The premise is slightly off because it is treating “Network MIDI” as a system-level service rather than as a device interface. On AROS, applications that want to use CAMD are expected to open a CAMD device. That makes the appropriate abstraction a

.device

, even if the implementation behind that device has to deal with multiple network peers, sessions, transports, and dynamically created endpoints. The fact that Network MIDI is stateful and manages multiple peers does not, by itself, make it a service. A device driver can quite legitimately manage that complexity internally. The important architectural boundary is the interface exposed to applications, not whether the implementation happens to maintain persistent network state.