[derek73]

Step 0 of the release checklist for 3.0.0: the stale-prose sweep. Each site the checklist names was read against the 3.0 tree, with every doubtful claim run. It changes no behavior: only docs, docstrings, message strings, comments and case notes.

What changed, by site

docs/migrate.rst

• "Changed in 3.0" listed only the render change. It now has 13 shapes, each with a doctest, for the 3.0 bullets whose 1.4 reading moved: Graf Lambsdorff, Sta. Maria / Delos Reyes, María del Carmen, طارق عبد العزيز, Ortega y Gasset VI, the bracketed clauses (PhD, MBA) / (JP), Dumas père, Pr / Profesor, KC / MSW, U+2063, the given names Maj / Gen / Pharaoh / Saint with Brig Gen, Silva, Junior Carlos, and the case repair of ST, DELA and MACHADO. Each 1.4 reading was measured against the 1.4.0 wheel.
• "Finding the sites" now says to upgrade to 2.8.2, not 2.8.1. It adds the config.surnames import and the unwarned read of a Lexicon set field.
• The intro no longer says "Further output changes come with the 3.0 alphas". There are no alphas.

README.rst

• The 3.0 announcement mentions reading-based Lexicon configuration, render in written order and the language packs.
• The upgrade advice says 2.8.2 and names the unwarned Lexicon field read.
• python -m doctest README.rst and PyPI's readme_renderer are both clean.

docs/usage.rst, concepts.rst, modules.rst, index.rst

• References to internal Lexicon fields (after_given_titles, suffix_acronyms_ambiguous, given_name_abbreviations, emptying titles) became the 3.0 calls (words(...), readings_of(...)).
• 1.x last_list became family in STABLE_TAGS.
• "2.0 API" framing was dropped where it no longer holds.

docs/customize.rst, locales.rst

• ES_AR is mentioned beside ES's given-name list and in the composition paragraph.
• render()'s description covers the views and the default layout in read order.
• The pack intro and "what a pack is for" include packs that choose languages.
• The CLI paragraph covers language tags and --strict.
• The rules_patch users are listed in full.

Autodoc docstrings and messages (_types.py, _locale.py, locales/__init__.py, locales/ja.py)

• AmbiguityKind docs no longer cite internal Lexicon.title_phrases and conjunctions_ambiguous, or "planned for 2.x".
• surname1's folded-middle example was wrong against rules/R11 (the first-surname views) and is fixed.
• Locale names its languages.
• The ja ImportError quotes 'nameparser[ja]' for zsh.
• About 50 raise and warn paths were triggered, and every snippet they offer ran (22 also under mypy --strict). The one not run is segmenter=locales.ja_segmenter(): namedivider-python isn't installed in this checkout.

Pipeline stage headers (_assemble, _assign, _classify, _post_rules, _script_segment) and ParseState's ownership map

• Consumes/Produces/Reads now match the code, checked with the checklist's grep plus aliased reads.
• _classify over-claimed given_names and missed title_case_name_words.
• _assign reads the title-phrase sets for H9 (the title phrase).
• post_rules writes S1's set-apart roles and R11's first-surname tag.

tests/v2/cases.py notes

• 9 notes were made stale by rules.md#R3's "a conjunction never initials" is unqualified: Juan de y has base y and initials J. #461 (connective initials), H9's justice, Set apart a bracketed clause of credentials, generations or titles (#792, #771) #1067's bracketed clauses (rules/S13's twin examples, rules/S1), R6's rendering and Should Doe, John MA read MA as a credential, as 1.4.0 did, or at least report the fork? #531. Each was rewritten from a live run, checked against the 2.8.2 wheel where a past reading is named.

docs/design/

• Open: lines were repointed from closed issues: Should Doe, John V report that the V might be a middle initial? #877 and Should a bracketed generation end there? "W. Russell (Jr.) Withers" reads family "Russell", suffix "Jr. Withers" #792 are decided, and Should common British post-nominals (KC, FREng, DL, FRS, PC…) be in the vocabulary, so Jane Doe KC keeps family Doe? #965's leftovers moved to Should a credential written in its own mixed case (FREng, CEng) end a name as the credential? Jane Doe CEng FIET reads family 'CEng' #1053.
• Standing lines moved from plan tense to shipped in 3-0-reevaluations, vocabulary-model and decisions/H9, with a Not standing: line for the v1 title list.
• rules/C3 (counting the words before the comma): its Accepted example moved from Graf, a particle since Should Otto Graf Lambsdorff read family Graf Lambsdorff, as Manfred Freiherr von Richthofen reads family Freiherr von Richthofen? #709, to Herzog.
• rules/H9 (the title phrase): its Precedence bullet is in 3.0 terms, remove(w, kind="given") and a caller's other title reading. Its budget record was re-recorded, 778 → 784 words; the clause reads no new signal (axis 12).
• The S Background notes that a multi-word suffix is refused at add().
• The design-docs review ran on the diff. Its three findings are fixed in the last commit.

AGENTS.md files and nameparser/vocab/README.md

• Root:
  • the honorifics view in the snapshot list;
  • a bullet for the views' kinds (GENERATION_WORDS and the rest);
  • _types importing _parser at call time;
  • three missing ADDED_IN fields;
  • the default Parser no longer called "(future)";
  • stale names in the frozen-constant, Cyrillic-regex and perf gotchas.
• nameparser/AGENTS.md: _INITIAL no longer said to be public.
• vocab/AGENTS.md: the saint and rank capitals from Should case repair write St Clair and San Martín with their capital? JOHN ST CLAIR repairs to John st Clair #1069.
• docs/AGENTS.md: no public vocabulary module left to :mod: link.
• vocab README: its default = false example was wrong (père has been on by default since Locale packs for Alexandre Dumas fils and Jiří Dienstbier ml.: generational words that collide outside their language #684) and is now cs ml. The phase-3 lines are in the past tense.

Decided items, applied

• rules/P9 (a particle that is also a suffix word, beside a family comma): it gains fenced examples for SMITH VD V, JOHN and its one-case smith vd v, john, each with deviates: #1080. Ask one predicate for every roman-numeral rule (#752, phase 2f) #886 moved both readings in 2.7, and the fix is in 3.x. The rules corpus and the snapshot gain the two names; no reading moved.
• decisions/O6 (two family names): its Open line carries only the ñ question (Should "Maria Begona Lopez" under locales.ES keep Begona out of the family, as "María Begoña López" does? #1081). The per-region clause is dropped: ES_AR answers the case Should two_family_names leave a known given name out of the family, so "José Luis Pérez" keeps family "Pérez"? #959 raised.
• decisions/vocabulary-model, phase 2's pool: the Standing line is now shipped tense, matching the entries that record it. 2a–2m shipped piecemeal, 2g's credentials landed through the pool, and the gate stays at its own sites (item 4). Since 4i-b a configured Lexicon compiles its pool. 3a's reading index is not that pool: it holds the vocabulary files' readings only.
• Retargeted to "unbuilt, 3.x":
  • The Policy .pyi stub, with a dated 3-0-reevaluations entry; the differential-ledger Standing line points at it.
  • The reserved vocabulary names bare_only, leading_only, trailing_only and separators, with a dated vocabulary-model entry. The change covers tools/vocab/build.py's RESERVED and its message ("reserved, unbuilt, for 3.x"), nameparser/vocab/README.md's table, the RFC's two present-tense lines and test_vocab's rows together.
• 2.x translation messages:
  • remove(suffix_acronyms={'ma'}) now offers lexicon.remove('ma', kind='credential'), keeping the Filipino same_as='maría'. The translator first tries leaving alone a kind one of whose readings already holds what is wanted, keeping the exact-compile check, and it carries a held same_as through. The 2.x pickle path keeps its old skip.
  • A bare Lexicon(surnames={'kim'}, titles={'dr'}) offers Lexicon(surnames={'kim'}).add(...), so the offered code no longer calls an unbound lexicon.
  • New tests run the offered code: a bare constructor, a replace of an edited lexicon, and ma. Three of the rows fail on the pre-fix code.
• API reference:
  • "Added in 3.0." is on every public member new since 2.8.0, found by diffing the public members against the 2.8.0 wheel.
  • Docstrings now cover the seven role properties, family_base, family_particles, surnames, given_names, as_dict, Lexicon.default, Lexicon.empty and PolicyPatch.two_given_names, so autodoc renders them.
  • The nameparser.locales.ja cross-reference is fixed.
  • modules.rst's "The 2.0 API" is now "Classes and functions".
  • locales.rst's Spanish and Portuguese titles section has a label.
• locales.rst's promise: it now says what the non-interference gate tests: over the test corpus and each pack's example names, every name a pack reads differently from the default is one its declared surface covers. That surface is a shape for a rule (three name words for the packs that read two family names), or the words a choice changes. Every pack passes the gate, ES, CS, SK and ES_AR included.
• cases.py notes: accepted as swept.

Checks

• uv run pytest passes on 3.11 and 3.15: 37,279 passed, 29 skipped, 5 xfailed. The -n0 -m serial pass also passes (54).
• uv run mypy and uv run ruff check are clean.
• sphinx-build -b doctest gives 0 failures, and -b html gives no warnings.
• The design-docs review ran on every docs/design commit, and a code review ran on the fix commit. Their findings are fixed in the later commits, except one, below.
• There are no open Dependabot PRs, and uv.lock pins namedivider-python 0.4.2, the latest on PyPI.

Per-region lists, declined

• decisions/O6 has a dated 2026-11-05 Sweep stale prose for the 3.0.0 release #1079 entry and a Standing line. They decline a per-region list, or another selection rule, for every Spanish-speaking country but Argentina, each staying on locales.ES.
• The entry cites the figures already recorded:
  • ES_AR's list beats ES's by at most 0.61 point on ORCID (the Add locales.ES_AR, Spanish with a given-name list cut for Argentina #1061 entry);
  • no switch combination beats locales.ES by more than 0.06 point, Peru's;
  • the six-country pooled list came within half a point of per-country lists everywhere but Argentina and Peru (Should two_family_names leave a known given name out of the family, so "José Luis Pérez" keeps family "Pérez"? #959);
  • ORCID PE 96.0% → 95.9% (178 fixed, 189 broken);
  • PSC Peruvian 81.0% → 83.6%.
• The measured population was the groups of at least 100 records; the decline reaches the smaller ones on no figure of their own.
• It names the three earlier passages that pointed at the Open line.

Not in scope here, as planned: the moved-names report, release notes (including the 3.0 entry's "under development" line), CITATION.cff, the version, the tag and the pickle check.

🤖 Generated with Claude Code