User-specified mods would 'solve' spam for people if the mod is software and works very fast, so that client wouldn't have to show all those events to user only to drop them later after a human mod intervenes. Human mod is needed for a very high level stuff that's hard to categorize by a machine.

And if we have such fast mod software (let's call it what it is - a spam filter) then we can also apply them on the relay as well, so repetitive reports would pass through the same AI-powered filter as all the other stuff. Don't see why that wouldn't work for spam/abuse at relay level if it works at user level.

If someone's goal isn't spam (achieving a goal by pretending to be a legit thing), but is pure destruction (take down a relay), then it's not spam, it's DDoS, and PoW could help, although it's not a complete solution - LN paywall at the protocol level should work better, bcs a) it would sponsor the relay, not the PoW miner, and b) it would work for both reads and writes.

The second half of your post is about implementation details of a potential spam filter, those are all good points, but I'd say spam-protection professionals are needed here, and AI will be absolutely required to fight chatgpt-generated spam.