Oh wow, this is an important question that deserves an answer.
In my experience, you don’t want to key in a linear space like ACES/ACEScg. Keying is all about managing imperfections, but if you key in a linear space the imperfections in darker areas will get amplified when the view LUT is applied.
So what you want to do is key in a video space instead, or maybe a log space like ACEScc. Anything but linear. Probably the closer to your final display, the better. This way the imperfections of the key are evenly distributed across perceptual space, which will look a lot better.
So you convert from ACES/ACEScg to ACEScc or whatever, do the key, and then convert back, right? Not so fast! You also have to do the actual A-over-B composite in non-linear space as well or you’ll get the same problem with the artifacts being more harsh in the dark areas. In After Effects you might sandwich your keyed footage between adjustment layers that go ACES->ACEScc below and then ACEScc->ACES above. Keylight should retain all your 32-bit values.
I came across this on my first linear show in the new AE 7.0 at The Orphanage way back when. The 30 second Keylight demo worked great before we started introducing linear into the equation. The answer was to bounce back to video space for the keying and then back to linear right after.