F182: i18n infrastructure (SPA foundation) — react-i18next, coordinated bootstrap, errorCodes, pilot

[mforce]

The SPA i18n foundation for #182 — the plumbing so screens add catalog keys instead of hardcoded strings, and the stored §4.5 language finally drives display. English-only; no second language ships here.

#182 stays open as the tracker for the full string sweep of the remaining ~20 screens (which blocks Phase 1.5). This PR is the foundation + a worked pilot.

What this adds

• react-i18next + a typed English catalog (src/i18n/en.ts) — namespaced (common/auth/sales/account/errors), CustomTypeOptions makes t("typo") a compile error, keySeparator:false (the errors keys are dotted API codes), init-once singleton with fallbackLng:"en" (missing key → English, never blank/raw key).
• Coordinated no-flash bootstrap (src/session/SessionContext.tsx) — inside ProtectedRoute, SessionProvider reads /me + /account via Promise.allSettled, resolves the UI language (me.language → farm-locale subtag → English, case-insensitive, pack-gated), calls i18n.changeLanguage, and gates the shell (if (!result) return null) until the language is set — so the authenticated shell never paints one English frame first. It hands the account to FarmProvider (new tri-state initialAccount) so /account is read once. MeContext/useMe expose the user.
• parseError → catalog (src/api/client.ts) — maps errorCodes (from i18n infrastructure (API half): validation error codes + users.language + GET/PATCH /me #45) to catalog messages, index-aligned with a per-slot English fallback (i18n.t(key, {defaultValue: serverMsg})); ApiError.message stays a single flattened string (all ~30 consumers untouched).
• Formatting stays farm-locale-driven (§4.5) — money/date/number never read UI language; a guard test pins the invariant so a future change can't wire it in.
• Per-user language selector — wired to PUT /me/language, hidden while English-only (renders nothing until a second pack ships).
• Pilot — every user-facing string in Login.tsx + SalesPage.tsx externalized (111 sales keys) as the worked pattern; t() for plain, <Trans> for JSX-interleaved, t(key,{var}) for interpolation.
• web/CONTRIBUTING-i18n.md — the t()/<Trans> convention, key naming, the formatting boundary, the singleton-for-non-render rule.

Deliberate decisions

• SessionProvider owns the single /account read and seeds FarmProvider (not two independent providers) — this is what closes the flash window and avoids a double fetch. FarmProvider's standalone self-fetch path is preserved for its own tests.
• Promise.allSettled, not all — a /me failure preserves a good /account (its timezone/locale drive date correctness) and vice versa; only an account-read failure surfaces FarmProvider's banner.
• Pilot = Sales (over daily-entry) — its test already flows through renderWithProviders (the i18n wiring point) and it exercises dialogs/tables/interpolation/role-gating.
• Coverage floors RAISED (statements 83→84, functions 77→78; lines/branches held) to the honest measured state — no locked path (src/lib/src/farm 100, src/auth, src/api/client.ts) lowered.

Out of scope (tracked follow-ups in #182)

• The full string sweep of the remaining screens (shared components, nav, status badges, a11y labels, dialogs, the Help page).
• Externalizing raw enum displays (StatusBadge, payment method, unit) — invisible today (enum === English label), a completeness gap for the first non-English pack.
• Any second language / translation (Phase 1.5, epic EPIC: Phase 1.5 — Egg product hardening #15). The Help-page (F18: In-app help — user guide + glossary in the SPA #71) i18n note rides with the first visible selector; GLOSSARY already defines UI-language (from i18n infrastructure (API half): validation error codes + users.language + GET/PATCH /me #45).

Testing

npm run test:coverage 643 tests green; coverage 84.24% stmts / 73.91% branches / 78.27% funcs / 87.66% lines (above floors); npm run typecheck clean; npm run build succeeds. New coverage: catalog fallback, resolution chain (case-insensitive), no-flash gating + independent-settle + cancellation (observable-effect assertion), parseError code→message + index-alignment + English fallback, formatting-independence guard, hidden + multi-language selector.

Review

Reviewed per policy — plan reviewed pre-execution (codex + pi + Claude; findings folded in before any code), each of the 9 tasks task-reviewed, then a whole-branch review by two Claude agents + pi (all ready to merge; codex's run produced no usable verdict). One real (dormant) bug the reviews caught — the language selector read a stale me.language and would snap back after a switch — was fixed in this PR (current = i18n.language) with a regression test, since the selector is unreachable while English-only but would bite the moment a pack ships. Persist/switch failures now log rather than swallow.