Read halfwidth katakana as katakana (#594)

[derek73]

Closes #594.

The bug. Halfwidth katakana (U+FF65–FF9F) fell outside the script table, so 山田 ﾀﾛｳ read given 山田, family ﾀﾛｳ where 山田 タロウ reads family-first. Legacy bank, payroll and CSV exports written for JIS X 0201 systems still use this encoding.

The fix is one range: Script.KATAKANA now covers (0xFF65, 0xFF9F). Everything keyed on the script table follows.

Input	2.3.0	This branch
山田 ﾀﾛｳ	given 山田, family ﾀﾛｳ	family 山田, given ﾀﾛｳ
高橋一郎 ﾀﾛｳ (JA + segmenter)	segmenter is asked	already divided
ﾀﾛｳ·ﾔﾏﾀﾞ	one given token	given ﾀﾛｳ, family ﾔﾏﾀﾞ
ﾀﾅｶ. John	title ﾀﾅｶ.	given ﾀﾅｶ.
ﾔﾏﾀﾞ ﾀﾛｳ	given ﾔﾏﾀﾞ	unchanged (W4)

The ﾀﾅｶ. John row closes a limit that decisions.md#cjk-full-stops had recorded and left open in #322/#323.

Why a range and not normalization. The T rules forbid rewriting the input text, so normalizing at tokenize was never an option. An NFKC fold at classification reaches far past the kana (ＪＯＨＮ, ㈱). It also folds an uncomposable voicing mark (ｱﾞ) into the hiragana block, which would turn a wholly-katakana name family-first. decisions.md#W4 has the measurement.

Wholly-katakana names keep the declared order, halfwidth included. W4's rationale is amended, not its behavior. Systems with no kanji wrote Japanese names in katakana too, so "predominantly a transcription" no longer holds. The script can't tell a Japanese reading from a transcribed foreign name, so it settles nothing. usage.rst now gives the opt-in recipe: add (Script.KATAKANA, FAMILY_FIRST) to script_orders.

Differential. No corpus line held halfwidth kana, so the gate was blind to this until the eight new case rows. With them:

• five contract movers and two tolerated radar movers, each classified at every baseline: under the fix(#594) rules from 2.1.0 on, and at 1.4.0 and 2.0.0 the five under the canonical CJK rule's widened span with the two under a fix(#594) title rule;
• the canonical CJK rule's span copy widens to the whole block;
• U+FF65 leaves _SANCTIONED_EXTRAS for the table;
• all five baselines exit 0.

Docs.

• rules.md: W Background, W4 rationale and examples, H2 Accepted.
• decisions.md: a new W4 entry, and the cjk-full-stops limit marked closed.
• AGENTS.md: the pure-katakana sentence.
• usage.rst and locales.rst; a release-log bullet.

Review. Two rounds before opening (a code reviewer, plus the design-docs reviewer on the docs/design changes and then on the fix commit; fixes in 6451052, 151ca96, 6803db6), then /pr-review-toolkit:review-pr with code, test-coverage and comment reviewers. That round added coverage for the span's ﾟ/ｰ edges, the unspaced 山田ﾀﾛｳ and the ja adapter (1ea6e68), and fixed comment and prose inaccuracies (15e0358).

Out of scope, found in review. Full-width katakana typed with an uncomposable combining U+3099 (ア゙イ) already takes the kana license today, contrary to W4. Filed separately as #596, together with a W4 wording fix that review also turned up.

Checks. uv run pytest 11,119 passed. mypy, ruff and both Sphinx builds are clean. The suite was also run without namedivider, and the halfwidth segmenter path was checked with the ja extra installed.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.98%. Comparing base (d989c5e) to head (357f685).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #595   +/-   ##
=======================================
  Coverage   98.98%   98.98%           
=======================================
  Files          45       45           
  Lines        4220     4220           
=======================================
  Hits         4177     4177           
  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.