Read the Maghrebi, Amazigh and Hebrew family-name particles; add locales.AR and locales.HE (#1257) - #1373

Implements the decision on #1257 (its decision comment, 2026-11-12): where a particle reading ships turns on its script, its bearers and the mid-name test, and its kind on what the phrase it opens is in the names its reading serves.

What changes

clan is the kind the decision's rule 3 asks for. The README defines it as "a lineage prefix, where the word after it is a family name shared by the lineage rather than a parent's name", and no view reads it. Latin ben and Hebrew בן take it alike. One test decides the kind: a reading is patronymic where, among the bearers measured, the phrase it opens names the bearer's own father more often than it is a family name passed down unchanged, and clan otherwise. It binds words a change adds or reopens; bar, bin, binti, ibn, بن and בת are not re-judged. By it ben and בן are clan (ORCID's Tunisian family boxes 993, the Knesset's בן 15 of 15, against the 148 register chains #1248 counted). The test reads ould as clan too (family boxes DZ 29, MA 8, FR 9 against MR 20); it is patronymic by a choice decided on 2026-11-12 for the Mauritanian use, with the Algerian and Moroccan family names filling the view as its accepted cost. R9, the patronymic view, and decisions/vocabulary-collisions state it.

locales.AR and locales.HE are each a choice of their one language and nothing else. Each is registered, with rotators, a contents and pickling test, and CLDR's ar and he files in the locale view (20/24 and 18/24, unchanged by either pack). Three costs of the packs, each stated in the docstring and the docs:

• Choosing ar drops en's credential ABD for ar's bound given name abd (Jane Smith, ABD → given ABD).
• The packs are for surnames, not register-style chains (Karim ben Ali ben Salah Haddad → family ben Salah Haddad).
• parser_for(locales.AR, strict=True) reads a leading Ben as the family, and drops ber.toml's ait/nait unless ber is chosen too; locales.HE keeps Ben given under strict=True.

Each Latin word carries a cased mask. Family boxes write them with a capital (Ben 1,294 to ben 1, El 2,849 to 13, Ait 203 to 1, Ould 36 to 0), and without a mask case repair lowered them as particles (Karim ben Salah).

el: on by default, a decided exception to rule 2

el fails rule 2 (Algerian compound given names) and is read by default anyway, on the measured trade decided after the gate. Rule 2 stands. Fixed/broken at this PR's head, the word removed as the comparator (config_diff.py score "expr:Parser(lexicon=Lexicon.default().remove('el', kind='article'))" default), against al's article measured the same way; each group is capped at 20,000 records unless marked:

Algeria's breaks are the article inside a compound given name (Nour El Houda Haddou → family El Houda Haddou), and a bound given name before it is the same shape (Abd El Kader Haddou → given Abd, family El Kader Haddou, where 2.2.0 through 3.0.1 read given Abd El, middle Kader, and 1.4.0 through 2.1.0 given Abd, middle El Kader). Both are Accepted under rules/P2 (particles join forward) and in decisions/vocabulary-collisions. #1378 asks for the vocabulary that lifts them.

Canaries, run against master:

• El Greco and El Mehdi Haddou keep given El, reported.
• Haddou, Sara El → family El Haddou.
• Sara El-Haddou, Juan de el Valle, Juan del Valle and Maria do Carmo Souza do not move.
• Maria El Salvador → family El Salvador.
• Lower-case Sara el Haddou → family el Haddou.

What moved

Measured at this PR's head against master's readings, rebuilt in the tree as an expr: lexicon. Its compiled sets equal those of master 161ebd28 and of the merge base 73718fd9 (36 sets compared). The shared tools cache was only read; match_diff.py wrote its kept base to a cache folder of its own.

• Committed snapshot: eight rule examples added (git diff 73718fd9 HEAD -- tools/differential/corpus_rules.jsonl). Three names built on Zein El Dine moved with el (family Dine → El Dine; neither is the whole name, so unclear), and no other. The ben given reading and the off-by-default particle were each byte-identical in their own commits.
• Local corpora (config_diff.py diff): JRC-Names 3,483 of 1,315,051 spellings and opensanctions-parliaments 16 of 24,192, nearly all the particle joining the family. Before el (commit 08629c8 and earlier) the figures were 287 and 15.
• Scored (config_diff.py score; ORCID and PatentsView by country, PSC by nationality, each group capped at 20,000 records; ORCID's extract holds at most 20,000 a country), fixed/broken:

Agreement in every box, ORCID Tunisia: 87.6% by default, 96.1% under either pack. Uncapped, ben read as a particle breaks 148 PSC British, French and Tunisian filers against 120 it fixes (--cap 5000000), British middle names Ben, which is why it is off by default.

• Name matching (match_diff.py --base <73718fd9>): 49 pairs move under each of the 21 settings, every one holding el. 22 become consistent: spellings of one person writing El and al, or the article with and without its hyphen. One becomes inconsistent only in the setting that removes particles-aside. No consistent pair shares no key. Recorded in decisions/name-matching.
• Group suggester (suggest_measure.py --base): five cells moved. Four are in the capitals table (orcid-famcaps deciding capitals_as_family at 2 names 77.3% → 77.0% on either base; families read right without the switch 88.9% → 90.0%; orcid-givcaps' families 4,256 → 4,260). One is in the order table: orcid-givcaps' given-first anti-goal row at 3 names, 8.3% → 8.7%. All are listed in decisions/group-suggestion; the user-facing claims stand.
• Other canaries unchanged: Ben Affleck, Van Johnson, Al Gore, Park In Hwan, Nguyen To Nga, John V. Smith.

Accepted

Recorded in rules/P1 (the lone particle fold), rules/P2 (particles join forward), rules/R9 (the patronymic view) and decisions/vocabulary-collisions:

• A listing written family first without a comma reads whole as the family where a never-given ait or ould opens it (Ait Haddou Karim), the fold's existing cost.
• A family name written with Hebrew בן first and no comma, under a family-first order, gives בן alone to the family (בן חיים יעל → family בן), where master read בן חיים. Should Hebrew בן opening a name be the given name Ben? (בן כספית gives family 'בן כספית' and no given name) #1209 counted 8 Knesset family names written so; a comma keeps them whole. Accepted (decided 2026-11-12) as a cost of the given reading.
• ould fills the patronymic view in an Algerian fixed family name (Karim Ould Haddou).
• An Arab name transcribed in Hebrew script with בן for the father leaves the patronymic view empty, as a Tunisian register's chain does with ben.
• A family name written alone with בן reads given בן (Should Hebrew בן opening a name be the given name Ben? (בן כספית gives family 'בן כספית' and no given name) #1209).
• אל inside a given name, and El inside an Algerian compound given name or after a bound given name, give the article and the rest to the family.
• The packs' three costs above, and a strict Arabic choice reading ben as never given (with cheikh in the reviewed table of words strict ungates; strict Hebrew keeps it given through he.toml's own given reading).

Not changed: moved to #1378

• Latin abd, abdul, abu and Hebrew עבד with rules/P12's (bound phrases) completed_by for the names after the article.
• The Arabic compound given names behind el/al as rules/P10 (a given name a particle completes) words.
• Arabic-script ايت/ولد, which Should Ait, Nait and Ould be family-name particles? (Youssef Ait Taleb gives middle 'Ait', family 'Taleb') #1249 asked about.

Also left out:

• N'Ait, which Should Ait, Nait and Ould be family-name particles? (Youssef Ait Taleb gives middle 'Ait', family 'Taleb') #1249 lists. It reads as one word: given N'Ait opening a name, middle N'Ait inside one.
• בת and the excluded בר, unchanged.
• Parser.revise never fills the patronymic view, bin included; this PR leaves that as it is.

Checks and reviews

• Run here: uv run pytest, -n0 -m serial, mypy, ruff check, sphinx-build -b doctest and -b html, python -m doctest README.rst and build.py --check.
• Python 3.15 in its own environment, without namedivider, plus the locale, CLDR, Wikidata-native and snapshot tests with the ja extra.
• Reviewed in this session:
  • the measurement gate before any vocabulary edit;
  • sweeps per word over comma shapes, the three orders, case, capitals_as_family, initials(), capitalized() both ways, Parser.revise and Parser.matching().check both sides;
  • the 1.4.0, 2.0.0, 2.1.0, 2.2.0, 2.8.2 and 3.0.1 wheels for every release-log and case-row before/after claim, re-run after the second review found the Abd El Kader Haddou claim wrong for 1.4.0 through 2.1.0 (fixed).
• The independent code, test and design-docs reviews were run on the PR, and this commit answers them.

#1210's parent issue is #1253 (decisions/family-without-an-inherited-name), which deferred it to #1257.

Closes #1209
Closes #1210
Closes #1211 (the Hebrew עבד with completed_by is in #1378)
Closes #1240
Closes #1248
Closes #1249 (N'Ait and the Arabic-script ايت/ولد are in #1378 or left out above)
Closes #1257

🤖 Generated with Claude Code