[derek73]

Closes #1237. Refs #1253.

Adds locales.MN, a Mongolian pack. It reads a name written wholly in Cyrillic with the father's name first, so the father's name is the family, as #1253's first statement has it. A romanized name keeps the caller's order.

from nameparser import parser_for, locales
mongolian = parser_for(locales.MN)
sukhee = mongolian.parse("Батын Сүхээ")
sukhee.family, sukhee.given            # ('Батын', 'Сүхээ')
mongolian.parse("Sükhee Batyn").given  # 'Sükhee': a romanized name keeps the caller's order

1.4.0 through 3.0.1, and the default still, read Батын Сүхээ as given Батын, family Сүхээ.

Changed since the first push (decided in coordination, 2026-11-12)

The first version declared FAMILY_FIRST_GIVEN_LAST for every name. Its own gate found that Mongolians write their own romanized names mostly given-first (65 of 87 two-word ORCID credits), so the order is now keyed to Cyrillic.

What the pack is

• A plain choice of mn, plus a script_orders entry for Script.CYRILLIC, set to FAMILY_FIRST_GIVEN_LAST.
  • The choice changes no reading, because mn.toml's readings are all on by default.
  • _INERT_CHOICES exempts MN from the two tests that expect a choice to change a word.
• Given name last, not second, so a word written before the father's name (a clan name, Боржигин Батын Сүхээ) leaves the given name in place. The clan name itself is not read as one.
• Not read:
  • the genitive, into the patronymic view (Decision 1);
  • the glued official form Ц.Сандаг-Очир (Should a Mongolian choice read a glued father's initial as the family? (Ц.Сандаг-Очир gives given 'Ц.', family 'Сандаг-Очир') #1306), which rules/T8 already divides: family Ц., given Сандаг-Очир;
  • the traditional Mongolian script, which is not Cyrillic and so reads as it does without the pack. Its default order is decisions/T6's open question.

What the Cyrillic order needed

• Script.CYRILLIC joins the enum and _policy._SCRIPT_RANGES: the Cyrillic block, plus Supplement and Extended-A, -B and -C.
  • The readers that meant "the East Asian scripts" when they read the whole table now read _EAST_ASIAN_SCRIPTS:
    • rules/T3's interpunct flank and the two-family-names run count (_classified_char);
    • the segmenter's neighbour test;
    • the CJK corpus builder;
    • the case table's CJK guard.
  • DEFAULT_SCRIPT_ORDERS and _NO_INITIALS gain no entry.
  • _classify asks the Cyrillic matcher only of a word that opens with a Cyrillic letter, so a non-ASCII Latin word costs nothing extra.
  • _assign._effective_order returns the declared order at the first word whose script, outside the kana licence, has no script_orders entry. Such a script resolves alone to itself or, beside any other, to None, so the answer cannot change. Under the default parser a Cyrillic name therefore costs what it did on master: Иван Петров 233 frames and Иванов Иван Иванович 353 (242 and 370 before this return), and John Smith 126 on both. A guard in tests/v2/test_benchmark.py pins one effective_script call per name from that walk.
• effective_script reads a Cyrillic word with a hyphen or an apostrophe inside it (Оюун-Эрдэнэ, Мар'яна) as Cyrillic. The table stays letters only.
• PolicyPatch.script_orders composes by script, where it was a scalar override.
  • A patch's entry replaces the base's entry for that script, and the other entries stand. So the pack adds its entry without erasing a caller's own: Policy(script_orders=()) with MN keeps only the Cyrillic entry.
  • parser_for warns where two packs set one script to different orders.
  • This reaches any caller's own PolicyPatch, not only MN's, so the release log carries it as a Change. In 2.1.0 through 3.0.1, Policy().patched(PolicyPatch(script_orders={Script.HAN: GIVEN_FIRST})) kept the Han entry alone; now it keeps the Hangul and hiragana entries too.
  • An empty patch no longer clears the table. Clear it with Policy(script_orders=()) on the base parser.
  • A malformed patch still reaches Policy's curated messages, an unhashable key included.
  • Nothing shipped patched script_orders before, which is why _policy.py's comment had said to revisit this when a pack needed it.
• A lone word stays rules/O5's. The Cyrillic entry orders a longer name and settles no lone word's field: EffectiveOrder.by_script is set for an East Asian script alone. So Сүхээ under MN reads given and reports given-or-family. A caller's own katakana entry still settles a lone word, as Han's does; a test pins both.
• No [rules] table. The table's keys are scalar Policy fields, and a mapping keyed by Script would need a value type that RULE_KEYS and _locale._RULE_VALUES lack. The pack carries its patch in code, as ZH and JA carry segment_scripts.
• The licence, worded so no one reads "Cyrillic means family-first":
  • rules/W4's When and Then now name "a script a chosen locale pack gives an order". Its Then bullet says the pack's order is the caller's choice, not the script's, and that O1 and O2 stand down under it.
  • decisions/W4 records the decision.
  • Root AGENTS.md's design-philosophy paragraph says a pack's script order is never a default, ending "Never read it as 'Cyrillic means family-first'".

Measured (2026-11-12, master 2b39f23)

Source	Script	default	FAMILY_FIRST	the pack
State Great Khural (mn_parliament, from the cache)	Cyrillic	0 / 126	126 / 126	126 / 126
JRC-Names, six Mongolian people, classified two-word spellings	Cyrillic	0 / 9	9 / 9	9 / 9
same	Latin	3 / 45	42 / 45	3 / 45, left to the caller
ORCID MN, 150 Latin free-text credits, against the holder's family box	Latin	73	25	73

• CLDR's mn sample renderings (all Cyrillic, rendered in every order) read 19 of 24 by default and 15 of 24 under the pack.
• With no pack chosen, no name moves:
  • snapshot.py reports no moves; the snapshot only gains rows for the new rules examples.
  • snapshot.py --local <each> --base <master checkout> reports "matches the base" for all 16 local corpora.
  • Two reviews compared 47 mixed-script and edge names under 10 configurations against master, and readings, ambiguities, initials() and warnings were identical.

Pack unions (asked for plainly)

The pack reads every wholly-Cyrillic name its way, Russian, Kazakh and Kyrgyz ones included.

Name	MN+RU	MN+TR_AZ
Иван Сидоров	family Иван	family Иван
Сидоров Иван Петрович	given Петрович, middle Иван, no patronymic	given Петрович
Нурсултан Абишевич Назарбаев	family Нурсултан, patronymic Абишевич	family Нурсултан
Бекмырза Чолпонбай уулу	given уулу	given уулу, patronymic Чолпонбай уулу

Proposed precedence, not built: a patronymic shape the active rule reads (an O1 ending, an O2 marker) keeps its name from a pack's script order. None of the Khural's 126 members carries either, so the pack would lose nothing there. It is recorded in decisions/W4's Open:.

Decisions for the maintainer

1. No genitive rule.
   • In Cyrillic it would only repeat family.
   • In Latin, now that the order is the caller's, it could order a romanized name the genitive marks: Should a Mongolian choice read the father's name first? (Ухнаагийн Хүрэлсүх gives given 'Ухнаагийн', family 'Хүрэлсүх') #1237 counted 515 Wikidata labels with the genitive on the first word against 1 on the second.
   • But the romanized endings collide with given names written first (JRC: Mervyn King, Carolyn Porco, Kathryn Gustafson).
   • It would also need a new PatronymicRule and a Policy change.
2. FAMILY_FIRST_GIVEN_LAST against plain FAMILY_FIRST. Given-last protects the clan-name shape. It costs a three-word Russian name written family first: Сидоров Иван Петрович reads given Петрович. Neither shape is measured.
3. The union precedence above.
4. Accepted costs:
   • Every wholly-Cyrillic name reads the pack's way, so the pack is for Mongolian data.
   • Сүхээ Б. reads family Сүхээ.
   • A Cyrillic letter with a combining stress accent, which has no precomposed form, leaves its word unclassified, so that name keeps the declared order.
   • Ноён (Mr.) is no listed title, and neither are Доктор and the honorific гуай: vocabulary candidates, not filed.
   • A Latin letter that looks Cyrillic inside a Cyrillic word (Cүхээ Батын, U+0043) leaves the name to the caller's order.
   • Matching. The father's name written as an initial, the usual short form (Батын Сүхээ / Б. Сүхээ), is the family under the pack. So the pair reads family-initial and conflicts. The default matcher calls it consistent, but only by reading the father's name as the given name. Accepting family-initial restores it; docs/languages.rst shows the accept as a doctest.
   • A pickle holding Script.CYRILLIC does not load in earlier releases.

Also in this PR

• tools/differential/suggest_measure.py reads each local corpus through local_corpora.load's checks, so a stale or edited corpus stops the run (with tests).
• mn_parliament was added to the parliaments tool here first, then split out as Add Mongolia's Khural to opensanctions-parliaments #1368 so every checkout could read the refreshed shared cache. Add Mongolia's Khural to opensanctions-parliaments #1368 is merged, and master is merged in here (e944bb3), so the duplicate is no longer in this diff.

Checks

• uv run pytest with the ja extra, on the head (1743756): 46,080 passed; -m serial 35; sphinx-build -b doctest 840. On the rework commit: without ja, 46,034 passed; on Python 3.15 in its own environment, 46,034 passed.
• pytest -n0 -m serial: 35 passed.
• mypy, ruff and tools/vocab/build.py --check: clean.
• sphinx-build -b doctest and -W -b html: clean.

Reviews

Each review ran as a subagent in its own detached worktree, on the pushed head of its round, with a read-only master checkout beside it.

Round	Review	Found	Done
1 (b4aee8e)	design-docs-reviewer	Enkhtaivan Erdene cited as written family-first, though its boxes say given-first. #1237's option 1 misquoted (it named family-first). The JRC "9 of 9" and "Latin and Greek" overstated what the recipe classifies. "Three in four" taken from the wrong population. The spaced-initial and population counts did not reproduce (8, not 6). The traditional-script example had lost its U+202F.	All corrected in f0bfd08, with each recipe's filter stated.
1 (b4aee8e)	code-reviewer	Missed roster entry in docs/contributing.rst. The inert choice is not recorded in decisions/deviates-registry. "snapshot unchanged" is literally false (one row added). The given-last order's cost to Russian names is unweighed. Plain-space traditional script is untested. docs/policy.rst names only vi.	All fixed except policy.rst, which is another PR's file. The Russian-name cost is now Decision 3 above.
1 (b4aee8e)	comment-analyzer	The glued initial is divided by script_segment (rules/T8), not tokenize. The U+202F example. contributing.rst. A wrong NFD test comment. mn.py claimed a "gate re-take" the entry did not record. "yet" in locales.rst. A misordered clause in languages.rst, with a romanized example that was not the Cyrillic one's transliteration. The strict-mode comment and the DEVIATES "so" were imprecise.	All fixed.
1 (b4aee8e)	pr-test-analyzer	No critical gaps; mutations confirm each test bites. The strict choice was untested. Readings the entry states were unpinned. The non-interference gate is vacuous for an all-declaring pack, as it already was for VI.	Added a strict test, a table of the entry's stated readings, and seven more default-equal comma and title rows. The vacuous gate is VI's shape and left as is.
round 2 (f0bfd08)	code-reviewer	One fresh error: the new Russian-name bullet blamed O1 for the empty patronymic view, but O1 stands down under either declared order; the cause is that R9's East Slavic rule reads only a middle word, and the given-last order puts the patronymic in given. Everything else re-verified (ORCID and JRC counts over the stated filters, the U+202F, T8's division, the snapshot row, the doctest and prose, the new tests' mutations).	Fixed in a3506a1.
rework (c9f11d2)	code-reviewer	Found no default reading moved through Parser() or any shipped pack, but three problems. A caller's own PolicyPatch(script_orders=…) now composes differently, with no Change bullet. An unhashable malformed key lost Policy's curated message. The Cyrillic matcher cost a frame per non-ASCII Latin word.	All fixed in 33efcf5.
rework	type-design-analyzer	The same malformed-key regression and missing Change line. The composition's metadata had no constant, and no test pinned it. The by_script docstrings contradicted the code. The katakana vs Cyrillic lone-word asymmetry was undecided. Stale autodoc docstrings. segment_scripts accepted CYRILLIC with no word on what it does there.	Fixed: a _BY_SCRIPT constant with a test, rewritten docstrings, a test pinning the katakana and Cyrillic rows, and a segment_scripts note.
rework	design-docs-reviewer	The pack clause was filed under rules/W4's Except, but it covers names outside W4's When. The composition change was undocumented. decisions/deviates-registry still named mn.toml's [rules]. The snapshot gains three rows, not one. O1 says nothing of the pack.	W4's When and Then now state it, with a Then bullet saying O1 and O2 stand down under it. The rest is fixed.
rework	pr-test-analyzer	Seven of fifteen conditions survived mutation. Two were reachable without the pack: the interpunct flank (Гал·ла Иванова would split) and the segmenter neighbour test (山田太郎 Иван would stop dividing). The per-script warning's filter was untested, and so were the joiner edges. An unreachable empty check. Stale comments: the Cyrillic range, EffectiveOrder on katakana, the joiners, build.py. Glued initials (Б.Б.) and stress marks escape the pack's order.	Fixed in 7ac557a: a test for each survivor (each checked to fail without its condition), the neighbour test routed through _EAST_ASIAN_SCRIPTS as decisions/W4 says, the empty check dropped in 33efcf5, the comments corrected, and the two escapes recorded as accepted.
fix (33efcf5)	code-reviewer	One fresh error: the new segment_scripts docstring called CYRILLIC inert there, but a configured segmenter divided Cyrillic words, and the construction warning asked for one. Everything else held: 13 malformed-patch shapes give master's messages, the joiner edges, the frame counts (José García 177 as on master), the release-log claim against 2.1.0, 2.8.2 and 3.0.1, and W4's citation.	Fixed in 3301938: a Cyrillic word is no segmentation site, and the warning leaves CYRILLIC out, with a test that fails without the change.
fix (7ac557a)	code-reviewer	Nothing at confidence 80 or above. The neighbour gate reads exactly as master's is not None on 46 tokens, Cyrillic aside; every new test fails under its mutation; the decisions sentence holds when run. Cosmetic: two stray blank lines, and effective_script's docstring missing the joiner path.	Tidied in e95d3d0.
merged (e944bb3)	independent review (coordinator)	No defect: with no pack chosen, 1,227,135 grid rows read as master. Three follow-ups. The default parser walked every word of a Cyrillic name (+9 and +17 frames). The pack's matching cost for a father's initial was undocumented. The user docs did not say the patronymic rotations stand down beside mn. Also: no direct test that the default keys no Cyrillic order; a homoglyph escape and two honorifics to record.	All done in 1743756: the early return with a frame guard, a matching doctest in languages.rst, a composition doctest in locales.rst, a default test, and the #1237 entry records the rest.

🤖 Generated with Claude Code