cli: platform-page-i18n-parity's "no silent English fallback" claim overstates what the test checks, and #8765 makes the gap real #8878

Restart-when: any PR touches packages/cli/test/platform-page-i18n-parity.test.ts, or closed #8827

Observation found while implementing #8765 (PR #8877). Filed unassigned, out of that card's file surface (packages/cli, and the card's surface was packages/platform-objects/src/apps/translations/).

What the test says vs what it asserts

packages/cli/test/platform-page-i18n-parity.test.ts:

it('translates every page in every shipped locale (no silent English fallback)', () => {
  const enKeys = Object.keys(pagesOf(EN)).sort();
  for (const locale of locales) {
    expect({ locale, pages: Object.keys(pagesOf(locale)).sort() })
      .toEqual({ locale, pages: enKeys });
  }
});

The assertion is a key-set claim over page NAMES — every locale carries an entry for each of the three plugin-contributed pages. The parenthetical promises something stronger and different: that no page renders English text in a non-English locale. A page entry that exists but whose leaf copy is English satisfies the assertion and defeats the promise.

That was already true before #8765 (an entry could always hold an English string — that is literally what --fill=default writes). What changes now is that it is no longer hypothetical: under the ruled Option B, a pages.* leaf whose source string has been edited is deliberately served as the source text, and this test stays green while it happens — correctly, since staleness must not fail a build.

Why file it rather than fix it in #8877

The claim is a test title, not a defect: nothing is mis-asserted, and no behaviour is wrong. But this directory has now paid three times in one day for in-tree prose that reads as a guarantee it does not implement (#8828, #8820, and the clause #8877 had to qualify), and a title is read more often than the body of the it it names.

Suggested shape

Narrow the parenthetical to what the assertion actually covers — an entry exists per page per locale — or, if the stronger claim is wanted, assert it: walk the leaves and require each translated locale's value to differ from en's except where a recorded source hash marks it as deliberately falling back (packages/platform-objects/src/apps/translations/source-hash.ts). The second is a real coverage claim and would want its own decision about pages.* locale coverage, which #8877 explicitly did not make.

Backlinks: #8765 (the ruling), #8877 (the implementation), #8827 (the adjacent pages.* coverage question).

Generated by Claude Code