Spam can be filtered effectively client-side with a good spam filter. This has worked well for me for decades without the need for any server-side spam filtering.
Using NLP / LLM spam filtering would presumably be either inaccurate or expensive or both. Someone would have to pay for it.
Bayesian filters are basically just a cheaper / worse version of what an LLM filter would do. Very easy to beat. Especially if the spammer is using an LLM to write a semi-unique email for each recipient.
As I've said, I'm not interested in theoretical arguments. All of my domains wildcard forward to the same email address, too. Filtering client-side has never been a problem.