refactor(frontend): replace Apollo Client with urql for the GraphQL transport [INFP-563]
What & why
The frontend used @apollo/client as a transport-only GraphQL layer — no Apollo hooks, no Apollo cache (fetchPolicy: "no-cache", queryDeduplication: false), no subscriptions. TanStack Query is and remains the sole caching and request-deduplication authority. Apollo was paying for a cache and a hooks layer the app never used.
This replaces it with @urql/core + @urql/exchange-auth, keeping the imperative graphqlClient.query() / graphqlClient.mutate() interface so the ~95 call sites are untouched by design.
Dependencies: -@apollo/client, -apollo-upload-client, -@types/apollo-upload-client, +@urql/core, +@urql/exchange-auth.
Bundle: 4,574,160 → 4,539,212 bytes gzipped JS (−34,948 bytes, ≈ −34 KB), measured by summing gzipped dist/assets/*.js across a production build on the same machine.
No user-facing behaviour change. Auth, priority header, error routing, token refresh-and-replay, and file uploads are all preserved.
Notable implementation details
Four places where urql's defaults differ from Apollo's and parity had to be restored deliberately:
preferGetMethod: false — urql's within-url-limit default sends small queries as GET, which the backend's /graphql answers with the SPA fallback index.html. Forced to POST, with a regression test.
__typename injection — Apollo's InMemoryCache added it automatically and several dynamic object queries relied on it. Restored with a mapExchange running formatDocument.
One urql client per endpoint — this is the subtle one. urql's Client merges concurrent operations by hash(query, variables) and ignores the URL, but this app carries branch and point-in-time in the URL. A single shared client therefore served one branch's data to both callers of two concurrent identical queries on different branches — a branch-safety hazard, not a caching nicety. getGraphqlClient(branch, date) memoises a client per endpoint, so each branch/date keeps its own in-flight operations. Verified against @urql/core directly: with one shared client the two-branch case produces 1 fetch and identical data for both callers; with per-endpoint clients it produces 2, each correct. Same-endpoint identical queries still share one request, as before.
Error routing — handleGraphQLErrors keeps the catalogue-code routing (TOKEN_EXPIRED / AUTHENTICATION_REQUIRED → /login, PERMISSION_DENIED silent, generic → the caller's processErrorMessage). Both network and GraphQL errors reject, so a resolved result carries data only — every use-case already threw on errors, and the ones that did not dereferenced null.
The transport also dropped things that turned out to be dead: the X-Priority header is a client-level constant (no call site ever overrode it), and GraphQLRequestContext.date is now Date | null (nothing ever passed a string).
Testing
pnpm test— 1096 passing (159 files), including new regression tests for POST-not-GET,__typenameinjection, endpoint targeting, the priority header, the refresh-and-replay contract (one refresh, one replay, then bail), cross-branch / cross-date concurrent operations, the multipart shape of aFileupload, andX-Prioritysurviving the replayed operation after a refresh. The last two cover guarantees that rest entirely on@urql/*internals and had no fast-suite coverage.pnpm exec biome ci .— clean.pnpm knip— clean.pnpm build— succeeds.pnpm exec betterer ci— clean (190 issues, unchanged).
🤖 Generated with Claude Code