People tend to get carried away in a heat of discussion, but eventually return to the mean of civility after cool-down period.
Other places just impose ban frivolously, which doesn't help long term, bans destroy the community, methinks.
Moderate moderation like yours is the way to go, and I'm not saying it as a kiss ass.
Edit: dang, what moderation tools do you think could be helpful. What part of your job can be automated via ML/NLP?
What is your least favorite, repetitive or time consuming manual algorithm as a mod?
Probably looking through the flagged comments is the least favorite, as well as most repetitive and time consuming manual activity, for all the mods who do it.
And if this ML labeling is successful then do all the unflagging, or whatever is the most easily automated , most frequent action, to reduce the queue for manual processing.