ci(i18n): untranslated-string ratchet baseline drifted above the real count in 10 locales #16976

What

repo_tests/i18n_untranslated_ratchet_test.py::test_untranslated_strings_only_shrink
failed on main (CI shard 6) — real translations landed in 10 locales
(ar de es fa fr he lv pl pt ur) without the frozen BASELINE dict being
lowered to match. Same ratchet-hygiene class as a fragmentation-ratchet
growth-not-caught failure, opposite direction: a shrink not recorded, so the
number stops being a deliberate claim (the test's own stated purpose).

Confirmed not caused by #16875 (the PR this shipped alongside): that test
counts placeholder strings identical to their English value; #16875 only
added new locale keys with real translations, it never touched an existing
key's value, so it cannot be what moved these 10 counts down.

Fix

Recounted with the test's own _untranslated() function, not by hand:

Acceptance criteria

• All 10 locales' BASELINE entries match _untranslated()'s current
  measured count exactly
• test_untranslated_strings_only_shrink passes for all 10 locales

Refs

Discovered as one of three independent main reds while unblocking base for
#16875/#16972's work; not attributable to any of them.