Split a Thai title glued to the start of the name off it (rules/W6) (#317) - #1324

Closes #317.

What changes

A listed title glued to the start of the name splits off and reads as the title (new rule W6, the glued-title peel). It mirrors W2's peel of honorifics glued to the end of a name.

How the peel decides:

• Which words peel. A word peels only where the vocabulary marks it attached = { side = "before" }. That is the slot the build had reserved, rejecting it with "phase 1 (the Glued CJK honorifics are absorbed into the name (山田太郎様 → family 山田太郎様) #308 peel)". It compiles into a new private GLUED_TITLES set.
• One peel, longest match first, falling back to a shorter title where the longest would leave a Thai vowel or a mark (นางสาวิตรี → นาง + สาวิตรี). A hyphen, a digit or the repetition mark after a title blocks the peel (นางสาว-สุดา stays whole). It never cuts a title that is listed whole, and never a title that may be a given name, a caller's included.
• What's left must open the way a Thai word can. That keeps Japanese names written in Thai letters whole (นางิ Nagi, นางาเซะ Nagase).
• Where it looks. The first word of the name, and the first word of the given part after a family comma, stepping past titles written apart (including dotted pairs like รศ.ดร., though never a word holding a name, such as ดร.สมชาย). Under a family-first order, or before a family comma, that first word is the surname. Surnames were vetted too: 1 of 12,073 Thai surnames opens with นาย, none with the others.
• A remainder still glued at a full stop (อาจารย์ดร.สมชาย) divides as T8 divides it written apart.

Vocabulary. th.toml lists the salutations นาย, นาง, นางสาว and the feminine rank ว่าที่ร้อยตรีหญิง, and marks them and #1166's spelled-out titles to peel. The short forms (นพ, ดร, ศ, พญ, ทพ) and คุณ are not marked, because each opens real given names.

How the words were vetted

The source lists aren't committed. They're cited in decisions/W6.

Measured

Against master ed7dfb8:

• orcid-researchers: 7 names move, and each one gains its title and keeps the given name. All 7 count as field-agreement "regressions" only because the filer typed the salutation into the given-name box.
• jrc-names: reads exactly as master. No other local corpus holds Thai names, and only a word opening with a glued title can move.
• Committed snapshot: no corpus name moves. It gains W6's examples.
• Frames: a name with no word opening with a glued title pays none. I caught a generator in the gate costing 2–3 per non-ASCII name and replaced it with a loop. The reference row is unchanged at 344.

Review rounds

The code review and the docs review found these, and each is fixed with a test row that fails on the first commit:

• A given name lost to the title. The remainder glued at a full stop wasn't divided, so อาจารย์ดร.สมชาย lost its given name to the title.
• No fallback to a shorter title. Without it, นางสาวิตรี stayed whole.
• A dotted title pair wasn't stepped past.
• A caller's word could be split with no title read. A title that is also a given name, once marked attached, cut the word and read no title.
• Lexicon.add accepted a multi-word attached title, which the build refuses.

A third review, of the fix commits, found two places where those fixes over-reached: the site walk skipped a title glued to a name (นางรศ.ดร.สมชาย kept the whole word as the title), and the fallback fired behind a hyphen (นางสาว-สุดา → นาง + สาว-สุดา). Both are fixed, with rows that fail on the previous commit. The Codecov failures were three uncovered lines: two build refusals, which now have mutation rows, and a dead check, which is removed.

The W6 and decisions text was tightened to match the code: Then says "marked to peel", and the Except is about a title listed whole.

Accepted costs:

• After a family comma, a glued title with a full-stop join leaves the given part empty, as T8 already does for ใจดี, ดร.สมชาย (Should a glued word in the given part after a family comma divide? (ARTHUR, L.MICHAEL gives given 'L.MICHAEL') #1311).
• Only one glued title peels per word.
• A name a caller lists whole is still cut, since the guard reads titles only.

Open, recorded in decisions/W6

• A Thai title with one name word. นายสมชาย reads family สมชาย, the way Mr. Smith does under the title-plus-one-word rule (H1). Thai addresses a person by the given name. addresses = "given" on the Thai titles would settle it, but that covers Should Thai academic and professional titles read as titles? (รองศาสตราจารย์ ดร.สุดา ใจดี gives given 'รองศาสตราจารย์') #1166's titles too, so it's left out of this PR.
• The dotted ranks and salutations (น.ส., ด.ช.) are Should Thai military and police ranks read as titles? (พ.ต.อ.ดร.สมชาย ใจดี gives title 'พ.ต.อ.ดร.สมชาย' and no given name) #1319's.

Tests

• tests/v2/cases.py: eight rows covering the four comma shapes (no comma, a name before the comma, one word before it, nothing before it), the listed-whole word, and the short-form and Japanese boundaries.
• tests/v2/test_lexicon.py: the off switch (attached=None), a caller's attached="before", and a check that every shipped glued title actually peels.
• tests/v2/test_vocab.py: build mutations for the wrong side and for a multi-word title, and the README's property table probed for both categories.
• Rules and decisions: rules/W6, decisions/W6, both indexes, the vocab README, the Lexicon.add docstring, AGENTS.md's set list, and the release log.

🤖 Generated with Claude Code