fix(C1): a particle surname before a comma is one name word (#575)

[derek73]

Closes #575.

C1 decides whether a comma introduces the listing form (Family, Given) or trailing credentials (Name, PhD). Part of that decision is counting the words before the comma: two or more name words mean the name is already complete. All three places that count, the ones listed under Changes, counted tokens or pieces. So a particle surname counted as several words:

text	2.3.0	master	this PR
De La Cruz, Ed	given Ed, family De La Cruz	family De La Cruz, suffix Ed, no given	given Ed, family De La Cruz
van der Berg, MA	given MA, family van der Berg	given van, family der Berg, suffix MA	family van der Berg, suffix MA
van der Berg, PhD	given van, family der Berg	same	family van der Berg, suffix PhD
Van Buren, Ed	given Ed, family Van Buren	given Van, family Buren, suffix Ed	given Ed, family Van Buren

The ambiguous-credential rows (Ed, MA) broke earlier in this 2.4 cycle and that breakage never shipped. The PhD row (an unambiguous credential) was wrong in 1.4.0 through 2.3.0.

The rule

A particle run and the one name word it attaches to count as one word. That's the reach P1's fold uses, not P2's whole chain, so de Mesnil Jean, Dr. under FAMILY_FIRST is still two words: family de Mesnil, given Jean.

• The comma settles P1's fork for a leading particle that could be a first name (Derek's call). Standing alone, Van Buren reads given Van, but the listing form puts a surname before the comma. Van Johnson, Dr. moves the same way.
• Particle chains only. Connective joins aren't counted as one word, because whether P3 joins depends on the whole name, which the count is helping to decide. Bound given-name pairs aren't either, because they build a given name. P3's statement and mechanisms.md#UNIT-PARTITION's Contract now name this count as their exception.
• A title-particle (Freiherr, St) is no particle only where it opens the part. Freiherr von Berg, PhD keeps its title; de St Pierre, Ed stays one surname. Abu counts as a particle, so Abu Bakar, Ed keeps 2.0–2.3's reading.
• De La Cruz, M.J. K.L. now reads given M.J., suffix K.L. (Derek's call), the same as Cruz, M.J. K.L..
• Freiherr von Berg, Ed reads family Freiherr von Berg, which is the right reading: since 1919 a former German noble title is part of the legal surname. That's also 2.0–2.3's reading. The same listing-form rule is a cost for Prof. Cruz, Ed, recorded as accepted.
• Out of scope: a title in front. Dr. van der Berg, PhD reads as its part does on its own.

Changes

• _vocab.unit_ends: the unit walk moved here from post_rules, with a chain flag. Post_rules' fold uses the full chain; the C1 counts use the fold's one-word reach.
• The C1 counts:
  • segment's legacy token count (surname_unit_count)
  • the ambiguous-class count (name_word_count, which has a no-particle fast path)
  • assign's positional-read test, which now walks every token so a suffix word still stops a particle
• Facts: segment builds them from the vocabulary (surname_unit_tags) and assign derives them from classify's tags (surname_unit_facts). An agreement test sweeps every single-word vocabulary entry in three casings and both positions, with a stored negative control. Case rows pin each count's position flag; mutating either flag to a constant now fails.
• Docs and ledgers: rules.md#C1 (plus P3's exception), decisions.md#C1, mechanisms.md#UNIT-PARTITION, the release log (including a stale Dotted initials after a two-word surname read as a credential: García Márquez, G.J. has no given name #563 example), and ledger entries at all five baselines.

Verification

• Tests: full suite 10732 passed; mypy and ruff clean.
• Gate: exits 0 at 1.4.0, 2.0.0, 2.1.0, 2.2.0 and 2.3.0. At 2.3.0, fix(#575) classifies one name.
• Blast radius: of the 1459 names in master's corpora, under all three name orders, exactly one moves: De La Cruz, M.J. K.L., the decided move. The corpus holds only two names of this shape. Outside the corpora, the class that moves against 2.3.0 is a leading ambiguous particle plus one word before an unambiguous credential or a title alone (Abu Bakar, PhD, bin Laden, PhD, Van Johnson, Dr.).
• Cost: John Smith, PhD goes from 209 to 217 frames per parse. Smith, John, John Smith, MA and the benchmark reference name are unchanged.
• Review: four rounds — design docs, code, then two reviews of fix commits. Each fix round found defects in the one before: title-particles folded into the family, a suffix not stopping a particle, the title-particle exclusion applied regardless of position, Abu described backwards, an unguarded flag, false blast-radius claims. All are fixed and pinned. Details are in decisions.md#C1.

Unblocks #564.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.89%. Comparing base (65afcdb) to head (733cc17).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #576      +/-   ##
==========================================
+ Coverage   98.87%   98.89%   +0.01%     
==========================================
  Files          45       45              
  Lines        4017     4070      +53     
==========================================
+ Hits         3972     4025      +53     
  Misses         45       45              

☔ 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.