The argument for doing this before launch is a calendar slot, not a cost model. Google Play holds a build in closed testing for fourteen days regardless, and the author's judgement in the CannyCart build log is that fourteen days of real people on translated builds beats fourteen days of them on English [2][18]. Nothing in the post prices the after-launch alternative. If your release path does not force a wait you cannot compress, the retrofit competes with launch work rather than hiding inside it, and the survey numbers are what you are actually buying down: no i18n infrastructure, roughly 660 user-facing text nodes, 53 files [1].
Those nodes came out as 568 English keys at launch [5]. Against the rounded 660, extraction removed about 92 duplicate strings, near 14 percent, which is the reuse you only find by reading every alert and toast in one pass [1]. The machine tier then inherits that surface: thirteen locales at 568 keys is 7,384 generated strings, from a diff-only pipeline that had to be forced into structured output because free-form JSON kept breaking on quotes [15][2]. That is a lot of copy whose first reader is a user.
The plural-rules polyfill is therefore load-bearing rather than tidy. The author reports that the Hermes engine ships a real `Intl` on some platforms and a stub on others, and that Arabic needs six plural categories [4]. A stub does not throw. It hands back a category anyway, and the Arabic dual reads as a fluent sentence that is wrong about how many items you have. Thirteen of the nineteen shipped locales get no human review, so nothing downstream catches it [3]. The same class of trap already hit this codebase once, on thousands separators [19].
First paint is the other place the design shows. Resources are bundled statically so the first frame already has copy instead of raw keys, `dayjs` switches locale alongside so dates move with the words, and a per-user AsyncStorage cache resolves the language before the profile query returns [3][6][8]. `UserProfile.locale` is the durable value that roams; on account switch the language resets to the device default and is then restored from the incoming profile [8][9].
Arabic is in the launch tier precisely because right-to-left finds every layout assumption, and the flip only lands at the next launch [16][12]. The one-shot restart prompt is mounted inside the onboarding gate, so a first-run Arabic choice finishes onboarding before the app asks to relaunch [12]. Onboarding also gained a language step ahead of currency, pre-selecting the device language [11].
What transfers here is ordering. The cache resolves before the profile query, and the polyfill lands before the unreviewed locales do.
