The question dock e2e tests used seedSessionQuestion which sends a
prompt to a real LLM and waits for it to call the question tool. This
is inherently flaky due to LLM latency and non-determinism.
Add withMockQuestion (mirroring the existing withMockPermission pattern)
that intercepts GET /question and POST /question/*/reply at the
Playwright route level, making the tests fully deterministic.
eb5c67de58