Case repair reads the unjoined-particle mark, and P6 reports its fork (#407, #405)

[derek73]

Bundle 4's head: two render/report fixes that share a review pass rather than any code.

Closes #407
Closes #405

Why R5 comes first

Writing #407's R4 examples exposed that rules.md never documented the case-repair gate, and that R4's statement was narrower than the parser's behavior — it promised "an already-correct name comes back unchanged" where the parser keeps any mixed-case name, correct or not.

Worse, R4's boundary example did not pin what it appeared to: "Juan McDonald" passes under force=True as well, so it was consistent with both idempotence and the gate and distinguished neither. Shirley Maclaine — mixed case, wrong, and kept — is the witness that separates them.

So the first commit extracts R5 and changes no behavior. Verified on the released 1.4.0 wheel: the refusal predates git (45a1539, 2011-02-03, the svn import), and the force override arrived in bf1e0a5 (2016-06-02, shipped in 0.4.0). capitalized_forced is a new example pseudo-field, because force is an argument rather than a policy and cannot ride an example's annotation slot.

#407 — case repair reads the unjoined-particle mark

post_rules marks a name part whose every word is particle vocabulary: nothing joins them to a name, so they are not doing a particle's work. Three views already read the mark. Case repair did not, so ANH DO capitalized to Anh do while family_base read DO as an ordinary name word.

Whole part, not standing-alone — rules.md#R2's test is "'Every word', not 'standing alone'", and anh van do has a two-word all-particle family that a narrower rule reads wrong on a real Vietnamese name.

This deviates from 1.4.0 deliberately and does not restore it: 1.4.0 returns Anh do. The accepted cost is that a degenerate all-particle family capitalizes too, so juan van der gives Juan Van Der.

How it is verified, given the gate cannot see it. compare.py compares the seven role fields plus _ambiguities; the string capitaliz does not occur in it. That is by design rather than a gap — the harness exists for parsing behavior, where role assignment is emergent across the stage fold and the blast radius cannot be read off a diff, while case repair is a four-branch pure function. The deviation is pinned by the rules.md#R4 examples and tests/test_capitalization.py, and decisions.md#R4 records the limit.

While rewriting R4's statement: McDonald was never a vocabulary exception. CAPITALIZATION_EXCEPTIONS is {ii, iii, iv, md, phd}; McDonald comes from the separate Mac/Mc convention. R4 had been illustrating its promise with the one mechanism it was not about, and now names both separately — each with a row that dies when its own mechanism is removed.

#405 — P6's attachment reports the fork it decides

rules.md#A1 requires a genuinely uncertain reading to carry a report. Van Johnson reports particle-or-given; Nguyen, Thi Van decided the same fork silently and lost the given name.

_assign's scoping — "after a comma that fixed the family, a leading given-position particle is not meaningfully ambiguous" — was true of the family, which the comma fixes, and not of the particle behind it, which P6 decides separately.

Two kinds, keyed on the reading P6 overrode rather than on the word's vocabulary:

what assign had read	kind
a name word (middle)	PARTICLE_OR_GIVEN
a post-nominal (suffix — vd, mc)	SUFFIX_OR_NAME
a never-given particle	silent — nothing was declined

Keying on vocabulary instead is backwards for both collision words: Berg, Jan do reports particle-or-given, because do sits in the AMBIGUOUS acronym half and no credential reading was ever taken; and Berg, Jan de vd reports nothing, because behind de the run was read as name words. Both now have case rows.

The arms are ordered, suffix first, rather than asserted disjoint. A legal custom Lexicon adding a word to both particles_ambiguous and suffix_acronyms makes both true, and an assert there would raise on a parse — a totality violation of A1, whose only sanctioned exception is a caller's segmenter.

Beethoven, Ludwig van now carries a report. The textbook-correct Dutch listing gains one because the parser genuinely cannot distinguish it from Nguyen, Thi Van — the ASCII transliteration is the same string in both traditions. That is the most user-visible consequence of this PR.

post_rules gains ambiguities in the stage-ownership map, and the ParseState docstring twin moves with it.

Ledger

Seven names gain an ambiguity kind. The plan called for a new _ambiguities-only rule per ledger; that explains nothing, because classify matches on fields being a superset of the whole diff and these names already diff in middle/family from #379/#380. The existing rules were widened instead, each still satisfying #452's declared == union check. 1.4.0 needs nothing — _ambiguities cannot enter a diff below baseline 2.0.

Verification

• 6116 passed, 156 skipped, 9 xfailed; mypy and ruff clean.
• Gate green at all three baselines: corpus 1094, intentional diffs 229 / 194 / 102, unexplained: 0. Per-heading counts diffed before and after: identical.
• Every example row mutation-checked. Two inert rows were caught in review and withdrawn rather than shipped.

Review round

A four-agent review (code, tests, comments, error handling) found no Critical issues and nine Important ones, all fixed in the final commit. Highlights: the emitted detail said the run is "read as the family's particle" where R2's later mark can make every other view read it as a name word — reworded to name the attachment instead; the detail text and reported token span were pinned by nothing (both mutations left the suite green) and now have assertions; the release log said the tussenvoegsel fix moved seven names where the gate says nine; and AGENTS.md had re-landed two stale claims in a line this PR rewrote.

The test pass ran 22 mutations and found no inert row — every new test, cases.py row and rules.md example is killed by a mutation of the exact expression it pins. Totality was checked across ~250,000 parses under adversarial lexicons, randomized vocabularies, four locale packs and 24 policy combinations: zero exceptions.

Three issues were spun off from the review: #459, #461 (both found while measuring), and #458 above.

Out of scope but adjacent, called out deliberately

• All three ledger headers still said fields "is optional", which Deleting a rule's fields escapes #452's check and makes the rule maximally permissive — the symmetric twin of #451 #456 made false on 2026-08-29. Corrected in the same sweep, each in its own wording.
• A drifting count — "all 39 ambiguous particles" — removed from AGENTS.md, mechanisms.md and a cases.py row. The set holds 37 today. decisions.md#vocabulary-collisions keeps its figures: they are labelled "as of 2.2.0dev" and followed by "recompute before quoting them", which is the dated-snapshot form the counting rule permits.
• Two API docstrings and docs/concepts.rst were made stale by A tussenvoegsel attachment decides a particle-or-given fork and reports no ambiguity, which A1 requires #405 and fixed here. concepts.rst had said "the comma paths stay quiet on purpose", the exact claim this PR reverses.

Spun off, not fixed here

• Render views should honor the parser's conjunction/initial tags instead of re-deciding from the word #458 — render views should honor the parser's conjunction/initial tags instead of re-deciding from the word. Measured a pure refactor: 0 of 1094 names move.
• HumanName('JOHN SMITH MBA').capitalize() gives John Smith Mba — a credential acronym is title-cased as an ordinary word #459 — a credential acronym is title-cased as an ordinary word (JOHN SMITH MBA → John Smith Mba).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.65%. Comparing base (0847ecf) to head (e555d21).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #460   +/-   ##
=======================================
  Coverage   98.64%   98.65%           
=======================================
  Files          45       45           
  Lines        3177     3186    +9     
=======================================
+ Hits         3134     3143    +9     
  Misses         43       43           

☔ 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.