OK. So feature extraction is a funny thing. It turns out that most of our models use a similar set of features and otherwise share many dependencies. In most cases, reverted, damaging, and goodfaith use the *exact same features*. But when you make a request to ORES that looks like this ...
https://ores.wmflabs.org/v2/scores/enwiki/?models=damaging|reverted&revids=1234
... features will be extracted independently for damaging and reverted. That's a waste. We could save a bunch of time by extracting features for one model and then passing the feature extraction cache to the next.