I like the idea, but it could be that the "edge cases" here are not quite so easy to handle. Part of the way social networks work and grow is in allowing users to discover new people. If you don't think carefully about discovery from that perspective, you might end up killing that feature. Granted, spam is also a big annoyance, so trading off slightly less reach for less spam may be a worthy trade.
Discussion
If you imagine a point cloud (social graph) of users, then the main network would be connected with follows and spam accounts would be atomised satellites.
If the spammers created a fake network of fake accounts and follows then it would be a separate cloud.
If the spammers bridge into the main network, then the main network would respond by unfollowing people who follow spammers until the spammers are ejected from the main network / point cloud.
New users would start as atomised satellites, but finding a human or someone to follow you is fairly trivial because you only have to do it once at first entry.
Spammers would soon give up targeting nostr, and the tolerance level would relax during times of low spam and increase during spam attacks.
But most of all, the whole thing is self policing and self correcting which means it can scale infinitely and is not restricted by moderator resources, or rigid rules.