`capitalize()` breaks a decomposed (NFD) word at its combining mark: `josé garcía` typed NFD gives `José GarcíA`

[derek73]

Case repair splits a word at a combining mark, so a name typed in decomposed form (NFD, the normalization macOS file names and some databases hand back) is repaired as if the letters after the accent were a second word.

>>> import unicodedata
>>> HumanName(unicodedata.normalize("NFD", "josé garcía")).capitalize()   # then str()
'José GarcíA'
>>> HumanName("josé garcía").capitalize()
'José García'
>>> parse(unicodedata.normalize("NFD", "renée o’néill")).capitalized()
'RenéE O’NéIll'

(Outputs shown NFC-composed for legibility; the repaired text keeps the input's form.)

Every release does this, 1.4.0 through 2.3.0 and master, plain and forced, on both surfaces. The parse itself is right: roles and tags for the NFD spelling equal the NFC ones, because _lexicon._normalize NFC-composes before every vocabulary lookup. Only the render view is wrong.

Cause. nameparser/_render.py:27: _WORD = re.compile(r"(\w|\.)+") is the sub-word splitter _cap_text runs per token. A combining mark is Unicode category Mn, which \w does not match, so garcía is the words garci and a with a mark between them, and a capitalized alone is A. The same split is why a mask misses an NFD word: pé.x. under a caller's péx → Péx mask gives Pé.X. composed and PÉ.X. decomposed, the pe half finding no key and taking the acronym clause. capitalized()'s docstring already records the output-side twin (ǰ upper-cases to J plus a combining caron, which a second pass reads as two words); this is the input side of the same limit, and it is reached by ordinary Latin names rather than by a Unicode corner.

Reach. No corpus or case-row name is written NFD (the only combining marks in tools/differential/*.jsonl and tests/v2/cases.py are the Bengali vowel signs of one caseless row and an emoji modifier), so the differential gate cannot see it, and capitalized() is not a compared surface in any case (decisions.md#R4).

Options:

1. Let the splitter carry marks. Python's re has no \p{M}, so either an explicit class of the combining blocks or a splitter that asks unicodedata.category(ch)[0] == "M". The output keeps the input's form, so rules.md#R4's "Repair changes case and nothing else" holds as tests/v2/test_properties.py::test_case_repair_changes_case_and_nothing_else measures it (a casefold comparison, which does not normalize). Recommended.
2. NFC-compose before repair. Fixes the split but changes the output's normalization form, which that property test would report as a change beyond case, and R4's statement would need a carve-out.

Either way a rules.md#R4 example line in NFD and a property-test row pin it; _apply_mask and _letter_run_ge2 already read a mark as a non-letter consistently and need no change for the mask case once the word reaches them whole.

Noted in #539's description as a follow-up candidate; measured 2026-09-24.