[derek73]

Part of #752: one shared predicate for each of three questions that #752's later changes widen. Neutral: no reading moves.

What changes

All three predicates are in nameparser/_pipeline/_vocab.py.

• split_pair(first, second) handles the split credential Ph. D. (decisions/phd-merge). Group's merge, _comma._pieces and is_wholly_suffix each had their own copy of the two-pattern test, and now all three call this predicate. The test runs on every adjacent pair, so each caller first asks SPLIT_PAIR_OPENS at C level. That is the predicate's own first-word pattern, so the pre-test can't drift from it, and a name without Ph. pays no frame. A later change for Should John Smith Ph. D. Jones read like John Smith PhD Jones, or is the split spelling's reading the better one? #827 widens both together, to merge split credentials such as Pharm. D..
• reads_as_generation(text, preceding, one_case) is is_trailing_numeral_suffix renamed, as the predicate every numeral rule is to ask. Today it is the fork's shape half alone. C12 keeps its suffix-word test beside the call, as on master, because its question is broader: any suffix word (Jr., Esq.) or a numeral that reads as a generation. The listed numerals belong inside the predicate, or each rule copies the vocabulary test again, which is how Should a known suffix right before a comma make it a suffix comma? Fabio Martini II, AD reads given AD, family Fabio Martini #656's first draft missed VI. The shape test already takes ii–iv, so reading the vocabulary would add only the one-letter i and v, and that moves readings: 375 at the peel's fork (John J. V from family 'V' to suffix 'V') and 144 new reports (John J. V MD), because the ones those sites meet carry the initial tag. That half waits for the change that has every numeral rule ask the predicate (Should John Smith VI MD read like John Smith III MD? #784, Should John Smith V, M.J. keep V as the suffix, as John Smith Jr., M.J. keeps Jr.? #774, Should Smith, John PhD I, Jr. keep I as a middle initial, as Smith, John V, Jr. does? #819). The decisions entry lists the sites that still recognize a numeral on their own, with a reading for each.
• is_cased_paired_initials(text) is S3's "paired initials in a script with capitals". Classify used to test the script half inline (Read a surname written as two dotted initials as a name, not a credential (#864) #872). The lift puts that condition in one place for any later caller. C5 keeps the bare is_paired_initials: moving it would change John Smith, พ.บ. (given พ.บ.).
• tests/v2/test_regex_sync.py's roster still listed _group._PH and _group._D, though both had already moved to _vocab. Nothing checked an entry with no config key, so nothing failed. A new test now refuses an entry that names no compiled pattern, and I confirmed it fails on master's roster.

Neutrality

Measured against this branch on 632685c, and again on 2dd0414 after phase 2a (#883) merged:

• Grid. 20,364 texts under all three orders read identically, compared on fields, ambiguity kinds and details, token roles and tags, initials() and capitalized(). The grid covers:
  • 22 heads: no comma, a whole name before a comma, one word before it, nothing before it, a nickname-only head, a maiden clause, a family comma with a title-particle head, and a quoted clause before the pair.
  • 33 spellings, among them MA, M.D., X.Y.Z., XYZ, Ph. D. and its case and spacing variants, V, VI, III, XI, M.J., K.C. and พ.บ..
  • Each text also written in lower and in upper case.
• The harness sees differences. With each piece knocked out by monkeypatch, the same harness reported 45,871 (split_pair), 18 (C12's call to the numeral predicate), 3,974 (the cased half) and 2,986 (the peel's call to it) readings differing.
• Direct comparison. The old and new predicates gave the same answer on 19,769 direct calls.
• Corpora:
  • tests/v2/test_snapshot.py doesn't move.
  • Every cldr_agreement.py record matches: CLDR, GND, Wikidata, conventions, native, and both locale views.
  • local_agreement.py psc-individuals matches its record. The record was already stale against 632685c (13 records moved before this change), so I re-recorded it from each base first.
  • The local snapshots of ctgov-officials, fec-individuals and tse-donors, each recorded from the base with --tree, all match on both bases.
• Frames (sys.setprofile):
  • Unchanged: _CALL_BASELINE's reference name (305) and Smith, John (158).
  • +1 per site that tests the split pair, for each Ph word with a word after it: John Smith Ph. D. goes 233 → 234, Smith, John Ph. D. 239 → 241.
  • Unchanged on C12's evidence: John Smith Jr., AD stays at 372, since the tag test runs ahead of the call.
  • The split test's frame is paid whether or not a D. follows (ph ph ph ph ph ph 282 → 286), so it grows only with such words. It's a rare shape that no guard counts. uv run pytest -n0 -m serial passed three runs in a row. It failed once under a load average of about 10 from a parallel session, a different clock-timed row each time.

Docs

• mechanisms.md#ONE-PREDICATE-PER-QUESTION names the three predicates.
• decisions/vocabulary-model has a dated entry for the change and its measurement. It points at items 6 and 8 of the agreed-design entry from Record vocabulary phase 2's agreed design in the RFC and decisions (#752) #881 for what the later changes widen. It also leaves one overlap for the split-credential PR to state: that entry plans for the split merge to ask is_cased_paired_initials, but its own item 6 refuses a split credential that opens with a single letter, so the length test answers first for A. B..
• Six decisions entries and AGENTS.md's spelling sweep now point at the new name of the numeral predicate. decisions/vocabulary-exclusions and a cases.py note no longer name the retired PH/D patterns.
• No stage's Reads: line changes, because no stage reads a new field.

Reviews

• Code review. Its own grid of 71,382 texts gave 1,070,730 parses across the default policy, lenient_comma_suffixes=False, unlisted_dotted_suffixes=False, CapsSuffixes.EVERYWHERE/OFF, a delimiter policy and all three orders. It found 0 differences and no missed copy.
• Design-docs review, then a second review of its fix commits. Between them they found the catalog overstating the numeral predicate, the sites it misses, and stale PH/D names. Each is fixed in its own commit.
• PR review (comments, tests, simplification). It tightened comments. It also found that no test held a split credential without its periods: making either period required passed the whole suite. This PR adds four parity rows (Ph D, Ph. D, Smith, John Ph D, John Smith, PH D, each checked against the 1.4.0 wheel) and unit tests for SPLIT_PAIR_OPENS and is_cased_paired_initials. It moved C12's suffix-word test back beside the call, as above.

Checks

• uv run pytest
• uv run pytest -n0 -m serial
• uv run --python 3.15 pytest -q
• uv run mypy
• uv run ruff check
• sphinx-build -b doctest docs
• python -m doctest README.rst

Part of #752; doesn't close it.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.08%. Comparing base (2dd0414) to head (db694cb).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #884   +/-   ##
=======================================
  Coverage   99.07%   99.08%           
=======================================
  Files          47       47           
  Lines        4781     4785    +4     
=======================================
+ Hits         4737     4741    +4     
  Misses         44       44           

☔ 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.