Localize version 0.44.0

[0.44.0] — July 4th, 2026

This release closes the TR35 conformance gaps in the date/time formatting layer and fixes locale-fidelity bugs surfaced by newly-wired conformance data.

Added

• Flexible day periods: the B format symbol selects CLDR day periods ("in the morning", "mittags", "noon", "midnight") from the day-period rules now in the supplemental data pipeline, and b renders noon/midnight at the exact points. Languages without rules fall back to AM/PM.

• Localize.LanguageTag.U.parse/1 and parse!/1 parse a standalone BCP 47 -u- extension string (with or without the u- singleton) into a validated Localize.LanguageTag.U struct, using the same canonicalization as Localize.validate_locale/1. Localize.LanguageTag.U.encode/1 is now a documented public function.

• Downloaded locale files are verified against a SHA-256 hash manifest bundled with the package (priv/localize/locale_hashes.etf, generated by mix localize.generate_locale_hashes at release time), before any decode or cache write. A mismatch or missing entry fails the download with Localize.LocaleIntegrityError; when no manifest is present a one-time warning is logged and verification is skipped.

• The rbnf conformance suite now runs the vendored ICU reference data for de, es and fr in addition to en, and the likely-subtags suite asserts the CLDR FAIL rows.

Fixed

• Localize.Validity.U.encode/2 emits the preferred BCP 47 spelling instead of a deprecated alias (:islamic_civil encodes as "islamic-civil", not "islamicc"), and decode/2 accepts struct-field atom keys and the canonical hyphenated calendar spellings ("islamic-civil").

• Week fields honour the locale's week configuration: w and Y compute week-of-year from the locale's first day and minimum days (2023-01-01 is week 1 in en, week 52 of 2022 in de), and W computes week-of-month accordingly. Previously all three were ISO-only.

• Multi-replacement territory aliases follow TR35 Annex C: hy-SU canonicalizes to hy-AM via the likely territory, uk-SU to uk-UA, sr-YU to sr-RS; previously the first entry (RU) was always taken.

• RBNF decimal-format substitutions (=#,##0=) format with the rule's locale instead of the process locale, so fr digits-ordinal renders "1 141e" and de "1.141." rather than the English "1,141".

• Likely-subtags no longer fabricates mappings for languages without one: private-use languages (qaa) pass through unmaximized per the CLDR test data instead of inheriting the und mapping.

• Locales whose language has no CLDR match resolve to root (und) data instead of the alphabetically nearest wrong language — tlh no longer formats with Afar data.

• Compact formats no longer raise on a negative :fractional_digits option.

Changed

• mix test --cover now excludes build-time tooling (the data-generation pipeline, mix tasks, compile-time parser generators) from coverage measurement so the number reflects the runtime library.