[derek73]

Adds locales.ES_AR, an Argentine Spanish pack: locales.ES with a longer given-name list. Also adds config_diff.py score, the per-country sweep the pack was chosen by.

Why Argentina, and why a list rather than turning two family names off

decisions/O6's 2026-10-27 ORCID table showed two_family_names (rules/O6) taking Argentina from 80.3% to 59.7%. That was the switch without the Spanish given-name list. With the list, as locales.ES ships it, Argentina reads:

Source	Records	No pack	locales.ES	locales.ES_AR
ORCID researchers	19,364	83.44%	90.19%	96.13%
PatentsView inventors	3,343	93.81%	86.18%	92.73%
PSC filers	320	81.88%	81.88%	91.25%

(records agreeing in every box with the filer's own split)

• What ES gets wrong. PatentsView's Argentine inventors mostly write two given names and one surname. O6 then takes the second given name into the family whenever Spain's list lacks it.
• Why the list lacks those names. INE's 90% bar leaves out names Spain gives but also bears as surnames in a seventh to a quarter of their uses: Nicolás, Matías, Sebastián, Andrés, Ramón.
• Why not O6 off. Turning the switch off for Argentina reads exactly as the default does, and loses 6.7 points on ORCID.
• What ES_AR does instead. It keeps the switch on and cuts the same INE tables looser: share ≥ 0.5, 5,000 words, decided in chat.
• The cost. PatentsView stays 1.1 points below the default.
• Why not change ES itself. The looser list costs Spain, Mexico and Peru 0.6 to 1.3 points, so ES keeps its own list.

Every other country

config_diff.py score was run over every country or nationality with at least 100 records, under default/pack × every on/off combination of O6 and two_given_names (rules/O7). The measure recorded in decisions/O6: a configuration must beat the country's language pack by at least 2 points in every source holding 300 or more of its records, and beat no pack summed over those sources. Only Argentina passes.

• Spanish: ES is beaten by no switch combination by more than 0.06 point on ORCID.
• ES_AR's list elsewhere in Latin America: it beats ES by up to 0.61 point in Chile, Colombia, Venezuela, Uruguay, Paraguay and Central America. That falls short of the 2-point bar, and is recorded as the Open line's lead.
• two_given_names: it helps nowhere PSC, the one source that can see it, measures it.
• Portuguese: O6 splits by source (ORCID BR +24.9, PatentsView BR −24.5), so PT keeps waiting for a list.
• Philippines: the default reads 97–99%.
• Morocco: O6 alone helps in two sources. That is a lead for 3.1.

Shape

• The pack: ES_AR = Locale(code="es_ar", lexicon=Lexicon(given_names=...), policy=rules_patch("es"), languages=frozenset({"es-AR"})).
• No es-AR.toml. A choice of es-AR already selects es.toml by ancestry.
• No [rules] inheritance. No country measured needs it. It would also need two_family_names = false, which build check 12 refuses, and it would make parser_for(ES, ES_AR) depend on argument order.
• Stacking is harmless. The new list holds every word of ES's, so parser_for(ES, ES_AR) reads as ES_AR.
• No separate preset API.
• The list is generated by tools/es_given_names.py --list es-AR from INE's tables (CC BY 4.0).

Changes

• New: nameparser/locales/es_ar.py, _es_ar_given_names.py, and the registry entry.
• tools/es_given_names.py: writes either list (a LISTS table, --list).
• config_diff.py score: scores configurations against the filed boxes, per country. --switches and an @field=0|1 suffix on any configuration read it under each on/off combination of named Policy switches. It refuses a group or language that selects nothing.
• Docs:
  • decisions/O6: a 2026-11-04 entry, a Standing line, a head note on phase 3g, and an updated Open line;
  • one sentence in mechanisms.md's LOCALE-PACKS-PURE-DATA;
  • docs/locales.rst: a table row, a section and doctests;
  • the README's available() output;
  • a release-log bullet;
  • AGENTS.md, vocab/README.md and build.py comments on which packs read es.toml's [rules].
• AGENTS.md also records a hazard this PR's review hit: a probe that plants a fixture corpus outside a test module's cache fixture overwrites the shared local corpus. It replaced psc-individuals with one record during review, and it has been restored byte for byte.

Checks

• Tests: full suite green; the touched modules also pass on Python 3.15.
• Lint, types, docs: ruff and mypy clean; Sphinx doctests and README doctest pass.
• Committed corpus: 120 names move between ES and ES_AR, each a listed word staying in the middle name. Comma shapes don't move. Under the family-first orders the list's veto acts symmetrically, as decisions/O6 records.
• Reviews: a design-docs review and a code review, then a review of the fix commit, which found three defects in the fixes; all are addressed in c22b3a7.

🤖 Generated with Claude Code