[derek73]

Part of #704: the default reading the first PR (#1057) left open. This moves default parses.

What changes

New rule P10 (a given name a particle completes). Behind a one-word given name read given-first, a particle chain that opens with a particle run and a word whose given reading is written after_particle reads that run and word as a middle name, and the chain's other words as the family.

Name	1.4.0 – 2.8.0 / master	Now
María del Carmen García	family del Carmen García	middle del Carmen, family García
María de los Ángeles López Pérez	family de los Ángeles López Pérez	middle de los Ángeles, family López Pérez
Maria do Carmo Silva	family do Carmo Silva	unchanged (carmo not marked)
Domingo Santa María González	family Santa María González	unchanged (saint clause)
María del Carmen	family del Carmen	unchanged (no word left for the family)

• Saint clause, decided in chat: a run holding a saint particle (san, santa, st, ste) never forms a phrase. Santa María is a surname: in ORCID, a saint particle stands before a marked word in 0 given boxes and 7 family boxes. The saint kind now compiles into an internal SAINT_PARTICLES set, the way PATRONYMIC_PARTICLES does. This keeps the de María given phrases, which dropping maría from the list would have lost.
• O7 (two given names) now takes the phrase through its ordinary middle path, so the phrase path Read a particle phrase as the second given name; leave Kaur and Singh to neither two-name switch (#704) #1057 added is gone. So are its vocabulary clause and its "split candidate" note. Under the switches, only the saint clause moves a corpus name: 2 of 26,576 readings across every order and switch pairing, both Domingo Santa María González.
• Ambiguity alternative fix: a word behind a particle chain now names the field of the chain's last word, where the word would join. P10 moves the chain's first word to the middle, so the old anchor, the first word, gave the wrong field. Example: Jairo do Valle Cardoso Filho reported Filho as "name as middle"; it now reports "name as family". The TSE run found this.
• Parse cost: a name holding no marked word leaves the new rule after one plain loop. The first draft added 25 frames per parse, which the benchmark band caught.

Measurements (2026-11-04, against master 3b293777 with --base)

Population	Fixed	Broken	Changed
PSC (local_agreement.py psc-individuals --base)	90	8	7
ORCID (local_agreement.py orcid-researchers --base)	3,881	46	164

• TSE (snapshot.py --local tse-donors --base): 1,561 of 854,531 Brazilian names move, in family and middle only (for example ADRIANA DO SOCORRO MARTINS XAVIER). That corpus has no fields, so this measures reach, not agreement.
• What breaks is the marked words borne as surnames: ORCID del Valle 18, del Rosario 12, de Jesús 5, de Fátima 4. P10's Accepted section records this.
• Snapshot: 5 names move. CLDR records: gain one Wikidata rendering, María del Pilar Hurtado Montaño, re-recorded.

Review round 1 (four reviewers; fixed in 280179d)

• shared now affects a default parse. P10 never completes a shared name, so Kaur del Carmen García keeps family del Carmen García by default. Every doc that said shared moves nothing with the switches off now names P10.
• The reach is mostly Portuguese. The 1,561 TSE moves are mostly de Lourdes 519, do Socorro 491 and de Fátima 228. ORCID's Portuguese-speaking countries: 53 fixed, 6 broken. The release bullet, rule and decisions now say so.
• P10's Signal line now names its saint and given-name lookups, and its patronymic clauses cite R9. The family-first example was replaced, since it showed the outcome P10's Then predicts.
• Alternative anchor: O6's family-first cut splits a chain the other way, so naming the field from the chain's last word was wrong there. See round 2.
• New test rows: O6 over a family a saint particle opens, and P10's guards under the default policy (a family comma, a patronymic, a mark).

Review round 2 (9598f27)

• Alternative anchor fixed rather than accepted. The second review found 42 rows that the first fix newly got wrong, where master had them right (FAMILY_FIRST + two_family_names + middle_as_family). Beside now carries both ends of the chain. Assemble names the last word's field only where that word stayed in the family under a head that left it, which is P10's split. A last word P6 attached to the family keeps the head's field.
  • On the review's grid of particle chains, 132 alternatives move against master, each to the right field.
  • Across every committed corpus name, under every order and switch pairing (66,470 readings), no alternative moves.
• Docs: stale O7 lines in decisions/O6 and O7; the Catalan dels question reopened under decisions/P10 (its old reason is void now that P10 reads the phrase by default); recipe filters stated; the singh note; ste added to the saint lists; P10's Signal line and its P3 citation; and the pipeline AGENTS.md lesson restated for both split directions.

Docs

• New: rules/P10 and decisions/P10.
• Amended rules: O7 (When and Then restored, phrase clause folded into the unit bullet, Signal), O6 (Except now cites P10), P2 (new bullet), and mechanisms.md (UNIT-PARTITION's list of chain cuts, DECLARED-SPAN).
• Decisions: decisions/O7 gets a new entry, plus a head note on the 2026-11-03 entry it supersedes in part.
• User and vocab docs: vocab README (saint row, after_particle row), AGENTS.md compiled-set bullet, nameparser/_pipeline/AGENTS.md (the Beside lesson), customize.rst (doctested), and a release-log Behavior Changes bullet.
• Budget: rules_budget.json re-recorded (P10 new; O7 shrinks).

Checks: full suite, serial benchmarks, mypy, ruff and sphinx doctest pass.

🤖 Generated with Claude Code