fix(streaming): stop uncaught error when tearing down a connecting socket - #173

Fixes #170

Problem

Closing a WebSocket that is still in the CONNECTING state makes ws abort the handshake and emit error on the next tick (abortHandshake → process.nextTick(emitErrorAndClose, …)). The SDK's teardown paths call socket.removeAllListeners() immediately before socket.close(), so that deferred emit finds zero error listeners — and Node's EventEmitter escalates it to an uncaughtException that kills the process. It lands outside the teardown's own try/catch and outside any caller's try/catch around await transcriber.connect().

Two triggers, both reproduced against assemblyai@4.36.4 / ws@8.21.1:

• Connect timeout → discardPendingSocket(): the timeout fires precisely because the socket is still CONNECTING. Worse than a second unhandleable error: failAttempt() discards the socket before rejecting, and the nextTick queue drains ahead of the promise microtask queue — so the process dies before the caller's catch ever runs, and the timeout rejection is never even observable.
• close() racing connect(): StreamingTranscriber.close() (and the deprecated RealtimeTranscriber.close()) run the identical removeAllListeners() + close() sequence with no guard at all.

Fix

After removeAllListeners(), re-attach a no-op error sink (socket.onerror = () => {}) before calling close(), in all three teardown sites. On ws, assigning onerror registers a real listener, so the deferred emit lands on the sink instead of crashing the process; the failure itself is still reported through the rejected connect() promise. Browser builds are unaffected — the sink assignment is harmless, and a browser close() on a CONNECTING socket is a no-op anyway.

Tests

New tests/unit/streaming-connecting-teardown.test.ts covers all three paths and is red without the fix, green with it. The existing unit-test ws mock could never catch this class of bug — its removeAllListeners() replaces handlers with no-ops instead of removing them — so the tests inject an EventEmitter-backed fake with real ws semantics (error emit with zero listeners throws; onerror assignment registers a listener).

Also verified end-to-end: the issue's repro (1ms connectTimeout, and close() immediately after an un-awaited connect()) crashes the process with the reported stack on the published package, and survives cleanly (exit 0, timeout rejection now reaching the caller's catch) against this branch's build.

🤖 Generated with Claude Code