• App Router: backends are no longer attached during SSR — I18nProvider is a Client Component, but the App Router also renders it on the server, once per request, and its useState initializer therefore built a throwaway i18next instance in Node on every render. With a backend passed via use, that instance got a Node-side backend, and backends that refresh on a timer never let it go: i18next-locize-backend defaults reloadInterval to 1 hour whenever window is undefined, so every render left behind a live timer that kept refetching its namespaces forever. In production this showed up as a request floor that ratcheted upwards between deploys (measured on one app: ~90 requests/hour climbing to ~8,500 before a pod restart, ~190k downloads/day against a ~10k baseline). Backend plugins, including the built-in fetch loader, are now applied only in the browser, and partialBundledLanguages is off during the server pass, so SSR renders purely from resources. Non-backend plugins (detectors, post-processors) still apply in both passes.
• getResources accepts a language filter — getResources(i18n, namespaces?, languages?). The shared server instance preloads every supported language, so the default return value serialized all of them into the HTML; on a four-language app that was ~440 KB per page. Pass [lng, fallbackLng] to ship only what the client renders.