[derek73]

Sub-phase 4h of #754: the shape inventory per order (#469). This lists the written arrangements each name_order is read in, each with what the tree reads today. No reading moves. It also records #938 as declined, decided in chat on a survey of real family-first names.

What changes

• tools/differential/shapes.py gains ARRANGEMENTS. It has 104 rows: 36 under the default order, 38 under FAMILY_FIRST and 30 under FAMILY_FIRST_GIVEN_LAST. Each row has:
  • its notation and a name written in it;
  • the positions it instantiates;
  • every role that is not empty, and each ambiguity kind;
  • where a person would not read the name that way, a mark: accepted (the document that decided the reading) or misread (an open issue).
• POSITIONS covers these, for every order:
  • a title at the front, between the family and the given name (family-first orders only), at the end, and after a full head's comma;
  • a suffix without a comma, before one and after one;
  • the comma heads (full, one-word, empty);
  • a middle name, a particle, a nickname;
  • nameparser/AGENTS.md's sweep items: a nickname-only head, a connective and a particle surname before a comma, two title-particles at the head, a dual in a title run, a trailing title after an empty head, a roman numeral, dotted and undotted suffixes, a period-shape title, and an NFD letter.
• Shape model. Each order still has one Shape, and each arrangement names the shape it is written under, inheriting its order and min_baseline. That leaves the corpus tag, build_conventions_corpus._shape_for, cases._SHAPE_IDS and compare.py's skip untouched. Under a family-first order, family-comma rows are tagged 4 or 5 for their order, and the shapes.py docstring says so.
• tests/v2/test_shape_inventory.py checks:
  • each row's roles and ambiguity kinds, read under its order;
  • that each order covers every position it has a slot for;
  • each position the text can show against the text (a comma position needs a comma; nfd needs a text that NFC changes);
  • that every mark cites something that exists.
• Marked rows are pinned to their current reading, not marked xfail. A non-strict xfail would let a change to the row pass unnoticed.
• Negative control: under the default order, 49 of the 68 family-first rows fail. Of the 19 that pass, 15 hold a family comma and 4 lay out no name word by order.
• Corpus and snapshot. corpus_shapes.jsonl grows from 300 to 395 lines. The snapshot gains 80 entries and changes none.

#938 declined

The decision is recorded in decisions/H5's (the trailing title run) 2026-10-30 #938 entry, with the survey figures and their dates and samples. Real family-first names write the title in front:

• Hungary, 2024 municipal candidates: 37,250 names, 0 with a title between the family and the given name.
• Japan, Kantei English pages: 16 of 16 in front.
• Vietnam, Assembly candidate cards: 52 of 52 in front.
• JRC-Names: 11,558 romanized CJK spellings, no real titles.

CLDR's 50 of 70 is a template artefact.

The eight rows in #938's shapes are now accepted = "H5", as is John Smith Prof:

Order	Input	Reads
FAMILY_FIRST	Wang Dr. Xiaoming · Wang Dr Xiaoming · Wang Insp. Xiaoming	given the title, middle Xiaoming
FAMILY_FIRST	Brühl Prof. Dr. Ada Cornelia	given Prof., middle Dr. Ada Cornelia
FAMILY_FIRST	Wang Xiaoming Senator	middle Senator
FAMILY_FIRST_GIVEN_LAST	Nguyen Dr. Thi Mai · Nguyen Insp. Thi Mai	middle Dr. Thi · Insp. Thi
FAMILY_FIRST_GIVEN_LAST	Nguyen Thi Mai Senator	given Senator, middle Thi Mai

The usage.rst paragraph after forms 4 and 5 now states the title-in-front convention as the supported reading.

Open, in the 4h entry:

• The Hungarian title after a married -né word (Bakainé dr. Korbács Melinda): 102 of 37,250 names, all misread today. It is a question for a future hu pack.
• Unweighed, so unmarked: Vega y Lopez Jr., PhD Jones under the family-first orders (given y or Lopez), and an ambiguous particle opening a family-first name (van Gogh Vincent reads family van).

Checks

• uv run pytest: 33,883 passed.
• -n0 -m serial: 42 passed.
• mypy and ruff are clean.
• sphinx-build -b doctest is clean.
• The design-docs reviewer's first-round findings were fixed in the second commit. The independent review's points 1–8 are fixed in the third.
• The second review is fixed in the fourth commit:
  • every position is now held to the row's text by a predicate, and both reviewer mutations fail;
  • marks are decisions keys only: van Gogh Vincent, du Ménil Thi Mai and the family-first Vega rows are unmarked and listed as unweighed, and the default Vega row is accepted="C1";
  • H5's Declined entry carries its own Recompute recipe;
  • the RFC states the decline.

🤖 Generated with Claude Code