[derek73]

After #745 gave every rule a Signal: line, eight rules had a line naming two signals, which marks them as split candidates under docs/design/AGENTS.md's "One rule per piece of evidence". Derek decided each one as recommended. After this PR, T4 is the only split candidate left (grep -l '^ Signal:.* + ' docs/design/rules/*.md).

Three new rules

• P8 (a single-letter connective against the initial it spells) is split from P3 (connectives). P8 takes the case contrast: the initial reading, its report and "connects nothing", the marked set, the views, the Should Jose E Maria Santos join like Jose e Maria Santos does? (the single-letter connective's Latin-capital veto was never adjudicated) #383 parity breaks, the declined-marker clause, and the case half of the own-words span. P3 keeps:

  • the join and its three-word count
  • the one-name-word count
  • the generational connective
  • the separator
  • the own-words span

  decisions/P8 takes the Should a bare 'И' parse as an initial or the conjunction 'and'? (e.g. 'Хосе И Мария Сантос') #267, Should Jose E Maria Santos join like Jose e Maria Santos does? (the single-letter connective's Latin-capital veto was never adjudicated) #383/JOSE E MARIA SANTOS reads the conjunction E as an initial; an all-caps name carries no case signal to read it from #479 and Should a conjunction inside a hyphenated surname keep its lowercase? Ortega-y-Gasset now capitalizes to Ortega-Y-Gasset #478 entries.

• P9 (a particle that is also a suffix word, beside a family comma) takes P2's (particles join forward) clause on the part before a family comma and P6's (the trailing orphan particle) third exception. Why does SMITH VD MA, JOHN read family SMITH MA with VD as a suffix? #573 decided both. P6 now cites P9 for that exception, the way it cites S6 and S7 for the other two. The Why does SMITH VD MA, JOHN read family SMITH MA with VD as a suffix? #573 entry moves from decisions/P6 to decisions/P9.

• C13 (a credential opening the part after the comma) takes Should a comma part opening with a credential be a suffix comma? John Smith, PhD Jones reads family John Smith #603's half of C8 (the listing form's part after the comma). C1's hub map names C13. The Should a comma part opening with a credential be a suffix comma? John Smith, PhD Jones reads family John Smith #603 entry moves from decisions/C8 to decisions/C13.

One clause moved, no new ID

• S2's "an ambiguous acronym written with its periods" clause, with its Rationale sentence and the M.A., M.Eng. and Jack Ma. examples, moves to S3 (credentials written with interior periods). The recommendation said S4. It went to S3 because S3's signal is the periods shape. S4's signal is the words-to-spare count, so putting the clause there would have made S4 the next split candidate. docs/design/AGENTS.md now records this check.

Kept, re-cited

• S10 (the trailing roman numeral) cites T4 instead of restating T4's case conditions.
• C9 (a one-character suffix word inside a comma part) cites C2 for the further comma.
• T4 (the bare letter as an initial) is unchanged and remains a split candidate. An intermediate commit reworded its line as one signal; it is restored, because the case read compares a letter with the rest of the name, which is a case contrast and not a shape. Whether to split off the period half is left for a later decision.

Other changes

• Every reference that meant a moved piece is retargeted: code comments (quotes still checked by test_doc_citations.py), tests, cases.py notes, other rules, decisions, mechanisms.md, AGENTS.md and docs/modules.rst. Each moved decisions entry carries a split note.
• tests/v2/rules_budget.json is re-recorded. P3 shrinks from 1026 to 628 words. P8, P9 and C13 are new records. S3 grows by the moved clause.
• The split-candidate test no longer asserts that the list is non-empty. It also checks the recipe against a synthetic rule of each kind, so it keeps testing the recipe once T4 is resolved.

Behavior: no change. Every example line is the same set, and corpus_rules.jsonl is unchanged.

Review: the design-docs-reviewer ran on the branch and again on its fix commit. Every finding was fixed except one, left for Derek: S10 still applies its own words-to-spare count for numerals. It cites S4 for it, per rules/README.md's "S10's count is S4's". The reviewer's other open point, that T4's one-signal line mislabeled its case read, is resolved by restoring T4's two-signal line.

Full suite: 21855 passed. mypy and ruff are clean.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.05%. Comparing base (d373055) to head (6d99ba5).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #755   +/-   ##
=======================================
  Coverage   99.05%   99.05%           
=======================================
  Files          47       47           
  Lines        4642     4642           
=======================================
  Hits         4598     4598           
  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.