Languages

The bundled base-language registries are the 49 base keys in the pinned num2words parity contract:

Codes
`am` `ar` `az` `be` `bn` `ca` `ce` `cs` `cy` `da` `de` `en` `eo`
`es` `fa` `fi` `fr` `he` `hi` `hu` `hy` `id` `is` `it` `ja` `kn`
`ko` `kz` `lt` `lv` `mn` `nl` `no` `pl` `pt` `ro` `ru` `sk` `sl`
`sr` `sv` `te` `tet` `tg` `th` `tr` `uk` `vi` `zh`

Explicit locale overlays inherit a base registry and are independently keyed:

Base	Locale overlays	num2words parity
`en`	`en_IN`, `en_NG`	v0.5.14
`es`	`es_CO`, `es_CR`, `es_GT`, `es_NI`, `es_VE`	v0.5.14
`fr`	`fr_BE`, `fr_CH`, `fr_DZ`	v0.5.14
`pt`	`pt_BR`	v0.5.14
`zh`	`zh_CN`, `zh_HK`, `zh_TW`	current master only

Language input is normalized by trimming whitespace, accepting both hyphens and underscores, canonicalizing base/region casing, and trying the exact registered locale before falling back to its base. For example, pt-BR selects pt_BR, fr_FR selects fr, and en_GB selects en. Common verified three-letter aliases and cz for Czech remain accepted. eo and es_NI are explicit keys; eu is not supported.

The added registries are intentionally conservative. Ambiguous short forms are guarded by numeric/name context or omitted, Russian multiword abbreviations accept ordinary, non-breaking, and narrow non-breaking spaces, and Turkish lexical entries are case-sensitive. Unit output is a canonical singular lemma; the stable API does not realize plural, case, numeral-government, or suffix morphology.

English, French, Italian, Portuguese, and Spanish structured quantities recognize the reviewed shared currency identities €/EUR, $/USD, and £/GBP in either numeric-prefix or numeric-suffix position. Portuguese additionally recognizes R$/BRL as currency-brazilian-real, while Czech recognizes Kč/CZK as currency-czech-koruna. These matches preserve the written numeric lexeme, source offsets, written symbol, language, category, and stable canonical ID; they do not speak the amount or choose locale-specific currency grammar. Downstream consumers such as spokenform own that semantic realization.

French dotted duration forms min. and sec. are context-sensitive structured quantity aliases. A numeric match consumes the complete dotted symbol, while standalone lexical min. continues to expand as minimum; numeric context is what selects the duration identity. Sentence punctuation is rendered by the generic abbreviation layer, and French number grammar remains downstream.

German quantity symbols include the reviewed electrical and frequency forms kWh, Wh, mAh, mA, GHz, MHz, kHz, Hz, W, and V, plus Stck., ltr., Tsd., Mio., Mrd., and EUR. Case-sensitive metadata is preserved: for example, 2 mA matches but 2 ma does not. Dotted German aliases and their base symbols share canonical IDs, so h/Std., min/Min., and l/Ltr. cannot acquire separate semantic identities.

Known ambiguity

Some inventories contain collisions whose effective winner follows the source registry order rather than a complete contextual model. Current documented examples include str. in Czech, mar. in Spanish and Italian, n° in French, and seg. in Portuguese. German Fr. has an explicit title-context override, but its default expansion remains Freitag.

Context policy is language-specific and uses bounded local windows. English profiles distinguish street suffixes, saints, titles, and addresses; the German profile distinguishes Fr. Müller (Frau) from Am Fr. (Freitag). Name evidence uses Unicode cased characters and supports accents, apostrophes, hyphens, and quoted names. Uncased scripts remain conservative without an annotation signal. Context-aware output is still not a comprehensive semantic disambiguator, so applications should review domain-specific entries.

Uppercase undotted initialisms such as English time zones, BCE, CE, and MIT are case-sensitive. This prevents ordinary lowercase words from being rewritten; reviewed title-case and lexical abbreviations retain their own registry policy.

Coverage tiers and generated inventory

Coverage metadata is checked in with the language bundles and registry shards. Run python scripts/generate_registry_snapshot.py after an intentional data change and inspect the shard diff.

Tier	Base keys	Content contract
Reviewed extended	`cs de en es fr it nl pl pt ru sv tr`	Mature lexical inventories migrated through declarative bundles with parity snapshots
Reviewed baseline	`am ar az be bn ca ce cy da eo fa fi he hi hu hy id is ja kn ko kz lt lv mn no ro sk sl sr te tet tg th uk vi zh`	Source-tagged references/titles, bounded numeric guards, localized neutral unit labels, and script-safe boundaries
Locale overlay	14 explicit locale keys	Base inheritance plus structured numeric identities; no identity lexical currency rules

The detailed generated count table is maintained in docs/language-coverage.md.

Baseline unit labels are neutral surface labels, not plural/case/gender or numeral-government realization. Consumers that need semantic quantities should use iter_unit_matches() and perform locale grammar downstream.

DATE requires nearby numeric date evidence and is deliberately not an unrestricted parser. Uncased scripts remain conservative without the cased letter title heuristic; CJK entries do not rely on Latin \b behavior.