I understand your concerns, but still think this would be feasible:
The NPCs aren't being controlled by the server, they're being controlled by the NPC superclient (NPCClient). That one currently runs on the same machine as the server, but it doesn't have to. Therefore, the server wouldn't have to do a single additional computation (except for distributing the NPC-generated lecture).
Also, as said, the NPCClient would only need to process combat stance changes, and possibly hit results. These already don't get sent to clients beyond a certain range, so the NPCClient wouldn't see a lot of them, especially not if the area is supposed to be a safe one (so if there's a fight, it'd be the exception anyway).
This leaves, however, the messages sent to non-guard NPCs, like rats. Right now, NPCClient likely just drops all fight messages (but already has to check against the list of what to drop), whereas in this case it'd have to filter them by NPC. Still, given that selective dropping by NPC wouldn't require that much processing (if (NPC is not a guard) then drop message), that NPCClient is designed to be independent, that NPCs are still perfectly fine with a response time in the decisecond range, plus also expecting an increase in computing power in the future, and knowing that at least last year the server's CPU was stated to be barely utilised (most lag is network or client (graphics) related), there doesn't seem to be that much of a problem.
Additionally, the NPCs will have to react to fighting and a lot of other player messages in the future, so that they can for example flee, or help each other, etc.. Also, the NPCs already react to the player's proximity, so they have to process the movement messages already, which still are much heavier traffic than combat messages.
In fact, this reminds me of the time when the NPC dialog system had been redesigned to work via normal /say, and require the NPCs to look for their name in the beginning of the sentence. Lower volume, but similar idea, and no speed concerns were raised.