Fix case repair splitting a decomposed (NFD) word at its accent

[derek73]

Closes #542.

A name typed decomposed (NFD: macOS file names, some databases) writes í as i followed by a combining accent (U+0301). Case repair's word regex (\w|\.)+ matches no combining mark, so the accent split the word and the letters after it were repaired as a separate word. josé garcía came out as José GarcíA. The parse was already right; only the capitalized() / capitalize() view was wrong.

The fix (nameparser/_render.py)

This is the issue's option 1. Option 2, composing the text to NFC before repair, was declined: it hands back text in a form the writer didn't use, which rules.md#R4 forbids and test_case_repair_changes_case_and_nothing_else would report. A decomposed word now repairs the way its composed spelling does, and the output keeps the form it was typed in.

• Word splitting. _sub_words replaces _WORD.sub: each match runs on through the combining marks after it, tested with unicodedata.category(ch)[0] == "M". A hand-written list of combining-mark ranges would drift; it would already miss Adlam. A mark with no letter before it starts no word of its own.
• Mask checks. The two places that ask about a letter's neighbours, _apply_mask's split-off-initial check and _letter_run_ge2, now look past marks via _beside. The issue said these needed no change. Measured, that was wrong: two mask shapes still disagreed once only the splitter was fixed.
• Shape checks. The Mac/Mc check and the two initial checks (inside hyphenated names, and for raw text spliced into a field) now test the word's composed spelling.
  • macée was MacÉe composed against MacéE decomposed before this change.
  • A decomposed й. read as the conjunction й where its composed spelling is an initial: ivan petrov-й.-sidorov.
  • The docs review found the й. case.
• Side effect. The documented ǰ limit is gone: a second forced pass now leaves J̌o unchanged instead of giving J̌O, because the mark stays with its letter.

Verification

The differential gate can't see this, because capitalized() is not a compared surface. Coverage instead:

• Corpus walk. test_a_decomposed_name_repairs_as_its_composed_twin re-encodes every corpus and case-table text that has accents, then compares its repair with the composed text's on both surfaces, plain and forced.
  • Recorded control: 144 disagreements over 137 texts with the old _render.py. A live control test re-breaks the code on purpose to show the walk can fail.
  • It sets aside 47 texts whose decomposed form parses differently. All are unspaced Hangul, which is matched as written by a documented decision, and a test checks that only Hangul is ever set aside.
• Unit tests. test_render.py covers the shapes no corpus name reaches: Mac, the two mask shapes, the hyphenated й., the spliced-in й., the ǰ repeat-pass case, and a leading mark. Each was confirmed to fail without the change; the mask rows were also checked against a version that ignores marks.
• Rule examples. rules.md#R4 gets two NFD example lines. Their expected values spell the accent as ́, so it's visible that the output stayed decomposed, and an editor that silently converts the file to NFC would make them fail. corpus_rules.jsonl is regenerated to carry them.
• Checks run: the full suite (11059 passed), mypy, ruff, the Sphinx and README doctests, and the differential gate at 1.4.0 / 2.0.0 / 2.1.0 / 2.2.0 / 2.3.0, all exit 0. The two new NFD corpus names produce no diff.
• Design-docs review. Two rounds. The first found the й. gap, and before-fix outputs that had been copied from an intermediate state. The second found three wording errors. All are fixed.

Out of scope, recorded in decisions.md#R4

• initials() takes a token's first character, so for a decomposed initial it drops the accent: émile zola NFD gives e. z.. That's a rules.md#R3 view, not case repair, so it gets its own issue.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.97%. Comparing base (97af1f0) to head (81b0481).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #584   +/-   ##
=======================================
  Coverage   98.97%   98.97%           
=======================================
  Files          45       45           
  Lines        4181     4210   +29     
=======================================
+ Hits         4138     4167   +29     
  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.