Measure the CLDR corpus with each locale pack too

[derek73]

Closes #672.

tools/differential/cldr_agreement.py gains a locale view. LOCALE_FILES maps each CLDR file to its language's pack:

• zh, zh_Hant, yue and yue_Hans map to ZH.
• ja maps to JA.
• ru maps to RU.
• tr and az map to TR_AZ.

Those renderings are parsed with parser_for(pack), over the same FAMILY_FIRST base where CLDR's order is surnameFirst. Their disagreements are recorded in the new cldr_agreement_locales.json, and test_cldr_agreement.py ratchets them per pack. The summary now prints each pack's before/after:

JA     ja: 7/24 default -> 12/24 with the pack
RU     ru: 17/24 default -> 17/24 with the pack
TR_AZ  az, tr: 32/48 default -> 32/48 with the pack
ZH     yue, yue_Hans, zh, zh_Hant: 22/96 default -> 42/96 with the pack

These figures reproduce the issue's table, except that ZH's default count is 22 where the issue measured 21 on 1f0b374.

How the issue's open questions were settled:

• The ja extra: there's no env switch. Without namedivider, the JA ratchet is a visible pytest skip. That's safe because CI's ja-extra job runs import namedivider before pytest, so it can't skip there.
• Re-recording: one --record rewrites both records. Without the extra, it keeps the recorded ja entries unchanged.
• Mapping completeness: a test fails on a registered pack that is neither in LOCALE_FILES nor in NO_CLDR, and on a mapped file with no renderings.
• Separate record: the locale view gets its own file. cldr_agreement.json is byte-identical.

Review found that record entries whose file leaves LOCALE_FILES fell in no pack's check and survived re-records. Fixed in the second commit: such entries are now reported as gone, and the re-record drops them. That commit also adds the lesson to AGENTS.md's filter-test paragraph.

Verified:

• ruff and mypy are clean, and the full suite passes (16,631 passed) without the extra.
• With the extra, the JA ratchet passes, and re-recording without it leaves both records byte-identical.
• Negative controls: when the locale view falls back to the default parser, the pack pin plus the ZH and JA ratchets fail. When the carry-over fix is reverted, its test fails.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.02%. Comparing base (c16df14) to head (14abd6a).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #673   +/-   ##
=======================================
  Coverage   99.02%   99.02%           
=======================================
  Files          46       46           
  Lines        4534     4534           
=======================================
  Hits         4490     4490           
  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.