[derek73]

Closes #1088.

A credential joined by a slash or a hyphen now reads, inside brackets or quotes, as its words written apart with a comma do. Before this change, John Smith (MD, PhD) read suffix MD, PhD, but John Smith (MD/PhD) read nickname MD/PhD, even though rules/S3 reads the bare John Smith MD/PhD as the suffix.

Input	master	this PR	Its words written apart
John Smith (MD/PhD)	nickname MD/PhD	suffix MD/PhD	suffix MD, PhD
Smith (MD/PhD), John	nickname	suffix	suffix
John Smith (Jr/Sr)	nickname	suffix	suffix
John (MD-PhD) Smith	nickname	nickname	nickname (S1's initials Except)
Justin (JP/MD) Stearns	nickname	nickname	nickname
John (NP-C) Smith	suffix	suffix	(listed whole: one word)
John Smith (MD/Bob)	nickname	nickname	(S3 refuses it bare)
Jean (M.A,) Smith	nickname, no report	nickname, reports suffix-or-nickname	as Jean (M.A) Smith

Rule changes

• rules/S1 (a bracketed credential is not a nickname):
  • Its set-apart word now asks suffix_as_written, the question a bare word is asked, so a joined credential S3 reads is a word of such a clause.
  • The initials Except asks each chunk of a joined word, so before a name word the two spellings agree. This was decided after review, as option A (parity).
  • A credential the vocabulary lists whole (NP-C, SHRM-CP) is one word and is never split.
  • A new Accepted bullet records where the two spellings still differ, each by a rule of its own: a member of the ambiguous class whose position S13 does not settle ((M.A./PhD) suffix vs (M.A., PhD) nickname), and a word that is a title too ((PhD/Sr) John Smith suffix vs (PhD, Sr) John Smith title Sr).
• rules/S13 (a bracketed member settled by its writing):
  • A member left the nickname is reported, and a comma the clause keeps on it is no part of it.
  • The member's dotted test is now one predicate, _vocab.dotted_member. Before, post_rules held two hand copies.
• decisions/S1 has a Standing line and a dated 2026-11-05 entry holding the decision, the options put, the moves and the measurements. decisions/S13 has a Log line.

What it moves

All measured 2026-11-05 against master 4109b02, each tree's nameparser.__file__ asserted.

• Snapshot: none of its 6,945 existing entries (6,736 distinct names) moves. None of its 181 names holding a delimiter pair has a joined word inside a clause, so the snapshot cannot see this shape. It gains the rule's 14 new examples as new entries.
• Grid: 30 clause words × 14 shapes × 3 name orders, 1,260 names. 399 move, 133 under each order. The exact templates are in the decisions entry. Every move is one of three kinds:
  • a joined credential now read as the suffix;
  • a period-final joined word before a name word now read as the nickname (John (M.D./Ph.D.) Smith read middle in 1.4.0 through 2.8);
  • the member report.
• Local corpora (snapshot.py --local <source> --base <4109b026>):
  • ctgov-officials, 3 moves: Franck Pages, Professor (MD-PHD) and two like it, now suffix where they read nickname.
  • fec-individuals, 1 move: FRANTZ (CAPT USN-RET), JOSEPH now reads title CAPT, suffix USN-RET.
  • psc-individuals and orcid-researchers: none.
• Lost report: John Smith (MD/PhD, MBA) now reads suffix with no report. Master reported suffix-or-nickname there, because mba was marked ambiguous on 2026-11-03; 2.8.2 did not.

Frames

The reference row is unchanged at 306.80 calls/parse on py3.11, so _CALL_BASELINE does not move and #1089 can lower it from the same figure. On the clause shapes, master → this PR:

Name	master	this PR
John Smith (Andy)	214	214
Smith, John (MA)	242	234
John Smith (PhD, MBA)	271	268
Jean (M.A) Smith	242	246
John Smith (MD/PhD)	221	244 (now reads a suffix)

Tests

• Case rows in tests/v2/cases.py:
  • every comma shape: after the name, inside it, before it, alone, before and after a family comma, before and after a suffix comma, an empty head, and in quotes;
  • a family-first order;
  • the period escape extract keeps for a dotted member;
  • the kept-comma members;
  • the listed-whole credentials;
  • the refusals.
• Guard tests in tests/v2/pipeline/test_post_rules.py, each with a recorded negative control:
  • A bracketed word after the name is the suffix wherever the bare word is one. Control: 7 words read differently on master. S3's refusals are pinned apart from the predicate the test asks.
  • A joined credential reads as its words written apart, over 2,310 (pair, shape, order) rows. Control: 540 rows read apart with the chunk loop off.
  • Two caller-lexicon tests for the title-and-member paths.
• Recorded counts updated: the field-sweep claim count (1615 → 1624 on the merged tree: Should Smith y V, John keep the V in the family name, as Smith Jones y V, John does? #1093 and this PR each added nine claims), test_layering's _UNGUARDED_READS (128 → 123 → 128 across the commits, net unchanged) and rules_budget.json (S1 and S13 grow; each new clause reads a signal its rule already reads, S3's joined chunks or N1's comma).

Merged with master at 510d1aa (#1093, #1094, #1096, #1099). On the merged tree, corpus_rules.jsonl rebuilds byte-identical, the snapshot matches, rules_budget.json re-records unchanged, and both negative controls re-take at their recorded values (540 of 2,310 rows; 7 words).

Checks run: uv run pytest, uv run pytest -n0 -m serial, uv run mypy, uv run ruff check, and py3.15 (uv run --python 3.15 pytest -q, in an environment of its own). All pass.

Reviews

Each ran as a subagent on the diff; every finding was fixed or is listed here.

Review	Outcome
Code review (pr-review-toolkit:code-reviewer)	Found that a joined word diverged from its split spelling before a name word. The maintainer chose parity (A), which is implemented.
Test coverage (pr-test-analyzer)	Found the rules corpus stale, extract's member check unguarded by any test, and gaps in comma shapes, orders and kept-comma members. All added; extract's check is now caught by two rows.
Silent failures (silent-failure-hunter)	Found that a caller lexicon listing m.a as a title lost the S13 report. Fixed with dotted_member, plus a test.
Type design (type-design-analyzer)	The "member" kind is the right shape, and mypy enforces that it never reaches set_apart. Took its optional SetApartKind alias.
Comments (comment-analyzer)	Every history claim verified. Fixed the parity wording, one comment about the escape, and a duplicated comment.
Design docs (design-docs-reviewer), round 1	Found the parity gap, a snapshot denominator copied from another entry, an unreproducible population count, and no issue link. Re-counted with a recipe; the issue is drafted.
/simplify (reuse, simplification, efficiency, altitude)	Applied inline necessary-half pre-checks, a gate on the initials walk, a flattened chunk loop and the member fold. Skipped two suggestions that would bring back an inline copy of the dotted test.
Second round on the fix commits: code review	Found the parity fix splitting credentials listed whole (NP-C, SHRM-CP), the "widened check narrows" shape. Fixed; a sweep of all 70 listed words with a joiner (560 names) matches master.
Second round on the fix commits: design docs	Found an unrecorded member-chunk difference, the title-too site stated too narrowly, a wrong period-final claim, a wrong 2.8 comparison, and no grid recipe. All corrected.
Third round on ed403c2: code review	Clean: the gate means exactly "listed whole"; 4,023 parses, of which 108 move, all CP-C, MT-BC, NP-C and SHRM-CP before a name word, back to master's suffix; every rule example and decisions claim checked against master, 2.8.2 and 1.4.0. It noted, out of scope: a joined word holding a listed hyphenated credential (PhD/NP-C) is refused by S3 bare, as on master.

Known gaps, left out

• An interior-period credential in brackets (John Smith (Msc.Ed)) still reads as a nickname, though bare it reads suffix. It goes through a different predicate (S3's any-chunk period test). An issue is drafted; the decisions entry's pointer gets its number once it is filed.
• A dotted member as a chunk of a joined word reports nothing: John (M.A/PhD) Smith reports no suffix-or-nickname, where John (M.A, PhD) Smith does. This predates the change.

🤖 Generated with Claude Code