeated words, the 成语
-  hyphen from a curated list of 117, 老王 → `Lǎo Wáng`, and GB/T 16159 5.1 in
-  full, so 毛泽东 is `Máo Zédōng` and 北京大学 is `Běijīng Dàxué`, on boundaries
-  CC-CEDICT's own capitalisation states.
-- **Typed rule overrides on the lattice** (`READING_RULES`), with the modal 得
-  and attested 儿化 as the first two.
-- **Tone colour in the terminal**, in MDBG's palette, on every command that
-  writes a syllable.
-
-### Fixed since 0.1.0-beta.0
-
-- **一 sandhi is about the 一 that counts.** 十一月 is `shíyīyuè` and 第一次 is
-  `dìyīcì`; the citation tone survives a final digit and an ordinal. 561
-  conversions change over 88,866 lines, 520 of them right.
-- **A colon between digits is a time**, not an identifier: 6:30 is
-  `liù diǎn sānshí fēn`.
-- **The documented proper-noun mechanism was wrong.** The capitals page
-  described a 500-entry surname list that has never existed; what produces
-  `Lǐ Huá` is two single-character dictionary entries each flagged a proper
-  noun.
-
-### Changed
-
-- **`engines` is now `>=22.0.0`**, down from `>=24.0.0`. The suite and the
-  packaged tarball are both exercised on Node 22 and 24 in CI.
-- **No more `beta` dist-tag.** `pnpm add @kensio/pinyinjs` finds it.
-- `src/romanization/` is `src/transcription/`, since bopomofo has a script of
-  its own and IPA is a transcription rather than a spelling. The published docs
-  path keeps the older name.
-- The build cleans `dist/` before compiling, and `pnpm pack:check` asserts the
-  tarball's contents. The beta tarball shipped 24 stale files from a directory
-  that had been renamed away.
-
-### Accuracy at 1.0.0
-
-Over the 130-case gold corpus, which is committed to the repository rather than
-to the published package, and scored by `pnpm accuracy`:
-
-|                  |   lattice | greedy baseline |
-| ---------------- | --------: | --------------: |
-| exact match      | **97.7%** |           86.9% |
-| reading accuracy |     99.7% |           97.9% |
-| spacing (F1)     | **99.8%** |           95.7% |
-
-Every figure in that table is asserted against the scorer by
-`src/changelog.test.ts`, because a number nothing executes goes stale, which is
-what happened to this project's README table across three releases.
-
-`pnpm polyphones` reports 89.04% for the lattice against greedy's 88.82% over
-CPP's 20,139 hand-labelled polyphones. That corpus is fetched rather than
-committed, so unlike the table above it cannot be asserted here, and it is
-reported rather than guarded.
-
-## 0.1.0-beta.0
-
-First published release: the lattice decoder, the tiered dictionaries, the
-syllable layer, sandhi, GB/T 16159 spacing and capitals, HTML output and the
-`pinyinjs` command.