[LHagfoss]

What

Two more causes for a turn the user does not have to restart by hand, in the
same code path as the usage-limit wait (state/limitResume.ts).

• Provider overload (feat(desktop): retry a turn after a provider overload and continue after a restart mid-turn #3388): a turn that stopped on HTTP 529/503 or
  overloaded_error is tried again after a doubling backoff — 30 s, 1 m, 2 m,
  4 m — with up to a quarter of it added at random so an account's threads do
  not wake together. Never a fifth time, and a credit or spend limit is never
  retried. The classifier, the backoff and the attempt cap are pure
  (overloadStop, overloadDelay, nextWait).
• A restart mid-turn (feat(desktop): retry a turn after a provider overload and continue after a restart mid-turn #3388): already shipped — store.ts resumeInterrupted
  sends "Continue where you left off" to the threads a quit, an update or an
  engine restart cut off. It is reused as it stands; nothing was duplicated.
• No network (feat(desktop): resume threads that failed because the internet dropped #3389): a turn that failed because the machine had no network
  (Claude Code "Can't reach the API server", Codex's dropped stream, Pi's
  "fetch failed"; classified in state/turnError.ts networkFailure, never one
  a provider answered) waits for the connection instead of dying. While any
  thread waits the app asks the webview every 5 s, and at once on online or a
  wake, then sends "Please continue." once the connection has stayed up for
  5 s, 30 s, 2 m and then 5 m on later retries, so flaky Wi-Fi cannot burn them.
  After 6 h it gives up and says so. The schedule is pure (state/internetWait.ts).

Both waits show the same row over the composer, now with Send now and
Cancel, in the existing type and colours. Both hold queued messages, both
end when the thread moves to another harness or account or when the user sends
anything, and both survive an app restart. Settings › General's row is renamed
"Continue threads on their own", since it now covers all three causes.

Changelog

A turn a provider overload stopped retries itself with a doubling backoff, and a turn the network killed waits for the connection instead of failing for good.

Testing

• bun run check in apps/desktop: typecheck, lint, 2068 tests, build — all pass.
• New unit tests: the overload classifier per harness message (and that credit
  and spend limits never retry), backoff and jitter bounds, the 4-attempt cap, a
  stop that changes kind, the network classifier (positive per harness, negative
  for 401/429/503/529, a limit and a spend limit), the stable-period escalation
  (5 s, 30 s, 2 m, 5 m) and the 6 h cap.
• New store tests: an overload error creates the wait with a ~30 s backoff and
  sends the resume turn when it falls due; a credit limit creates no wait; a
  network error waits, and sends "Please continue." only once the connection has
  been up 5 s.
• No app, browser or mock window was launched.

Closes #3388
Closes #3389