Testing Strategy

Commands are in backend/CLAUDE.md, frontend/CLAUDE.md, and the e2e/ section of the root CLAUDE.md.

Test levels

Level	Tool	What it covers
Backend unit	JUnit 5	The neighbour policy in both modes: edges, corners, non-desk cells
Backend integration	JUnit 5 + Testcontainers PostgreSQL	Booking and cancel rules, every error code, constraints, authorization
Backend concurrency	JUnit 5 + Testcontainers	The race fix (recipe below)
Frontend unit	Vitest + Testing Library + MSW	Store rules, cell rendering, optimistic booking and rollback
End to end	Playwright, in the top-level e2e/ package	Critical flows across the real backend, frontend, and database

Integration and concurrency tests never use H2. Its locking behaves differently from PostgreSQL's.

Concurrency test recipe

The race fix isn't done until this passes repeatedly against real PostgreSQL.

1. Seed a floor, a set of users, and a target date.
2. Create N threads (for example 20), each with its own user.
3. Hold every thread at a CountDownLatch start gate, then release them together.
4. Scenario A, same desk: every thread books the same desk.
5. Scenario B, neighbouring desks: threads book a desk and its neighbours.
6. Assert:
   • Scenario A: exactly one booking exists, and every other request got 409 DESK_TAKEN.
   • Scenario B: no two booked desks on the floor and date are neighbours. Every rejected request got a 409, and there were no 500s or deadlocks.
7. Repeat each scenario many times (for example 50) with @RepeatedTest. One pass proves little.

Scenario B is the real write-skew race. Scenario A alone would pass with just a unique constraint.

Backend integration cases

• Every error code in Domain Rules, each with its HTTP status and body.
• Both neighbour modes, and a desk surrounded only by walkways.
• Cancelling someone else's booking → 403. Unauthenticated requests → 401.
• The WebSocket publishes after commit, and a rolled-back booking publishes nothing.

Frontend cases

• Socket updates arriving out of order or twice: the older seq is ignored.
• After a reconnect, the snapshot reloads and the map is correct.
• An optimistic booking that gets a 409 rolls back and shows the reason.
• A single desk update re-renders only the affected cells.
• Keyboard navigation and aria-live announcements.

End-to-end cases

• Two browser contexts race for neighbouring desks. One wins, and the other sees the conflict message and the updated map.
• Booking in one browser appears in the other within moments.
• Cancelling frees the desk in both browsers.

Test quality rules

• Tests are deterministic: no sleep waits, no real clock or network dependence, and data is seeded explicitly.
• A test for a bug fails before the fix and passes after it.
• Flaky tests get their cause fixed. Don't add retries or skip them.