feat(locale): CJK calendars, postpositional bounds, and number-word arithmetic fixes

[Aymericr]

Wave 2 of plan 033. Wave 1 taught Chinese and Japanese to read numbers; this teaches them to read calendars, and fixes three silent wrong answers in the Romance number engine.

The review that produced this started from "how good is our multi-language support, really" — so the notable finding is that two of the gaps were silent: input parsed successfully with the wrong value, which for a validation library is worse than failing.

The silent wrong answers

input	locale	before	after
5キロ未満	ja	5 kg (bound dropped)	< 5 kg
mille cinq cents metres	fr	100500 m	1500 m
mil millones de kg	es	1,001,000 kg	1e9 kg
between one thousand and two thousand meters	en	failed to parse	1000..2000 m
cento e vinte metros	pt	failed to parse	120 m

5キロ未満 is the one that motivated the tokenization work: 未満 was glued to the unit token, so nothing ever saw the comparator.

Four mechanisms, all pack-data-gated

English pays nothing for any of this — each path is entered only when a pack declares the relevant field.

1. Glued-grammar splitting. Scripts without word spaces write grammar and content as one token, so tokenization now cuts word tokens at the profile's non-Latin grammar vocabulary. The subtle part: the cut must be suppressed inside a unit alias, because 間 is a range word but 時間 is the hour unit — otherwise 一時間半 splits into a range. That's why splitGluedWords takes an alias-length callback.
2. Postpositional bounds. GrammarBoundPhrase.suffix declares comparators that follow the quantity (以上, 未満, 以内, 超), consumed by parseTrailingBound in both the quantity and range paths.
3. Suffix-delimited dates and clocks (date/suffix.ts, date/numeral.ts): 2026年3月5日, 3月5日, 午後3時半, 三点一刻, 下午3点, plus date+time written without a space (明天下午3点) split at day-period and clock-suffix anchors. Years spelled digit-by-digit (二〇二六年) read as a digit run — positional reading would give 2+0+2+6.
4. Number-word arithmetic. Hundreds now multiply only the 1..99 group in front of them; a banked smaller scale multiplies the next one; the and-word after a bare scale links its remainder.

Fixing (4) exposed its mirror image: inside between A and B the and-word belongs to the range, so the existing noAnd flag now threads from range.ts through ValueCtx into the word engine. It deliberately does not suppress the fraction tail — between five and a half and ten kg still reads 5.5..10. (Blanking andWords wholesale looks equivalent and breaks that case; there's a regression test.)

Also in here

• zh/ja currency defaults — the shared ￥/¥ now resolves per locale (CNY under zh, JPY under ja), plus 元/块/人民币 and 円 aliases. Previously ￥100 read as JPY in Chinese text.
• Romance depth — long scales (milliard/billion, mil millones/billón), ordinal days (le 1er mars), Portuguese cento, and Portuguese locative date fillers so na proxima segunda-feira resolves.
• Performance — resolved profiles and detection scans are memoized per pack-array identity. Measured on the memoized functions in isolation (fresh array literal reproduces the old cost exactly): resolveLanguageProfile 15.1 → 0.04 µs per call, detectLocale 85 → 2.5 µs per call. Profile resolution previously ran on every parse, allocating ~40 Sets/Records each time.
• A per-locale benchmark suite, so multi-language throughput is measured rather than assumed: explicit locale with 5 packs loaded is 5.66 µs/op, auto-detect romance 13.98 µs/op, auto-detect CJK 8.10 µs/op.

Two things worth a second opinion

Size budgets were recalibrated (approved before landing, per the D19 escalation rule). The growth is capability, not drift, and ./ai moved purely as a cascade — no /ai code changed:

entry	before	after	budget
main	38.19	39.18	39.4
./core	25.46	26.56	26.8
./date standalone	40.85	43.01	43.3
./date marginal	14.29	15.46	15.7
./ai marginal	17.27	18.39	18.6
zh / ja packs	1.23 / 1.31	1.82 / 1.79	2.0 each

The noAnd threading touches the shared English path. It's guarded by the corpus contract (zero changes to any pinned row) and by explicit guard tests, but it's the one change here that isn't locale-isolated, so it's the place to look hardest.

Verification

• bun run check green: typecheck, 980 tests, build, size, corpus gate, zero-deps gate.
• Corpus contracts: zero breaking changes across all six (en/es/fr/pt/zh/ja) — the whole wave is additive: +96 locale rows, +5 English rows.
• Site typechecks and production-builds; playground examples and the agent-facing llms.txt updated (it previously advertised the CJK clock as deferred).
• Every code snippet added to the README, llms.txt, and site docs was executed before landing, not written from memory. One caught a real mistake: a toISOString() example that renders March 4 in UTC+2.

Docs

D70 in wiki/decisions.md records the mechanisms, the rejected alternatives, and the measured budget numbers; the two non-obvious invariants (alias-guarded splitting, noAnd sparing the fraction tail) are in wiki/architecture.md under key mechanisms. Plans 031/033 updated, and the wave-1 deferrals this retires are marked in plans/backlog.md alongside five new wave-3 items found along the way (Japanese era years, CJK suffix-date ranges, 3月5日(木) weekday agreement, positional CJK numerals in dates).

Made with Cursor

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
lingo	Ready	Preview, Comment	Jul 27, 2026 11:23am

[github-actions]

React Doctor found no new issues. 🎉

_{Reviewed by React Doctor for commit 71fd784.}