One way to mitigate the model's slowness could be to trigger this only on edits done by non-trusted users like IPs and non-autoconfirmed users. The other piece is that models rely on rev_id to look up features (content added, content removed, etc.), with pre-save models, you have to send all of them over the wire. That might be doable to use stash edit API and open it up to lift wing models somehow. @Krinkle might be interested.

Revert Risk Language Agnostic should do the work, the inference time goes below 200ms.