docs(ui): pin the scope of the socket's server-disconnect retry bound

[lstein]

Summary

Resolves #9542 by option 1 from the issue: keep socket.io's own unbounded reconnection as the app's recovery path, and make the scope of the #9540 retry bound explicit in code and tests.

The five-attempt bound in useSocketIO governs server-initiated disconnects only. When one of those retries fails at the transport level (server restarting, proxy down), socket.io reports it as connect_error rather than disconnect, and socket.connect() has already re-armed the manager's reconnection, so the manager's default backoff (infinite attempts, 1s doubling to a 5s ceiling, ±50% jitter) takes over. Our counter never sees it.

That handoff is deliberate, and this PR says so where the next reader will look:

• Rationale on MAX_SERVER_DISCONNECT_RECONNECTS: the server drops sockets for authorization changes and expects the client back, so reconnection: false or a reconnectionAttempts cap would strand the tab exactly when the server is briefly unreachable after such a change, and would hand the same dead end to every ordinary outage the manager rides out today. Also notes why the two drivers cannot race: a retry here fires once per server disconnect (which needs a connected socket), and Socket.connect() does not open the manager while it is between its own attempts.
• Two tests pin it. A connect_error during a retry is left to the manager (the hook adds no second driver), and the socket is built with manager reconnection unbounded.

No behaviour changes. The thread's other conclusion stands too: raising only reconnectionDelayMax would not slow the manager's 1s/2s/4s opening steps, so a rate-limit, if ever wanted, needs its own specified policy and reproduction (the four-attempts-in-180ms burst has not reproduced with the lockfile's client versions).

QA Instructions

Frontend only. pnpm vitest run src/services/events/useSocketIO.test.tsx (12 pass). Each new test was checked against the mutation it guards: a connect_error handler that retries fails the first, and either reconnection: false or reconnectionAttempts: 5 fails the second, with nothing else affected. All five pnpm lint:* scripts pass.

Checklist

• The PR has a short but descriptive title, suitable for a changelog entry
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable) — in-code rationale; no user-facing docs cover socket recovery

🤖 Generated with Claude Code

https://claude.ai/code/session_014Z8LWHPW4TfpNCtwUo5yr9