You're right. I think what Tim is getting at is (and he can correct me if I'm wrong) that, sometimes, when folks use all_spam_to, they also want to ensure that those messages don't get autolearned *at all*, regardless of their spaminess or haminess. Traffic from this mailing list is a good example.
Tim, what you want can be done with 3.0 and the bayes_ignore_to directive. It works like a charm. However, you're using 2.64, which doesn't support bayes_ignore_to.
Unfortunately, you're kind of out of luck doing this with 2.64.
True, but as Theo suggested, several tools that call SA can be configured to skip users like that.. Not only do you avoid tagging/learning/whatever but you save CPU time by not even trying to scan.
With procmail you just modify your procmail rules such that the user gets delivered without the rule that calls spamc/spamasassin.
MailScanner can do this quite easily with it's rule lists.
There are others, but I'm not familiar with them, be sure to check if your tool can do this.
If your tools can support it, this is by far a better way than any kind of modification to SA's configuration.
