State the conventions corpus's parts, and score them like CLDR's

[derek73]

Follow-up to #686 / #691: the conventions corpus gets a fields file after all, so a change that moves one of its names shows whether the reading moved toward the expected one or away from it.

What.

• 30 of conventions.py's 35 rows carry fields in CLDR's vocabulary, beside their expect note.
• The other 5 carry an unstated reason. The builder refuses a row with neither or both, so a row can't lose its fields unseen.
• build_conventions_corpus.py writes conventions_fields.jsonl.
• cldr_agreement.py scores it as a fourth source, recorded in conventions_agreement.json.
• The scorer gains one field, honorific, scored inside suffix.

Where the readings come from. The research the rows were written from: the 2006 UK naming-practices guide and the Cultural Atlas report, both in the gitignored docs/superpowers/, read back from the sessions that wrote #686. The fields follow the Wikidata source's conventions: a patronymic is written with its connective (bin Yusof) and scored inside given_names, and a surname2 inside family.

Your calls.

• An honorific written after the name (Bey, ssi, san, -san, úr) is a suffix. That matches the design notes and nameparser's reading of さん; CLDR's さん-as-title disagreement stays recorded.
• Kaur is given2.
• Péterné is the given name.
• H. Budi Santoso is title + one personal name.
• Ma.'s Cruz is a surname2.
• Vda. de Mora sits inside the surname.
• wa is left unstated beside arap/ole.

Left unstated (still in the corpus, still watched by the snapshot), since the scorer reads an unstated part as an empty role:

• the kunya Abu Karim
• Szabóné Kiss Éva
• arap/ole/wa, whose last word is a parent's name or a family name depending on the family

Guards added in review, since this source is hand-written:

• A field name no rule reads now raises, for every source; CLDR's locale and given-informal are skipped on purpose. A mistyped honorfic had been passing as agreement.
• test_the_fields_partition_their_rendering: every word of a non-CLDR rendering is stated by exactly one field. GND, Wikidata and conventions already held it.
• The scorer has a FAMILY_FIRST_GIVEN_LAST parser, so such a row no longer crashes it.

Measured.

• 5 of 30 rows agree in every role: Harpreet Kaur Dhillon, Szabó Péterné, Kim Min Ho, Ivan Petrov Dimitrov, Lim Mei Ling Jessica.
• All 25 disagreements are real misreadings. The largest group is a patronymic misplaced, usually read as the family name (Hassan bin Yusof → family bin Yusof), which makes the record a worklist for the 3.0 patronymic field. Titles read as given names (U Kyaw Win) and honorifics read as name words are the next groups.
• Re-recording leaves the CLDR, GND, Wikidata and locale records unchanged, and the corpus file is byte-identical.
• Each new guard was checked against a bad row and fails on it.
• Full suite: 19,330 passed; mypy and ruff clean.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.03%. Comparing base (c89629b) to head (4876318).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #692   +/-   ##
=======================================
  Coverage   99.03%   99.03%           
=======================================
  Files          46       46           
  Lines        4541     4541           
=======================================
  Hits         4497     4497           
  Misses         44       44           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.