# Bundle Vendor-Split + Test-Gap Fills — Design **Date:** 2026-06-09 **Status:** Approved (brainstorming) — ready for implementation planning. **Issue:** #67 (vendor-split the bundle for cache stability/headroom; fill the audit's unit-test + story gaps). ## Context The production build emits a single ~216 KB-gz entry chunk (the largest, measured by `scripts/check-bundle-size.mjs`, budget 250 KB) that bundles the framework deps (react-dom, react-router 7, @tanstack/react-query, @base-ui/react, i18next) together with app code — so every app edit invalidates the whole chunk's cache. `vite.config.ts` has no `build.rollupOptions`. Separately, four well-isolated units have no direct tests, and the composed combobox primitive has no story. All changes are additive/low-risk; `check:size`/`check:colors`/the existing tests are the guards. ## Components ### 1. Vendor split — `vite.config.ts` Add a top-level `build` block (sibling of `plugins`/`resolve`/`server`/`test`) with `manualChunks`: ```ts build: { rollupOptions: { output: { manualChunks: { react: ["react", "react-dom", "react-router", "react-router-dom"], "base-ui": ["@base-ui/react"], query: ["@tanstack/react-query"], i18n: ["i18next", "react-i18next"], }, }, }, }, ``` - The app entry chunk then carries only app code; the listed deps land in cache-stable vendor chunks (`react-*.js`, `base-ui-*.js`, `query-*.js`, `i18n-*.js`). `react-router` **and** `react-router-dom` are both listed so router code (v7 re-exports `react-router` internally) stays in one chunk and React isn't duplicated. - Only affects `pnpm build` (production); the Storybook/vitest browser builder is independent, so the test suite is unaffected. - `check:size` measures the largest emitted chunk; after the split the largest is a vendor chunk well under the 250 KB budget (verified in implementation by running `pnpm build` + `pnpm check:size` and reporting the chunk sizes). If a chunk is unexpectedly oversized or React duplicates, adjust the groups. ### 2. Test-gap fills (4 unit tests) **`objects/prune-fields.test.ts`** (new) — `pruneFields(fields, localizedTextKeys: Set, defaultLang)`: - scalars pass through; `undefined`/`null`/`""` top-level values are dropped. - a non-array object value whose key is in `localizedTextKeys` keeps **only** the `defaultLang` entry; other-language entries are dropped. - a non-array object value whose key is **not** in `localizedTextKeys` keeps all (non-empty) language entries. - empty inner entries (`""`/`null`/`undefined`) are filtered, and a map left with zero entries is dropped entirely. **`components/delete-confirm-dialog.test.tsx`** (new) — the delete-in-use flow: - render ``, open it (click the trigger), click the confirm action. - when `onConfirm` rejects with `InUseError(3)`, the dialog shows the `actions.inUse` message (containing the count `3`) and **stays open** (the confirm action is still present / the dialog isn't dismissed). - when `onConfirm` resolves, the dialog closes. (`DeleteConfirmDialog` routes its catch through `errorMessageKey`, which maps `InUseError` → `actions.inUse` with the count — so this also exercises the shared error mapping.) **`lib/labels.test.ts`** (new) — `labelText(labels, lang)`: - exact `lang` match wins; else falls back to the `"en"` label; else the first label; else `""` for an empty array. **`lib/format-date.test.ts`** (new) — `formatDate(value, lang)`: - a valid `"YYYY-MM-DD"` formats via `Intl.DateTimeFormat(lang, { dateStyle: "medium" })` and is **not** shifted off its calendar day (assert the rendered string contains the expected day number for a fixed date + a fixed `lang` like `"en"`). - `null` → `"—"`; a non-date string (e.g. `"not-a-date"`) is returned unchanged; a non-string non-null (e.g. a number) is `String()`-ified. ### 3. Combobox story — `components/ui/combobox.stories.tsx` (new) A visual/interactive story for the composed combobox primitive, mirroring the existing `ui/*` story format (`@storybook/react-vite` `Meta`/`StoryObj`, single-quote + no-semicolon, `tags: ['ai-generated']`, a `play` using `canvas` + `storybook/test`). A small controlled wrapper renders `ComboboxRoot` with a handful of string options (`ComboboxInput`/`ComboboxClear`/`ComboboxTrigger`/`ComboboxPopup`/`ComboboxList`/ `ComboboxItem`/`ComboboxEmpty`), matching how `OptionsCombobox` composes them. A `Default` story; the `play` asserts the input renders and (optionally) typing filters the list. (Runs as a browser test in the Storybook vitest project — a few seconds in CI.) ## Error handling / edges - `manualChunks`: Rollup orders chunks by the import graph, so vendor chunks load before the app chunk that imports them — no load-order/duplicate-React risk when react+react-dom+router share one named chunk. - `formatDate` parses `\`${value}T00:00:00\`` (local midnight) to avoid a UTC day-shift; the test pins a fixed `lang` to keep `Intl` output deterministic (assert a substring like the day number, not the full locale string, to stay robust across ICU versions). - `delete-confirm-dialog` test drives the Base UI AlertDialog in its portal (`within(document.body)` for the confirm action), mirroring existing portal-aware tests. ## Testing - The 4 unit tests run in the jsdom project; the story adds one Storybook browser test. - **Gate:** `typecheck`/`lint`/`test`/`build`/`check:size`/`check:colors` green; `check:size` reports the new largest (vendor) chunk under budget; no new dependency; no new i18n keys; no codename. en/sv parity unaffected. ## Acceptance criteria 1. `vite.config.ts` has `build.rollupOptions.output.manualChunks` splitting react/base-ui/query/i18n into their own chunks; `pnpm build` succeeds and `check:size` passes (largest chunk reported, < 250 KB gz, no React duplication). 2. New tests exist and pass: `prune-fields.test.ts`, `delete-confirm-dialog.test.tsx` (InUseError → `actions.inUse`, stays open), `labels.test.ts`, `format-date.test.ts`. 3. `components/ui/combobox.stories.tsx` exists with a working `Default` story (browser-test green). 4. Full gate green; no new dependency; no new i18n keys; no codename; existing tests unchanged. ## Out of scope → follow-ups - The buildkit/Dockerfile CI migration (the robust fix for the slow native runner; overlaps #25). - Deeper treeshaking/bundle analysis; route-level code-split changes beyond the existing `lazy()` boundaries.