Fix requests hanging ~2 minutes and disconnect not working on a dead server

[taylorcox75]

Summary

Fixes #254. Reporter's repro: connect over Wi-Fi, switch to mobile so the server becomes unreachable, open torrent detail or Transfer — infinite loading, and the disconnect button doesn't work.

Root cause chain:

1. Only get() retried (withRetry), with 4 attempts (retryAttempts defaults to 3) each carrying the full axios timeout (default 10s), plus 500/1000/1500ms backoff between them — ≈43s per logical GET against a dead server.
2. TanStack Query's own retry: 2 on the torrent/transfer polling queries stacked another 3x on top — ≈129s (~2 min) before any error surfaced, and the polls kept restarting so the UI never settled.
3. ServerManager.disconnect() awaited a logout POST with nothing to cancel it if the server was unreachable, and nothing aborted the in-flight GETs — they landed afterward.

What changed:

• services/api/client.ts
  • withRetry now treats connectionTimeout as a total budget for the whole logical request rather than a per-attempt allowance: each attempt's axios timeout is capped to the remaining budget (with a 1s floor so the last attempt isn't handed ~0ms), and a retry is skipped rather than slept if the backoff alone would blow the deadline. Net effect: a GET against a dead server now fails in ~connectionTimeout, not ~43s.
  • Added a session-scoped AbortController (abortInFlight()), used by default in get/post/postFormData/postUrlEncoded whenever the caller doesn't pass its own signal, and wired into setServer()/clearCookies() so every session teardown (server switch, disconnect) cancels in-flight requests.
  • ERR_CANCELED now gets its own normalized message ("Request canceled.") instead of falling into the generic branch. It was already correctly excluded from retry (isRetriableError) and from RECONNECTABLE_MESSAGES (hooks/useReactiveReconnect.ts) — this just makes the thrown error identifiable rather than reusing the timeout message.
• services/server-manager.ts — disconnect() no longer awaits logout() to completion; it fires it (errors ignored, as before) and calls apiClient.abortInFlight() immediately, so disconnect returns promptly instead of waiting out a hung logout against an unreachable server.
• context/TorrentContext.tsx, context/TransferContext.tsx — set retry: 0 on the two polling queries (refetchInterval 2s/3s). The client already retries/backs off internally; a query-level retry on top of that was the multiplier that turned a bounded client-side failure into ~2 minutes. Did not touch the global default in services/query-client.ts — other one-shot queries rely on it.

Per AGENTS.md scope discipline, app/(tabs)/(torrents)/index.tsx (Task B territory) was not touched.

Test plan

• npx tsc --noEmit — exit 0
• npm test — 78/78 suites, 1140/1140 tests passing (added coverage in tests/services/client.test.ts for the retry budget/floor behavior, ERR_CANCELED not being retried and getting its own message, and abortInFlight() cancelling in-flight get/post/postFormData/postUrlEncoded calls that didn't pass their own signal — including that a caller-supplied signal is left untouched; plus tests/services/server-manager.test.ts coverage that disconnect() doesn't hang on a never-resolving logout)
• npm run lint — 0 errors, 37 warnings (documented baseline)
• npm run format — clean
• Could not run the app (no simulator/device in this environment) — this is a pure logic/timing fix with no UI change, so nothing here needed visual verification, but the owner may still want to confirm the real-device repro (Wi-Fi → airplane/mobile → open Transfer, then tap disconnect) resolves.

🤖 Generated with Claude Code