refactor: convert course recommendations to React Query

[brian-smith-tcril]

Summary

Converts the course-exit course recommendations feature from Redux to React Query. This is the pattern-setter PR for the wider Redux → React Query migration tracked in #1946 (per OEP-0067 ADR-0010), following the same path frontend-app-authn and frontend-app-learner-dashboard took.

Behavior is unchanged — recommendations still fetch from discovery and render the same table (or the catalog-suggestion fallback for < 2 results / errors).

What changed

• React Query infra: add @tanstack/react-query; an app-level QueryClient/QueryClientProvider in src/index.jsx (bare defaults, matching the frontend-base shell's client); and a QueryClientProvider in the shared test render().
• Colocated data layer: course-exit/data/queryKeys.ts (query-key factory rooted at a new appId constant in src/constants.ts) + course-exit/data/apiHooks.ts (useCourseRecommendations, a useQuery over the existing getCourseRecommendations).
• Redux removed: delete course-exit/data/slice.js (the recommendations slice) and data/thunks.js; remove the recommendations reducer from src/store.ts and the test store. CourseExit now calls postUnsubscribeFromGoalReminders directly from data/api.js.
• Consumer: CourseRecommendations branches on React Query flags (isPending/isError/isSuccess) instead of a Redux status string; the sendTrackEvent call moves to a colocated track.js.

Decisions

Full decision log

Scope

• Pattern-setter = course-exit course recommendations. This PR stands up the
  React Query plumbing and converts the smallest genuinely Redux-backed read
  end-to-end, as the reference the wider Redux → React Query effort (Convert Learning from redux to Context + react-query #1946) copies.
  We deliberately did not use an api-only leaf (preferences-unsubscribe,
  celebration, enrollment-alert) — those use no Redux, so converting them
  wouldn't demonstrate the actual removal. Recommendations is one recommendations
  slice + one thunk + a single consumer (CourseRecommendations.jsx) that reads
  data → useQuery.

Dependency

• @tanstack/react-query@^5.90.19 as a direct dependency. Major v5 matches
  the reference apps (authn ^5.90.19, LD ^5.90.16) and frontend-base's peer
  requirement (^5.81.2); resolved to 5.101.4. A dependency (not
  peerDependency) because the app bundles its own deps.

React Query client config

• App-level client is bare: new QueryClient() (no staleTime/retry
  options). Matches the frontend-base shell's own bare client (shell/site.tsx:
  new QueryClient()). We rejected copying the reference apps' values because
  neither is principled and they disagree:

  • LD sets staleTime: 5 * 60_000. Traced to PR chore(i18n): update translations #786: it originally shipped
    staleTime: 60 * 60_000 (1h); reviewer (arbrandes) pushed back ("quite a
    large stale time … default gcTime is 5 min, so it's kind of pointless …
    what's the intended behavior?") and it was reduced to 5 min in a "chore: minor
    improvements" commit — review-nudged, not a real product decision.
  • authn sets retry: false, but for an auth-specific reason (arbrandes: "a
    failed login retried 3 times would be confusing") that doesn't apply to
    reading recommendations.

• Test client (createTestQueryClient in setupTest.js): retry: false on
  queries + mutations; no gcTime. retry: false is universal across the
  references' test setups (authn createWrapper, LD test wrapper, and the shell's
  own test files) — without it a failing query retries 3× with backoff, slowing
  tests and flaking error-path assertions. We dropped gcTime: 0 (LD's test
  wrapper sets it, but authn and the shell tests don't): render() creates a
  fresh client per call, so cache isolation is already guaranteed. The
  bare-app-client / configured-test-client split mirrors the shell (bare
  site.tsx client, configured test files).

• QueryClientProvider placement. Nested just inside AppProvider
  (AppProvider > QueryClientProvider > …) in both src/index.jsx and the shared
  test render(). Matches LD's ordering; AppProvider stays the outermost app
  wrapper.

Query keys

• Colocated per-feature, rooted at a shared appId constant. Keys live in
  each feature's data/queryKeys.ts (here course-exit/data/queryKeys.ts), rooted
  at appId from src/constants.ts ([appId, '<feature>', …]). We chose this
  over LD's fully-centralized src/data/ data layer:
  • Learning already organizes by feature (≈a dozen existing feature data/
    dirs), so colocation matches the codebase and keeps each conversion PR local
    and low-conflict; a central keys file would be a hot shared file and a large
    upfront structural move.
  • Rooting at a shared appId constant (authn's discipline —
    import { appId } from '../../constants') gives collision-safety and a
    consistent namespace without the centralization churn.
  • Keys can be lifted into src/data/ later if cross-feature invalidation
    actually becomes common (it rarely does).
  • Added export const appId = 'learning'; to src/constants.ts (value matches
    the legacy APP_ID env var).

Consumer conversion (CourseRecommendations.jsx)

• Branch on React Query's flags directly (isPending/isError/isSuccess),
  not a derived legacy status string. Matches the reference repos, which use RQ
  flags in components (e.g. LD Dashboard: const { data, isPending } = …;
  MasqueradeBar: !isError && !isPending). An earlier draft derived a
  recommendationsStatus (LOADING/LOADED/FAILED) to keep the old branch
  conditions; dropped as un-idiomatic. Behavior is identical
  (LOADING→isPending, FAILED→isError, LOADED→isSuccess).
• Tracking extracted to a colocated track.js (feature root), following
  authn's src/recommendations/track.js pattern: trackRecommendationsViewed({ courseKey, isError, length }) owns the sendTrackEvent call and the
  FAILED/LOADED status-string mapping — the one place the legacy status strings
  are still needed, purely for analytics continuity (the event reports the same
  recommendations_status values as before). Consequence:
  CourseRecommendations.jsx no longer imports sendTrackEvent or
  @src/constants. Chose this over LD's heavier central src/tracking/ +
  useCourseTrackingEvent machinery — overkill for one event.
• courseId still read from the courseware Redux slice via useSelector —
  that slice isn't part of this conversion, and reading a not-yet-converted slice
  is expected during the incremental effort. Dropped useDispatch, the fetch
  useEffect, and useModel('coursewareMeta').recommendations (the query owns the
  data now).
• Used recommendations.length directly (dropped the recommendationsLength
  local and the old recommendations ? … : 0 guard): the query's data = []
  default guarantees an array, so the guard was dead code.

Data layer

• Deleted data/slice.js (the recommendations slice) and removed its reducer
  from store.ts and setupTest.js's initializeTestStore.
• Deleted data/thunks.js entirely. After removing the recommendations thunk,
  its only remaining export was unsubscribeFromGoalReminders — a redundant
  wrapper around postUnsubscribeFromGoalReminders (already in api.js), and
  CourseExit.jsx only passed courseId. So CourseExit.jsx now calls
  postUnsubscribeFromGoalReminders directly from data/api.js. The unsubscribe
  test asserts on the POST, so behavior is unchanged.
• Reused data/api.js getCourseRecommendations unchanged as the queryFn
  (it fetches recommendations + enrollments and filters). Pact/api tests keep
  testing it.

Out of scope

• This PR converts only the recommendations read (useQuery). The
  useMutation pattern and the rest of learning's Redux surface are converted
  separately under the wider effort tracked in Convert Learning from redux to Context + react-query #1946.

Test plan

• npm run lint, npm run types, npm run build — clean.
• npm test — full suite green (100 suites / 868 passed / 3 skipped), including CourseExit.test.jsx's recommendations coverage: success→table (with already-enrolled / same-course filtering), and < 2/error→catalog-suggestion fallback.
• Manual, against a live Tutor + discovery backend: on the course celebration page, the recommendations table renders from a real discovery course_recommendations response through the converted useQuery; also verified the fallback path when discovery returns < 2 / errors.

Refs #1946

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.55%. Comparing base (c694ac6) to head (1604f36).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1967      +/-   ##
==========================================
+ Coverage   91.48%   91.55%   +0.06%     
==========================================
  Files         353      354       +1     
  Lines        5838     5824      -14     
  Branches     1356     1391      +35     
==========================================
- Hits         5341     5332       -9     
+ Misses        478      473       -5     
  Partials       19       19              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.