[derek73]

Part of #761.

Migrates P2–P9 in docs/design/rules/ from the block format to the skeleton format (When / Then / Because, then Except / Precedence / Accepted bullets, --- footer), one commit per rule. No rule changes meaning.

• Examples: every example line is kept verbatim; the per-rule multisets of example lines match master exactly (P2 24, P3 16, P4 6, P5 41, P6 39, P7 6, P8 12, P9 6). P3's two NFD-typed lines (John й Smith, É. y É. Smith) are copied byte for byte. corpus_rules.jsonl and the snapshot do not move.

• Code quotes: every code-comment quote of P2–P9 still matches (the phrases code quotes were kept verbatim), so no code comment needed retargeting.

• Size budget: tests/v2/rules_budget.json is re-recorded, one rule's line per commit, each line moving from the block row to the skeleton row.

• Decisions: one dated bullet appended to decisions/P6 (2026-10-20). P6's block text carried three passages of argument or history. The 52 / 366 draft measurements and the deleted Nguyen Thi Van FAMILY_FIRST line are already in decisions/P6's 2026-08-30 Ménil Christophe de under FAMILY_FIRST drops the tussenvoegsel, and de Mesnil Juan reads Juan as part of the surname #467 entry, so the rule now points there. The test_order_correspondence.py note moves into the new bullet.

• docs/design/AGENTS.md: axis 13 (the migration check) gains one sentence. A block rule scopes sentences by the paragraph they stand in, while a skeleton field reads rule-wide, so a moved sentence's site has to be written into its lead. This is the trap the third review round found in P6.

Small wording changes (no change of meaning)

• P3 — (#395, landed) → (#395).
• P5 — Then names the reserve and its two lifts in a clause ("where the name has a name word to spare or a family comma or a given-name title has lifted that need (below)"). The reserve sentence itself is kept verbatim in the first Then bullet.
• P6 — Then carries the old rationale's "where something has already named the family and left the particle in a position that means nothing — after a family comma, or in the middle of a FAMILY_FIRST reading", so Because is shorter. "attaches as it did before" → "attaches" (history). The one-case Portuguese clause moves from the statement to an Accepted bullet with its example (SMITH, JOHN DO). The do / vd / mc report split and the several-word run's report move into the Then "What reports" bullet, and the mc-inherits cost stays Accepted. The old "every other spelling of such a word attaches" sentence, which followed all three stand-downs, is now the first sentence of that bullet and names its word (do) and the two stand-downs that can reach it (S6, S7). When says "trailing words that hold no name behind it aside", matching the run bullet. Two · boundary inputs where no particle ends the name (Maria Luisa y de la Cruz, de Anke van y) are in the main fence.
• P8 — "HumanName.initials() re-derives the conjunction decision instead of honoring the parse's tags: john e smith gives j. s. from the facade and j. e. s. from parse() #528 closed the split and decisions/P8 records it" → "(decisions/P8, HumanName.initials() re-derives the conjunction decision instead of honoring the parse's tags: john e smith gives j. s. from the facade and j. e. s. from parse() #528)". The Cyrillic example moves from the end of the Accepted clause to the bullet about Cyrillic capitals.

Old interacts: IDs

Rule	Named in the rule now	Dropped, and why
P2	P1, P4, M2, S2, S4, S6, S10	H5: H5's own Accepted clause states that a particle chain stops in front of the trailing title run; P2's old text never named it. C2: no chain in a tail part is C2's own bullet ("no particle chain (P2) is made in it"). P9: P9 cites P2's join itself; P2's text says nothing of P9.
P3	H1, P1, M2, C3, C10, S10, T4, P8 (+ N1, M1 from the text)	R1, C7: each names P3 in its own text (R1's render, C7's title-word join); P3's old text named neither. C4, C6, S6: the shared case span is P8's to state ("the suffix slot (S6) and the post-comma slot (C4, C6)"), and P8 still does.
P4	P1, H2, H3 (+ O4 from the text)	P5: P5 states its own precedence over P4 ("the join outranks the leading-position reading (P4)").
P5	C2, S2, M2, H1, H5, P2, P4, P6 (+ H3, S10 from the text)	none
P6	S2, S4, S6, S7, P5, P9 (+ R2 from the text)	P1, C7, C11, H5: each names P6 in its own text; P6's old text named none of them. A1, C1, M2: noise. Nothing in P6 reads them beyond "a family comma", which is C1's structure, and A1's general report contract.
P7	P1, P2, P6, T5 (named in the every-position bullet)	P3: one-letter connectives are P8's subject, not P7's. R3, R4: an initial's render follows from its role; R3 and R4 state it.
P8	P3, T4, M2, R3, R4, S10, S6, C4, C6	M4: noise. The marker's effect on the case span is M2's, and M4's lone-word family reading never meets the letter.
P9	P6, S4, H5, T4	P2: a deliberate drop. P9's before-the-comma clause came from P2 in the 2026-10-19 split, and P9 states it without leaning on P2's join (the code comment's "as P2's join does" is the code's analogy). S2, S6, S7: the old text names none of them. "whatever case" is the rule's own point, and S7's company is P6's stand-down, named in P6's Precedence.

Rules that look wrong

None was changed. One observation for Derek, not fixed here because the old text never covered it:

• Under FAMILY_FIRST, the no-comma site gives do two reports. Doe John Do (family-first) reads family 'Do Doe' and reports both suffix-or-name ("read as a family name rather than a post-nominal") and particle-or-given. Doe, John Do reports only particle-or-given. Old P6 stated the single particle fork for the comma site only, so the rule is silent on the family-first case, and the migration now scopes that sentence to the comma site. Whether one token should carry both reports there is open; it looks like A1's one-report-per-decision territory.

Verification

• uv run pytest (full), uv run mypy and uv run ruff check pass.
• The docs-review agent ran twice. Round 1 found five things, all fixed in the first fix commit: P6's "every other spelling attaches" sentence placed where the S7 bullet contradicted it; P6's When excluding post-nominal shapes; a comma-only lead on the post-nominal bullet; P4's "this is why" pointing at the wrong cause; two out-of-When boundaries in a Then fence. Round 2, on that fix commit, found three more in P6, all fixed in the second: the moved sentence had lost its subject and claimed every ambiguous credential attaches (Doe, John Ma does not); it listed P9, which never reaches do; When still left out trailing titles (Beethoven, Ludwig van Prof.).
• A third round (/pr-review-toolkit:review-pr: code-reviewer plus a prose-vs-parser accuracy pass) found that four P6 clauses master had scoped to the comma site by their paragraph read as rule-wide in their skeleton fields, and the family-first site falsifies each: the S2 precedence (Berg Jan vd under family-first keeps suffix 'vd'), the do / vd / mc report sentences, and the only-given-word guard. Conversely, R2's base guard, master's family-first clause, read as covering the comma site (van, Jan de attaches). Each lead now names its site. Also fixed: the post-nominal bullet now carries the Berg Jan de Jr. lines that show a post-nominal; P5's abdul V Smith (where the join fires) moved out of an Except fence; and P5's vd/mc Except now states its scope in the lead.
• Kept by choice (round 1, finding 6): P3's two single-letter carve-outs stay inside Then's sentence rather than becoming Except bullets, because code quotes that whole sentence verbatim (_group.py, _pieces.py). Its boundary Juan y Garcia stands under the carve-out's Then bullet.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.05%. Comparing base (a0edfed) to head (a64a0a5).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #777   +/-   ##
=======================================
  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.