fix: reject a buffered response when the request was aborted in flight

[bengeorge]

Closes #232.

Problem

fetch() can resolve as a normal success after its AbortSignal has fired. Cronet's UrlRequest.cancel() and URLSessionTask.cancel() are best-effort and are documented no-ops once the request has completed, so when the cancel loses that race nitroFetchRaw hands back a response the caller has already abandoned — which then goes on to parse a body it discarded. nitroStreamFetch already checks signal?.aborted in several places; only the buffered path is missing it.

Full diagnosis and the device measurement are in #232. In short: on a Fire TV stick, 4 of 4 superseded search requests completed and were JSON-parsed with signal.aborted === true; with this change, 0 of 4.

Fix

  const res: NitroResponseNative = await client.request(req);
+ if (signal?.aborted) {
+   throw createAbortError();
+ }

Placed before NetworkInspector._recordEnd so an aborted request falls into the existing catch and is recorded as aborted, rather than as a 200 the caller never received.

The existing abort mid-flight cancels a slow request harness case passes without this change — aborting 100 ms into /delay/20 lets the native cancel win comfortably — so it never covered the window where the response is already buffered. Added a case alongside it that blocks the JS thread to hold that window open.

Testing

• bun test — 10 pass, 2 existing todos, 0 fail
• bun typecheck — clean
• eslint "**/*.{js,ts,tsx}" — 0 errors, the 2 existing no-shadow warnings in the Expo plugin files. Run under node; bun lint hits an unrelated structuredClone is not defined in @eslint/config-array on macOS that CI on Linux does not
• tsc --noEmit in example/ — the new harness case adds no errors; the 4 in that file are identical to those on main
• The new harness case itself is not run — test:harness needs the example app built onto a device, which I have not done

[vercel]

@bengeorge is attempting to deploy a commit to the Margelo Team on Vercel.

A member of the Team first needs to authorize it.