[derek73]

D15 of the vocabulary phase 2 spec (#752): the semicolon becomes a default suffix delimiter, and a delimiter glued to a suffix word counts.

Behavior

Input	1.4.0 – 2.5.0	Now
John Smith, MD; PhD	given MD;, family John Smith, suffix PhD	given John, family Smith, suffix MD; PhD
John Smith, MD;PhD	given MD;PhD	suffix MD;PhD
Smith, MD; PhD	given MD;	family Smith, suffix MD; PhD
John Smith, MA; PhD	given MA;	suffix MA; PhD (reported, as MA PhD is)
John Smith, MD ; PhD	title MD, given ; (2.5.0)	suffix MD, PhD (the lone delimiter is dropped, as a declared one is)
Berg, Jan van; MD	middle van;	unchanged: a glued particle joins nothing, backward included (accepted, Derek)
John Smith MD; PhD	family MD;	unchanged: a semicolon before the first comma is rules/T1's question, out of scope

Design

• The default is a constant beside _vocab.delimiter_cores, not a Policy.extra_suffix_delimiters default. That field is a set callers replace (the facade's suffix_delimiter replaces it too), so a default inside it would vanish for anyone declaring their own delimiter. No policy removes the semicolon, as none removes the comma.
• A glued word reads as its stem. The stem is "the word without the delimiter". One predicate, _vocab.glued_stem, answers it, and four sites ask it:
  • classify tags the word as its stem, after the first comma only;
  • _comma.decide uses the stem in its shape tests;
  • is_wholly_suffix uses it in the run test;
  • segment's C2 title test uses it.
• Why the stem is read everywhere. Reading it in the run test alone, as the spike did, put 365 reversed names right. Tagging the stem raised that to 412, and reading it in decide's shape tests (MBBS;) to 431.
• Guards.
  • Behind a head that is not a whole name, the glued word takes the strict test, so Smith, V; MD keeps given V;, as Smith, V MD does.
  • A glued particle, connective or bound given name loses its joining tags, so nothing joins across the delimiter (Berg, de; Smith keeps given de;); P6's backward attachment goes with it, an accepted cost.
  • Every gluable core is tried in sorted order, so a word like MD/PhD; reads the same whatever the hash seed.
  • Only a core with no letter or digit glues: under " and ", Stand is not St with and glued on.

Measured against 638685b

Corpus	Result
ClinicalTrials.gov officials (local, data 2026-10-21)	494 names hold a semicolon. 443 move: 431 reversed names put right, 7 titles now read as titles, 2 standing-alone semicolons now rendered as a comma, 3 others. All read by hand, none worse.
FEC (15 semicolon names), TSE, PSC agreement, committed snapshot, CLDR/GND agreement	Nothing moves. The snapshot only gains the 12 new rules examples.

• Twin sweep: 269,136 generated glued texts across seven comma heads and three orders. No text that read as its space twin on the base reads otherwise now.
• Invariants: test_a_glued_semicolon_reads_as_its_space_twin (990 texts, 726 disagreed at base) and test_a_semicolon_standing_alone_in_a_tail_reads_as_its_comma_twin (129 texts, all disagreed at base).
• Frame cost: reference name 307 → 305; Smith, John 161 → 158.

Docs

• rules/C10 is renamed to "a suffix delimiter parts a suffix run" and rewritten, gaining a Precedence bullet and two Accepted bullets. Its rule-size record grew.
• C1's map line and R1's separator sentence are amended.
• decisions/C10 has a new entry.
• AGENTS.md's spellings paragraph gains the glued-separator lesson.
• Also updated: customize.rst, the Policy docstring and the release log.
• A design-docs review found 8 issues, including two wrong readings (Stand under " and ", and Berg, Jan van; MD), all addressed. A second review of those fixes found a hash-seed-dependent reading (MD/PhD;), fixed, and a forward join across a glued particle, resolved by the decision above.

Decided

• No opt-out switch: the semicolon is built in like the comma (Derek, 2026-10-24; recorded in decisions/C10).
• A glued particle's tags are dropped, so Berg, Jan van; MD keeps 1.4.0's middle van; (Derek, 2026-10-24; rules/C10's Accepted bullet).

Left for release

• migrate.rst entries are left for the 2.6.0 release step, as with the cycle's other PRs.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.06%. Comparing base (638685b) to head (29f3a80).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #871      +/-   ##
==========================================
+ Coverage   99.05%   99.06%   +0.01%     
==========================================
  Files          47       47              
  Lines        4669     4730      +61     
==========================================
+ Hits         4625     4686      +61     
  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.