refactor(i18n): merge the shared catalogue under web's, no keys moved yet

[mrviduus]

Slice 2 of five. The machinery, with web's file still a full superset — so web wins at every leaf and the golden fixture's values are unchanged. This isolates "does the merge work" from "did we move the right keys": functionally the riskiest PR of the set, content-wise the safest.

apps/web/src/locales/catalog.ts deep-merges translations.en from @textstack/shared under locales/en.json, once at module load. No build config changed — the alias already exists and web already imports from that package.

Three properties that are load-bearing, not tidy

• The merge builds a new tree. Object.assign(shared.en, web) would mutate the module-cached shared catalogue for every other importer in the process — the shared package's own t(), and under vitest every other test file in the worker. That is the easiest way to write a merge that passes its own tests and corrupts somebody else's. Asserted both as a unit and against the real module graph.
• A string-vs-object collision throws rather than picking. There is no correct answer: the result would depend on read order and the loser's subtree would vanish silently.
• Arrays replace, never concatenate. None exist today, but tArray is real API and "append" is a surprising default for a translation.

The collision guard earned its place on its first run

highlights.empty is a string in shared and an object in web. A leaf-by-leaf comparison cannot see this — a key that is a leaf in one file and a branch in the other never appears in both leaf sets, which is why my earlier count of divergences missed it.

Both are live: web reads highlights.empty.title/.subtitle/.cta; mobile reads highlights.empty next to its own highlights.emptySubtitle. And this exact shape has already shipped a bug once — apps/mobile/src/lib/i18nKeys.test.ts:60 records it:

Web shipped exactly that — t('highlights.empty') where the locale held { title, … }

Resolved by renaming the shared string to highlights.emptyTitle, pairing with the emptySubtitle already beside it. One mobile line, one fixture line.

The type

TranslationData = typeof en is gone. en.json is an overlay now, so the literal type could not describe the catalogue. Nothing consumed it structurally, and missing-keys.test.ts — every literal t('…') checked against the real catalogue — was always the stronger guarantee. Traded deliberately.

Verification

The golden fixture gains 123 keys and changes zero values: web can now resolve what it inherits from shared, and pinning that is the point.

Browser-checked, because types do not prove a screen still reads right: home, /en/books, /en/highlights, /en/terms and /en/privacy all render real copy with no key literals leaking through — and that last page gates the production deploy via smoke.spec.ts.

🤖 Generated with Claude Code

https://claude.ai/code/session_011rgEMvYYi4Egj99dVtvm6E