Connected to reflector 9559 today, and after a while, thought things were a bit quiet. Discovered that the audio path had become disconnected at the reflector end, and the security system had blocked the node's attempts to reestablish audio.
Worked around this by calling the "kicknode" script to disconnect the control side of the IRLP node, if the audio connection ends with an "rtcp_timeout" (which means either a network problem, or the reflector initiated a disconnection). This also means that if a sysop or admin accidentally uses .disconnect or .kick instead .ikick, the node will eventually be disconnected completely.
However, I'm not 100% satisfied with this solution, because it means that connections with transient minor packet loss may disconnect, rather than recover gracefully. I will monitor and see how important this issue is. At least the node will be completely kicked if the audio path dies, which is better than before.