Transport plugins deadlock if they write too much to stderr
launch_managed_proxy, via tor_spawn_background, opens a pipe from the child process's stderr, but never reads from the pipe. If the child process writes too much to its stderr, eventually an OS buffer fills up and the child process hangs. This manifests in the tor log as "No running bridges."
Seems like this has always been a problem, but it only showed up recently with Snowflake, which by default logs to stderr and is more chatty than past transports have been. See #25600 (moved). The problem went away when instructing snowflake-client to log to a file instead of to stderr.
Ccing ahf as suggested by arma.