Thread names are very useful to quickly diagnose which threads are causing problems (eg. from htop). However, there were always a few threads whose name was missing, which was caused by buggy set_thread_names function. New version uses a counting latch to ensure each thread gets its name assigned, which requires bumping the target macOS version. Generally a thread pool can be implemented by using boost::io_context and a nano::thread_runner which would give us much better control of underlying threads and avoid some code duplication, but that is a TODO.