perf(openai): reduce websocket transcript merge allocations