Skip to content

Research-driven quality + locale idiom coverage + 0.2.1 prep#1

Merged
Aymericr merged 3 commits into
mainfrom
improve/research-driven-quality
Jul 9, 2026
Merged

Research-driven quality + locale idiom coverage + 0.2.1 prep#1
Aymericr merged 3 commits into
mainfrom
improve/research-driven-quality

Conversation

@Aymericr

@Aymericr Aymericr commented Jul 9, 2026

Copy link
Copy Markdown
Contributor

Description

Three passes, adversarially cross-reviewed (Opus 4.8 ↔ Codex 5.5, Fable orchestrating):

  1. Research-driven quality (D67) — seeded property round-trip suite (41k+ cases, plan 010 layer 2) which found and fixed real two-way breaches (scientific coefficients, narrow-style unit gluing, time-grain anchored-range humanize); ARIA combobox contract for DOM completions; isNumber() guard; compile-time-checked .to(); MCP envelope fix; /ai output schemaVersion fix; +3–8% parse throughput.
  2. Locale idiom coverage wave 1 (plan 033, D68/D69) — CJK number engine (三十五公斤, 三百五十万, 5〜10キロ, 两公斤半), Romance composition (treinta y cinco, quatre-vingt-dix, dos coma cinco), localized date grammar (las tres menos cuarto, quinze para as tres, lundi en huit, Monday week, 月底, 再来週), deepened es/fr/pt/zh/ja packs, and per-locale corpus gates enforced by bun run check.
  3. Release prep — entropy/dedupe pass (−17 net LOC, behavior pinned), README/llms.txt/CONTRIBUTING synced with execution-verified examples, YAML issue forms + Discussions enabled.

Checklist

Gates

  • bun run check green (912 tests, all 24 size budgets, main + 5 locale corpus gates, zero-deps, docs-sync, schema)
  • bun run lint green
  • Corpus diff ADDITIVE only — interpretation of previously-valid input unchanged

API surface

  • Semver impact: minor-level additions only; shipping as 0.2.1 patch-line release per owner call
  • TSDoc @example on new exports; README / llms.txt / site synced together

Knowledge layer

  • D67–D69 decision entries; plans 010/033 updated; wiki/research + inspiration + backlog updated in-change

🤖 Generated with Claude Code

Aymericr and others added 3 commits July 9, 2026 02:08
…ombobox, perf, DX (D67)

Multi-agent research + audit pass (lodash/es-toolkit, Drizzle, AI SDK, Base UI,
FFmpeg, competitive landscape) distilled into wiki/research; the actionable
findings implemented and adversarially cross-reviewed.

Added
- Seeded property-based round-trip suite (plan 010 layer 2): 41k+ cases over
  every kind × unit × format style; found and fixed real two-way breaches.
- DOM completions ship the headless ARIA combobox half (role, aria-autocomplete,
  aria-expanded incl. blur collapse, listboxId → aria-controls).
- isNumber() result guard; compile-time unit checking on Quantity/Range .to().
- Perf: Intl.NumberFormat cache, binary-search token lookup, allocation-free
  unit-match dedupe, regex-free locale detection (+3–8% parse throughput).

Fixed (D67)
- Scientific coefficients with 3 decimal digits (3.493e-4 m) re-parse; exponent
  clears the AMBIGUOUS_NUMBER issue and its stale alternative.
- Narrow style keeps the space for K/M/ft³/ft²/kΩ so output re-parses.
- Time-grain anchored ranges humanize re-parseably; anchored path threads
  trailing zones through finishRange semantics; absolute anchors need no now;
  anchored-range offsets survive normalization shifts.
- /ai output JSON Schema schemaVersion 2 → 3 (matches the v3 wire shape).
- MCP lingoTool callback accepts bare args and the JSON-RPC envelope.
- LOCALE_NOT_LOADED / NOW_REQUIRED copy now names the fix.

./date standalone budget recalibrates 38.2 → 38.3 kB for the anchored-range
correctness weight (D67, D14 pattern). Corpus additive-only; 789 tests green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Codex GPT-5.5 <noreply@openai.com>
…, localized dates (D68/D69)

Research pass across es/fr/pt/zh/ja/en (335 test-precise gap idioms,
wiki/research/locale-idioms.md, plan 033) implemented as language-neutral
engine seams consumed through new optional pack-data fields — no per-language
parser branches (the chrono-node anti-pattern).

Engine (D68)
- CJK number walker: 三公斤/三十五キロ sub-token segmentation, 万/亿 grouping
  (三百五十万), elliptical 一百五=150 / 三万五, mixed 3万5千, wave-dash ranges
  5〜10, adjacent-number ranges 七八天, post-unit 半; 万万/亿万 rejected;
  D47 glued-1M guard holds.
- Romance composition: treinta y cinco, bare scales (cien gramos), composed
  table (quinientos, exhaustive vigesimal quatre-vingt-dix), spoken decimals
  as digit sequences (dos coma cinco seis = 2.56; English two point five),
  article-led chains (un millón quinientos mil), multi-word approximants
  (más o menos) with longest-first matching.

Dates (D69)
- Pack-owned spoken clock (las tres menos cuarto, quinze para as tres,
  quarter of five), period edges (fin juillet, 月底), weekday offsets
  (lundi en huit, Monday week), afterNext/beforeLast (再来週, 先々週),
  day-part compounds (tomorrow morning), localized durations (2 horas).
  English clock grammar migrated to the same fields, behavior-identical.

Packs + gates
- es/fr/pt/zh/ja deepened (~200 locale corpus rows); per-locale corpus
  contracts discovered generically and enforced in bun run check; locale
  auto-detection scores the new number tables. Traditional units 斤/两
  deliberately not registered (D4 hazards, backlog).
- Budgets recalibrated once for the wave (D68/D69 entries): full 38.3,
  core 25.6, date 41.0/14.4, ai 17.4, pack budgets raised for restored
  breadth. Main corpus: zero interpretation changes (additive only).

Cross-reviewed adversarially (Codex ↔ Opus); both review blockers fixed
(digit-sequence decimals, article-led chains) plus 万万 rejection,
detection scoring, veintiuna, longest-first phrase contract.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Codex GPT-5.5 <noreply@openai.com>
- Dedupe parallel-agent redundancy: shared eatAnyPhrase (parse/config),
  canonical stripDateFillers (date/state), dead number-word re-exports
  removed (-17 net LOC, behavior pinned by 912 tests + 6 corpus contracts).
- README/llms.txt/CONTRIBUTING synced to post-wave reality with
  execution-verified examples (treinta y cinco kilos, 三十五公斤,
  las tres menos cuarto, Monday week, 月底); site data refreshed.
- Contribution surface at editor parity: YAML issue forms (bug/feature),
  blank issues off, Discussions enabled with Q&A/Ideas links.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Codex GPT-5.5 <noreply@openai.com>
@vercel

vercel Bot commented Jul 9, 2026

Copy link
Copy Markdown

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

Project Deployment Actions Updated (UTC)
lingo Ready Ready Preview, Comment Jul 9, 2026 11:13pm

@Aymericr Aymericr merged commit 735fd85 into main Jul 9, 2026
11 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant