[derek73]

Under a family-first name order, rules/P1's fold takes two more pieces opening the name.

A lowercase ambiguous particle in a mixed-case name (#670, new rules/P11). The order has already settled that the opening slot is the family. The writing settles whether the particle is a one-word surname or the start of one.

name_order=FAMILY_FIRST	before (2.0.0–2.8.0)	now
van den Berg Cornelia	family van, given den Berg Cornelia, reported	family van den Berg, given Cornelia, no report
Le Van Thanh	family Le	unchanged (capitalized)
VAN GOGH VINCENT	family VAN	unchanged (one case)
le Van Thanh	family le	family le Van Thanh (accepted, rules/P11)

The default order is unchanged: van Gogh still reads given van. Folding a capitalized run of two or more particles (#670's third question) is declined, because Le Van Thanh is that shape. Both family-first orders are covered.

A particle chain behind a title-particle, cut as the run is (#828, rules/P1 and P4's Except). St de Mesnil Jean family-first gives family de Mesnil, given Jean, as Dr. de Mesnil Jean reads. Before, it gave family de Mesnil Jean.

The chain is left whole in three cases:

• the fold would take it whole anyway: H1's given name behind Sir (St Sir de Mesnil → given), or P5's bound given name;
• it holds a word that group's trailing read weighed as a suffix (St de Mesnil Jean V keeps its 2.8 reading);
• Under a family-first order, should St de Mesnil Jean read given 'Jean', as Dr. de Mesnil Jean does? #828's second question, multiple surnames under a family-first order, is answered by rules/O6 (two_family_names), which composes with the cut.

Measured

On 2026-11-04, against origin/master c620754. Recipes are in decisions/P1 and decisions/P11.

• Under a family-first order, should a lowercase von/van den opening the name start the family? van den Berg Cornelia gives family van #670:
  • Corpus names opening with a lowercase ambiguous particle, read under FAMILY_FIRST: 153 of 280 move, all read as improvements. Of the 127 that do not, 73 are written in one case and 54 carry a family comma.
  • CLDR agreement: 240 field lines now agree and 51 changed while still disagreeing. None regressed.
• Under a family-first order, should St de Mesnil Jean read given 'Jean', as Dr. de Mesnil Jean does? #828: names holding St/Freiherr and two more words, read under both orders: none of the 1,053 moves. No real name opens with a title-particle and a particle.
• Under a family-first order, should St de Mesnil Jean read given 'Jean', as Dr. de Mesnil Jean does? #828 with Should Otto Graf Lambsdorff read family Graf Lambsdorff, as Manfred Freiherr von Richthofen reads family Freiherr von Richthofen? #709's ranks (against master e599289): the 44 corpus names holding Graf, Gräfin, Fürst or Prinz before two more words read alike in both trees, each a rank mid-name (Alfred Graf VON Matuschka). The cut reaches a rank only in a name opening with one: Graf von Berg Otto under a family-first order reads family von Berg, given Otto, where master reads family von Berg Otto.
• Snapshot: 37 family-first entries move, 35 of them names only corpus_cldr.jsonl holds. The rest of the snapshot diff is new rules examples.

Also in the diff

• When a report's other reading would join a chain the fold cuts, assemble now names the field that reading lands in (Freiherr von J. ma).
• _state gains reported_suffix_fork and the private PAST_REACH_TAG, so post_rules' decline and assemble's alternative ask one predicate over the same words.
• config_diff.py's expr: scope holds the three name orders. The recipes in decisions/P1 and decisions/P11 name FAMILY_FIRST and raised NameError as written; a test now holds the scope.
• Rule examples that used lowercase van under family-first (O6, R11, P6) now use Van, which keeps what each example tests.
• Docs: root AGENTS.md's particles paragraph, customize.rst, and the vocabulary-model 4h entry's head note.

Reviewed by a code review and a design-docs review, with each fix commit reviewed again (the first fix keyed the decline on suffix vocabulary and over-fired; it now reads the report), then a comment-accuracy review and a mutation-based test-coverage review. 3.11 and 3.15 suites and the serial benchmarks pass.

Synced with #1064's ranks (master e599289): three rows in test_family_first_fold_reads_the_writing_and_the_chain show the cut reaching Graf, Gräfin and Prinz, each failing on master. snapshot.jsonl, corpus_rules.jsonl and rules_budget.json are re-recorded; the snapshot gains no move from the ranks.

Closes #670
Closes #828

🤖 Generated with Claude Code