I agree that would probably be a better idea, though you'd still need to keep it around for cidr thanks to backwards compatibility.

If we do this, this would be a breaking change for both cidr and inet (both support the value tuple). Unfortunately we can't obsolete this feature before removing it - which would have been better - since this isn't a specific API or type we can slap [Obsolete] on - it's a generic type argument we support on NpgsqlDataReader.GetFieldValue. However, I've very rarely seen anyone use this particular feature, so I don't think it would be a big problem.

Also, just so I understand how everything fits together here correctly. The code that turns inet -> (IPAddress, int) is in the ADO.NET layer, and the type mapping code in this PR is more telling EF Core "this is a valid conversion by the way"?

Pretty much. EF's TypeMappingSource is what defines what types are supported at the ADO.NET layer, and which CLR types map to which (string) store types. The type mappings it resolves also provide some extra services, like rendering a SQL literal for a given value, so we can embed literals in e.g. queries.