Bun.connect: native write() returns -1 (not raw -errno) on a peer RST and close carries ECONNRESET - #36422

For `Bun.connect({ allowHalfOpen: true })`, this early return relies on "let `on_close` report the error", but with `allow_half_open` set loop.c never auto-closes after `on_end` — it only re-arms WRITABLE — so `on_close` never fires and the promised `close(ECONNRESET)` never arrives. The socket instead spins on level-triggered EPOLLHUP (each poll: WRITABLE→`on_writable` no-op, then eof→`on_end` suppressed→re-arm WRITABLE), and pre-PR the user at least got an `end` hook to call `socket.end()` and

The last two assertions (`code === "ECONNRESET"` and `syscall === "write"`) encode Linux errno behavior; on macOS, `send()` to a peer-RST'd socket returns `EPIPE` without consuming `so_error`, so the close error surfaces as either `{code: "EPIPE", syscall: "write"}` or `{code: "ECONNRESET", syscall: "read"}` — the conjunction cannot hold and the test will fail on macOS CI. Either gate these two lines on `isLinux`, or relax to `expect(["ECONNRESET","EPIPE"]).toContain(code)` / `expect(["read","wr

`fatal_write_errno` is never cleared when a `NewSocket` wrapper is reused for reconnect (`detach_for_reconnect`, `connect_finish`'s `maybe_previous` branch, and the Windows named-pipe `prev` branches all skip it). A node:net socket that saw a peer RST, then reconnects — the code comments cite the MongoDB driver as a real user of this path — will on the *new* connection suppress `end` on a clean FIN (line 1593) and fabricate a spurious `ECONNRESET` / `syscall: 'write'` in `on_close` (line 2068) f