[derek73]

What this changes

locales.CS and locales.SK now read two family names (Policy.two_family_names, rules/O6). A Czech or Slovak woman who keeps her own surname beside her husband's writes the two without a hyphen, and the packs keep that double surname whole:

>>> parser_for(locales.CS).parse("Alena Pešová Řehořová").family
'Pešová Řehořová'          # default: middle 'Pešová', family 'Řehořová'

This was decided in chat on 2026-10-30, after measuring three populations. The parser gains in all three, and the cost is the same in each: a name with two given names reads its second given name into the family.

>>> parser_for(locales.CS).parse("Petr Pavel Novák").family
'Pavel Novák'              # a family comma keeps both: "Novák, Petr Pavel"

locales.SL and locales.HR are unchanged. two_given_names stays off. The default parser does not move.

How

• The switch. nameparser/vocab/cs.toml and sk.toml each gain a [rules] table with two_family_names = true. tools/vocab/build.py's PACK_RULES gains cs and sk, the build rewrites nameparser/config/_rules.py, and each pack carries policy=rules_patch(tag). This is how locales.ES carries its switch.
• The declaration. CS and SK now have a patch beside their choice of language, so the pack contract requires them to declare DEVIATES, covering what the patch changes. ES's declaration is exactly rules/O6's: three or more name words, or three or more runs of letters. I moved it into nameparser/_locale.py as deviates_by_two_family_names, and the DEVIATES of all three packs return it, so the three declarations can't drift apart. ES's behavior is unchanged. The second commit also fixes a gap the shared predicate inherited from ES: a word the tokenizer splits, beside a word with no letters (김민준 1, Jan\u200bPavel -), moved undeclared. Each word with no letters now counts as a unit, which fixes ES too.
• Tests (tests/v2/test_locales.py).
  • The CS and SK rotators for words the language choice turns on are rewritten as two-word names (MUDr Novák, Novák PhDr., Jana Nováková (roz. Svobodová)), because the patch alone moved the old three-word ones. The patch gets rotators of its own.
  • The 3f table gains pinned rows under both packs:
    • the two-word surname: plain, behind a title, under FAMILY_FIRST and before a family comma;
    • the two-given-name cost, plain and before a family comma.
  • St Clair Smith under CS now reads family Clair Smith.
  • A new test_cs_and_sk_pack_contents checks what each pack holds and that the three DEVIATES agree.
  • The non-interference gate, the choice-rotator test and the registry contract all hold, with and without the ja extra.
• test_vocab.py. Four build fixtures named a synthetic cs.toml without a [rules] table, which the build now refuses for a file a pack reads. They move to pl.toml, as Add the Spanish and Portuguese packs, locales.ES and locales.PT (3.0, 3g) #964 moved its es.toml fixtures.
• CLDR per-pack record. tools/differential/cldr_agreement_locales.json is re-recorded for cs and sk only (checked key by key). Whole renderings agreeing went from 14/24 to 12/24 for cs and from 15/24 to 11/24 for sk: the native double surnames are gained and the given-given-surname renderings lost. The default's records (cldr_agreement.json, GND, Wikidata) and the committed snapshot are unchanged.
• Docs.
  • docs/locales.rst: the cs/sk rows, a paragraph with doctests (the gain, the cost, the comma, the measurement, and how to switch it off while keeping the vocabulary).
  • The 3f release-log bullet, extended rather than adding a second bullet for one user-visible pack.
  • Decisions entries in vocabulary-model (dated, after the 3f entry), O6 (a Standing line and a Log entry) and deviates-registry.
  • One stale sentence in AGENTS.md.

Measured on the built tree (2026-10-30)

Run with local_agreement.py SOURCE [--locale cs] --base <bfb91163 checkout>. Each figure counts records agreeing in every field:

population	records	default	locales.CS before	locales.CS now	fixed	broken
cz-candidates (Czech election registers' boxes; data kv2022 2026-03-28, ps2025 2025-10-05)	170,181	96.96%	96.97%	97.47%	1,295	442
cz-candidates, against the default	170,181	96.96%	—	97.47%	1,318	440
ORCID researchers, country CZ (data 2025-10-17)	20,000	94.99%	94.99%	96.41%	635	352
ORCID researchers, country SK	12,962	94.84%	94.84%	96.40%	347	144

• Every record that now disagrees, in all three populations, has two or more words in the given box.
• Under locales.SK, cz-candidates scores 165,878. On ORCID the two packs read alike.
• Both corpora's default --base checks report "matches the base".
• A Czech or Slovak given-name list would win back most of the cost. A circular ceiling (a list built from each population's own given boxes) puts it at 97.72% for cz-candidates, 97.87% for ORCID CZ and 97.39% for ORCID SK.
• No population-wide source of Czech given names is published today. The Ministry of the Interior withdrew its frequency tables after the 2016 data, and ČSÚ publishes only the top 100 newborn names, under CC BY 4.0. The decisions entry records this as Open.

For the reviewer

A cost found in review: unlisted birth markers. A birth marker the chosen language doesn't list, standing between two surnames, is now taken into the family. Under both packs, Jana Nováková rod. Svobodová reads family rod. Svobodová, where the default reads middle Nováková rod.. Slovak rodená behaves the same way, and so does Czech roz. under SK. The listed roz. (under CS) and rozená (under both) still route the maiden name. A Slovak rod./rodená reading is vocabulary work, recorded as Open in the decisions entry rather than added here.

Packs agreeing on a scalar no longer warn. Before this PR, parser_for warned whenever two packs set the same scalar policy field, even to the same value, as docs/locales.rst documented. With CS, SK and ES all turning two_family_names on, the most natural stack, parser_for(locales.CS, locales.SK), would have warned although nothing was overridden, and failed callers running with warnings as errors. It now warns only where the later value differs. test_parser.py pins both sides, and docs/locales.rst and the decisions entry say so. This changes the warning's condition for every pack, not just these.

Checks

• uv run pytest: 33,647 passed (with the ja extra), plus -n0 -m serial (42).
• uv run mypy, uv run ruff check: clean.
• The suite on 3.15: 33,617 passed.
• test_locales.py and the CLDR tests with the ja extra installed.
• sphinx-build -b doctest docs: 570 tests, 0 failures.
• python -m doctest README.rst.
• The design-docs-reviewer agent found five problems, all fixed in this PR:
  • deviates-registry's Standing line still listed CS and SK as choice-only packs;
  • O6's Standing line claimed a gain "in every population", though CLDR's fictional renderings lose;
  • the reason given for CS and SK reading ORCID alike was false (now stated as measured: they differ on no record);
  • the RFC's surname-span line was stale;
  • _letter_runs gave a reason that no longer applies.
    It also surfaced the birth-marker cost above.

Resolves the draft issue "Should locales.CS turn on two_family_names?", which was never filed.

🤖 Generated with Claude Code