fix(export): close leaked fds and time out hung API requests

[harsh-thakkar7]

Description

Follow-up to the closed fd/temp-file + request-timeout PR, incorporating review feedback from karaaslanz on the request-timeout half. The fd/temp-file cleanup (backend/routers/export.py, tests/test_export_fd_leak.py) is unchanged — the reviewer confirmed that half is well-scoped.

Request timeouts, revised per review

The previous version applied one 30 s default to every apiFetch/apiFetchBlob call, which would have broken legitimately long-running requests:

• AI.chat, AI.work, AI.translate, AI.translateSentences, AI.translateSentenceBatch and provider model listing can exceed 30 s (slow/local models, long documents).
• Conversions such as PDF/HTML → Markdown and DOCX/PDF export can also exceed 30 s for large files.

Changes in frontend/src/lib/api.ts + new frontend/src/lib/http-timeout.mjs:

• Endpoint-specific timeouts. Settings/metadata/file-listing calls keep the short default (DEFAULT_REQUEST_TIMEOUT_MS, 30 s). Long-running AI/conversion/export endpoints pass LONG_REQUEST_TIMEOUT_MS (5 minutes) explicitly.
• Composed abort signals. fetchWithTimeout previously did init.signal ?? controller.signal, which silently dropped the timeout whenever a caller supplied a signal (e.g. translateSentenceBatch(..., signal)). It now composes caller + timeout signals via composeAbortSignals (uses AbortSignal.any, with a manual-wiring fallback for older webviews), so both caller cancellation and the timeout stay effective.
• Explicit no-deadline opt-out. timeoutMs <= 0 disables the automatic abort.
• The helper lives in a plain ESM module (http-timeout.mjs) so the Node-based regression suite can exercise the exact code the app runs.

Regression tests (frontend/tests/api-fetch-timeout.test.mjs, 8 tests)

• A hung request aborts at the configured budget instead of waiting forever.
• A request that outlives the short default succeeds with the long-running budget.
• A caller-supplied AbortSignal aborts the request even when a generous timeout is set.
• The timeout still fires when a caller supplies a signal (the specific regression the reviewer flagged).
• timeoutMs = 0 disables the deadline.
• composeAbortSignals unit cases (single/absent signals, any-input abort, already-aborted input).

Why this matters

Without the per-endpoint budgets, the 30 s default would turn previously-working slow AI/conversion requests into client-side aborts. Without composing the caller signal, cancel-and-timeout would not coexist.

How Has This Been Tested?

• node --test tests/api-fetch-timeout.test.mjs → 8/8 passing.
• frontend/src/lib/api.ts type-checks (transpile clean); the new http-timeout.mjs is dependency-free plain ESM.
• Backend fd/temp-file suite unchanged and previously green (tests/test_export_fd_leak.py).

Related

Supersedes the earlier fd-leak + timeout PR after review. Standalone PR — no matching open issue.

[petertzy]

Thanks for addressing the earlier timeout review and cleaning up the export file descriptors. I found one remaining timeout gap: the deadline was cleared as soon as fetch() received response headers, so a stalled JSON or Blob response body could still hang indefinitely. I fixed this in f76c86f; the deadline now stays active until the API response body has been read. I also made the timeout tests independent of a local HTTP port.

After the fix, all 16 frontend tests, ESLint, TypeScript, Ruff checks on the changed backend files, and the diff check pass.