Before: the script diffed each locale against EN at the flat-key level,
so the only plural variants ever generated were the two EN itself
defines (_one, _other). Locales whose CLDR plural rules require
additional categories (_few/_many in Slavic; _zero/_two/_few/_many in
Arabic; _many in Romance for fractional counts) silently shipped with
their non-_one/_other count values falling back to the English string.

Now: per non-EN locale, the script consults
`Intl.PluralRules(loc).resolvedOptions().pluralCategories` to discover
the required CLDR categories, fans out each EN pluralized base
(any `<base>_one` + `<base>_other` pair) into one expected key per
category, and adds the missing ones to the translation batch. The
prompt grew one rule (#8) telling the model that suffixes name the
CLDR category whose morphology it should inflect — Claude already
knows CLDR semantics so no per-locale lookup table is needed.

Helpers added:
- getPluralCategories(loc): wraps Intl.PluralRules; defaults to
  ['one','other'] on any unknown tag for safety.
- findPluralBases(enFlat): returns the bases where EN has both _one
  and _other (the only shape EN can express); the same set is reused
  for every locale.
- expectedKeysForLocale(enFlat, pluralBases, categories): the
  per-locale expected-key set used both by the missing-key detector
  in the main loop and by the post-write completeness scan.

Net behaviour on a locale that's already correct: zero change (the
existing _one/_other keys satisfy expected and `missing` stays
empty). On a fresh locale or any locale that's missing a CLDR
category, only the absent keys are filled — existing translations
are never overwritten (same as before).

Reviewer (PR #3908 P2/P3) caught two real issues:

bg.json — fallbackBadge contained 'živоте серии', a Latin-Cyrillic
mixed-script word ('živ' Latin + 'оте' Cyrillic). Now uses standard
Bulgarian 'серии на живо' (live series). Also fixed the same key's
fallbackTooltip where 'семе' (plant-seed) was a misread of the
tech-sense 'data seed' — replaced with 'Живите данни' (live data).

th.json — fallbackBadge / fallbackTooltip both contained 'สดใจ'
('cheerful' — literal: fresh-hearted), an artefact where the model
appended 'ใจ' (heart) to 'สด' (fresh/live). Replaced with the
unambiguous Thai-tech loanword 'เรียลไทม์' (real-time).

Programmatic sweep for the bg-style mistake (adjacent Latin/non-Latin
chars within a single word) across the full 6-baseline-keys × 20-
locales matrix surfaced only the bg one; ko 'ECB가' tripped the
scanner but is just an acronym + Korean particle (false positive).
Thai 'สดใจ' is a semantic error within Thai script — not detectable
by mixed-script heuristics; relied on the reviewer's read.

Prompt rule #8 (P3): the previous version said _few is "2-4" and
_many is "5+" for Slavic locales, which silently glosses over the
CLDR teen-case exceptions (e.g. Russian/Croatian _few requires
n%100 != 12-14). Reworded to explicitly tell the model to follow
CLDR rules exactly and NOT use the simplified rules of thumb. The
code path already delegates to Intl.PluralRules for category
selection — this just stops the prompt from biasing the model
toward incorrect teen-case morphology in future runs.

Three tooling fixes + two data fixes in response to Greptile review.

Tooling (scripts/translate-locales.mjs):

1. getPluralCategories null guard. The previous try/catch only caught
   constructor throws (RangeError on unknown locale tag), not the case
   where `resolvedOptions().pluralCategories` itself is absent (older
   Node where the property predates the spec). That `undefined` then
   crashed the next `for (const cat of categories)` mid-run with a
   TypeError. Now uses `?? ['one','other']` to cover both paths.

2. validateTranslation URL preservation. The model previously could
   silently paraphrase or drop URLs/paths embedded in translation
   sources (e.g. the methodologyLink `/docs/methodology/cii-risk-scores`)
   and the validator wouldn't catch it. Added a URL/path multiset check
   alongside the existing interpolation-token and HTML-tag checks. The
   path pattern requires the first segment after `/` to start with a
   letter so it doesn't fire on number fractions (`50/100`) or
   interpolation tokens (`{{count}}/{{total}}`). Verified against 6
   cases (preserve/drop/rewrite path; ignore fraction/interpolation;
   match https URL).

3. Skip leading-underscore "private" keys. Convention: any dotted-path
   segment starting with `_` is a translator-instruction key meant to
   stay in English so translators reading the raw files can understand
   it (e.g. `_methodologyLink_translatorNote` is a TODO note about its
   sibling `methodologyLink`). Sending those through the model produced
   the visible mistranslations in ar/ja/pt/th in the previous run.
   Added `isPrivateKey()` filter at the expected-keys layer so private
   keys never enter the missing-batch or the post-write coverage scan.

Data:

4. Reverted `components.cii._methodologyLink_translatorNote` to the
   English source in ar.json, ja.json, pt.json, th.json (the 4 of 20
   locales that the model translated; the other 16 already kept EN).

5. Aligned time-ago plural variants to their `_other` form across cs,
   es, fr, ro (22 strings total). The previous backfill chose more
   idiomatic prepositional forms (`před {{count}}h` / `hace {{count}}h`
   / `Il y a {{count}}h` / `acum {{count}}h`) while the pre-existing
   `_one`/`_other` use postpositional forms (`{{count}}h zpět` /
   `{{count}}h atrás` / `il y a {{count}}h` / `{{count}}h în urmă`).
   i18next picks exactly one variant per render, so users would have
   seen mixed-style timestamps depending on count. Greptile flagged
   only the fr capitalization symptom (`Il y a` vs `il y a`); the
   project-wide scan I ran surfaced the same wider issue in cs/es/ro
   so this fix covers all four locales.

Note on radiationWatch/thermalEscalation time-ago strings in fr.json:
the pre-existing `_one`/`_other` values are still English ("{{count}}h
ago") rather than French — a separate pre-existing translation gap not
in scope of this PR. Aligning `_many` to match `_other` keeps the
consistency rule but inherits the English text for those 5 keys;
fixing the root French translation there should be a follow-up.