ci(i18n): let a protected term be satisfied by a declared equivalent form

[bloxster]

Why

Some protected terms are the same word in two shapes — a singular and its plural. Whether an English loanword takes an -s is a decision each language makes for itself, not one English makes for it. Italian writes "gli Unified Address", because Italian does not pluralize borrowed nouns; Japanese, Korean and Chinese do not inflect plurals at all.

Demanding one exact shape therefore forces English grammar into 18 languages that do not share it, to assert a term the other shape already carries.

This is not hypothetical. ZK-SNARKs is protected and its singular was not, so on all four English pages using it, ja/ko/zh/it carry the English plural verbatim. They had no choice — the gate required it.

What this does

A translation may now satisfy a protected term with any form declared equivalent to it, via a new optional key in translation/protected-terms.json:

"equivalentForms": [
  ["ZK-SNARK", "ZK-SNARKs"]
]

A group means "these are one term", so membership is symmetric — any member asserts it. The source side still tests the exact term, because the English page really does use that form; only the translation side accepts an equivalent.

Declared, never derived

An earlier draft inferred the pairing by stripping a trailing s/es. Two reviewers independently rejected that, and they were right:

• it makes Argos interchangeable with Argo, and zec.rocks with zec.rock — different things that merely look related;
• it breaks on acronyms (CSS → CS);
• it silently misses irregular plurals such as Technologies / Technology;
• and it got its own motivating case wrong, deriving Unified Addresse from Unified Addresses, so the relaxation would never have fired for the term that prompted it.

A maintainer now states which forms are one word; the checker never guesses.

Every malformed declaration is fatal: a member absent from preserveVerbatim, a term in two groups, a duplicate within a group, a group of fewer than two, a non-array, a blank or whitespace-padded entry. A group that was quietly ignored would weaken this gate in the way least likely to be noticed — and a blank entry is the one shape that could do real damage, since a term of " " matches nearly every page.

It also closes a false negative

satisfiedIn is applied to classify()'s baseline test, and that tightens the gate rather than merely preserving it. If a base translation carries only the singular and a change strips it, the previous exact-match baseline made violatedAtBase true — bucketing the result pre-existing and exiting 0, silently passing a change that removed the last protected form. It is now correctly introduced, exit 1.

Base content is judged by the base declaration, read from git at the merge base, not by this change's config. Otherwise a PR that removed an equivalence group could never fail the gate: the base would look as though it had already been violating. A base config that cannot be read or parsed falls back to the head declaration rather than failing a PR over history it did not write.

Verification

check	result
unit tests	51 pass / 0 fail (38 existing + 13 new)
whole-tree violations	72 before and after
scoped gate on this branch	exit 0
remove the group and strip the last form	exit 1, reported as introduced
malformed declaration	fatal, exit 1, names the offending entry

The logic lives in translation/lib/term-forms.mjs — pure, no filesystem, process or clock — which puts its tests where the existing hash-lib-tests job already globs (translation/lib/*.test.mjs), so no workflow change is needed. They cover the declared pair in both directions, the -es case the old derivation got wrong, look-alikes that must not pair, and each validation failure.

termRegExp now memoizes; it was recompiling per (page, locale, term).

Note on scope

This changes nothing on the current corpus — 72 violations before and after — because the harm already landed: the locales were forced to carry the plural, so no violations remain to clear. The value is preventive, and it is why the companion PR adds ZK-SNARK rather than removing ZK-SNARKs.

Merge order

Stacked on #1965, which adds the ZK-SNARK singular that the declared group refers to. Merge that first; the diff here collapses to the checker change once it lands.

Known limitation

A PR that edits equivalentForms while also changing translations is judged partly by its own declaration, so removing a group cannot itself fail the gate in every case. Reviewers should treat a change to equivalentForms as needing a whole-tree audit (node scripts/check-protected-terms.mjs with no --base) rather than trusting the scoped result alone.