v0.4.0

Added

• Calendar ranges in parseDateRange (@pascal-app/lingo/date), which
  previously covered only clock slots and anchored durations:
  • Date to date — July 1 to July 5, Aug 3 - Aug 9,
    between Aug 3 and Aug 9, from tomorrow to friday, Mon-Fri,
    2026-08-01 to 2026-08-05, and open ends (from monday, until august 9).
    The end resolves against the start, so a pair never reads backwards.
  • Calendar periods — next week spans Mon–Sun, next month the 1st to the
    last, this year and 2027 the whole year, August the whole month. A
    coarse endpoint widens when it closes a span too, so July to August ends
    on August 31 and until August does the same, while from August still
    opens on the 1st.
  • Weekends — this weekend, next weekend, last weekend span Saturday
    through Sunday. parseDate gained the next/last weekend modifiers too.
    On a Saturday or Sunday, this weekend is the weekend in progress.
  • humanizeDateRange renders these as dates rather than clock times, so they
    round-trip. Time slots are unchanged.
  • A descending dated pair such as 2026-08-09 to 2026-08-03 is swapped and
    reported with the existing RANGE_REVERSED warning instead of being handed
    back backwards. Overnight clock slots (9pm to 5am) are untouched.
• Chinese and Japanese calendar and clock grammar: numeric dates written with
  suffixes (2026年3月5日, 3月5日, and years spelled digit-by-digit as
  二〇二六年), weekdays (星期三, 周三, 水曜日), clocks closed by
  点/時/分/秒 with day periods (下午3点, 午前9時半, 3点一刻), and
  date+time compounds written with no separating space (明天下午3点,
  明日午後3時).
• Postpositional bound phrases for scripts that put the comparator after the
  quantity: 5キロ未満, 5キロ以上, 5公斤以下, 5公斤以内, 5キロ超.
  These previously parsed as a bare 5 kg, silently dropping the bound, so
  they now return a range where they used to return a quantity — check any
  code that reads .quantity off a zh/ja parse of one of these forms.
• Chinese and Japanese default currencies, so the shared ¥/￥ symbol
  resolves per locale (￥100 is CNY under zh, JPY under ja), plus the
  元/圆/块/人民币 and 円/えん aliases.
• Spanish billón (10^12), Portuguese cento, and Romance ordinal
  day-of-month forms (le 1er mars, el 1º de marzo).
• A per-locale benchmark suite (bun run bench) so multi-language throughput is
  measured rather than assumed.
• Per-locale corpus contracts gained 96 rows and the English contract 5,
  covering the grammar above plus the number-word fixes below.

Changed

• Loading a locale pack now resolves its currency symbol instead of warning
  about it: under zh or ja, ￥100 returns CNY or JPY with no
  AMBIGUOUS_UNIT warning, where it previously returned JPY plus the warning
  in both. Selecting a locale is the disambiguation. Parsing without a pack
  loaded is unchanged — bare ¥, $, and cents still warn.
• dateRangeField() advertises its real grammar to models. The JSON Schema
  description promised only time slots, so a model had no reason to emit
  "next week" or "Aug 3 - Aug 9" into a field that accepts them.
• Resolved language profiles and locale-detection scans are memoized per pack
  set, so repeated parses with locale/auto-detection no longer re-merge packs
  or re-tokenize for detection on every call.
• Size budgets recalibrated twice for this wave — once for the locale idiom
  engine and once for calendar ranges (D70 and D71 in wiki/decisions.md carry
  the measured numbers and rationale); scripts/size.mjs remains the source of
  truth. Every corpus row pinned at 0.3.0 is byte-for-byte unchanged — this
  release only adds rows. Five readings outside the pinned set do change, all
  of them silent wrong answers being corrected, and each is called out in its
  own entry: postpositional bounds, ￥ under a loaded pack, French
  mille cinq cents, French billion, and Spanish dos mil millones. If you
  depend on any of those, read them before upgrading.
• README (npm): repo-relative links (docs/recipes.md, plans/, wiki/) now
  point at absolute GitHub URLs so they resolve on npmjs.com, and the docs
  site plus the agent llms.txt index are linked from the top.
• Root README: links the docs site, the npm package, and the agent llms.txt
  index directly, with an install one-liner.
• Docs site: every markdown section now also renders as a standalone,
  indexable HTML page at /docs/<section> (agent markdown stays at
  /docs/<section>.md); legacy demo redirects (/forms, /escalation,
  /coverage, /integrations) became permanent (308); the landing page
  gained server-rendered parse-example, forms-vs-LLM, tool-boundary, and FAQ
  content (FAQPage/TechArticle/BreadcrumbList structured data included); docs
  section headings no longer leak kicker text into the extracted HTML outline.
• Docs site: three demos show what a canonical reading is worth downstream —
  one date field that picks its own picker (day, two-month range, or time slot)
  from the shape parseDateRange returns, a LaTeX view that typesets a reading
  as real notation because the unit id and the value are separate fields, and a
  data grid whose columns are quantityFields, so mixed cell notation
  normalizes into one unit and the totals row can just add numbers. All three
  are keyboard-operable and the LaTeX view renders MathML for screen readers.

Fixed

• Agent-facing docs now cover calendar ranges. llms.txt, the package README,
  the skills/lingo skill, the site's markdown mirror, and the parseDateRange
  / dateRangeField TSDoc all described time slots only, so an agent reading
  any of them would not have found the feature. Each now states the three
  shapes, the closing-side widening rule, the RANGE_REVERSED swap, and what
  is deliberately absent (quarters, Aug 3-9, dash-joined ISO dates).
• /docs/dates.md (and therefore /llms-full.txt and /llms.md) documented
  2026-08-03..2026-08-09 as a supported span. .. is not a separator in the
  grammar and that input returns UNSUPPORTED_DATE; the example now uses a
  form that parses.
• /llms.txt never advertised @pascal-app/lingo/react-native, and the site's
  markdown mirror never mentioned @pascal-app/lingo/complete, hiding two
  published entry points from agents that read only those files.
• The docs gates now catch that class of drift: check-docs-sync.mjs asserts
  both llms.txt copies against package.json exports and every issue code
  and verifies the served copy is not stale, and check-llms-index.mjs asserts
  the generated index advertises every published entry point.
• The /docs sidebar no longer contradicts the page it scrolls: it listed
  Find values in text before Autocomplete anything, and Two-way slider
  after the two framework entries, in both cases the reverse of the document
  order the same nav highlights while scrolling.
• /docs heading levels are visually distinct again. SectionHeading (h2) and
  SubHeading (h3) both rendered at 14px, so the outline read flat and
  disagreed with the 16px/15px scale the standalone /docs/<section> pages use
  for the same content. Autocomplete anything also appeared twice in a row
  (section heading, then demo panel title), and Kinds hand-rolled its own h3,
  leaving #coverage-kinds as the only depth-3 anchor with no anchor link.
• Docs sidebar group labels and their subtitles met WCAG AA: at
  text-muted-foreground/75 and /55 on 11px text they measured 2.9:1 and
  2.1:1 in light mode (3.0:1 dark). Both now use the unmodified
  --muted-foreground token — 4.5:1 light, 7.1:1 dark.
• Number words: hundreds now multiply only the 1..99 group in front of them, so
  French mille cinq cents is 1500 (was 100500). A banked smaller scale
  multiplies the next one, so Spanish dos mil millones is 2×10^9 (was
  1,002,000). Both were silent wrong answers.
• French billion is 10^12, the long scale French actually uses; it read as
  10^9 before, a silent factor-of-1000 error. milliard was already correct.
• Scale words joined by de/e parse instead of failing on the connector:
  Spanish mil millones de metros, French un milliard de mètres, Portuguese
  mil e quinhentos.
• between A and B now keeps the and-word for the range when both sides are
  spelled scale words: between one thousand and two thousand meters,
  entre mille et deux mille, entre mil y dos mil (all previously failed to
  parse). between five and a half and ten kg still reads 5.5..10 — the
  fraction tail binds tighter than the range separator.
• Portuguese cento e vinte (120) and mil e quinhentos (1500) parse; the
  and-word after a bare scale word links it to its remainder.
• Portuguese locative contractions are date fillers, so
  na proxima segunda-feira resolves.
• Scripts written without word spaces no longer swallow grammar words glued to a
  quantity (三至五天, 5公斤左右). Unit aliases stay atomic through the split,
  so 一時間半 remains 1.5 hours rather than splitting at the range word 間.