fix(e2e): reset per test in excalidraw.spec

[christianhuening]

The two "flaky" excalidraw specs were deterministic failures, not flakes.

Cause

All three tests bootstrap through /setup, but the file used test.beforeAll(reset) — one TRUNCATE for the whole file. POST /auth/setup returns 410 (auth.setup_closed) once any user exists (setup.rs:79), and SetupPage renders that error and returns without navigating (SetupPage.tsx:22-31).

So tests 2 and 3 sat on /setup until the timeout:

Test timeout of 30000ms exceeded.
Error: locator.click: Test ended.
Call log: - waiting for getByTestId('new-doc')
  > 122 |   await alice.getByTestId("new-doc").click();

They passed on retry only because a retry restarts the worker, which re-runs beforeAll and truncates again. That explains the signature shape exactly:

• 30s on first attempt, 2–7s on retry
• always exactly 2 flaky — the number of tests after the first
• present in every run going back to the last green ones on main in June

Fix

beforeEach(reset). Each test is self-contained (own /setup, distinct email), so per-test reset is safe.

The repo already knew about this hazard: change-password and invite-password call reset() inline in their second test, and create-placement, new-doc-edit-mode and tree-reorder use beforeEach. excalidraw was the only multi-test file calling /setup in every test without re-resetting.

Verification

Docker Desktop isn't running locally, and the convergence test needs collab (which needs toxiproxy, broken on Docker Desktop macOS), so CI is the verifier here — and it reproduces this deterministically. A run reporting 0 flaky instead of 2 flaky is conclusive.

🤖 Generated with Claude Code