1) predictive policing is racist by either of your definitions, isn't it?
Predictive policing isn't racist by any definition.
Given an imaginary perfect predictor, one could detect all the potential criminals. Now, it may happen than a part of society (e.g. the pink people) could be over-represented in this group, and algorithm would target them more often. Would it mean that the predictor is racist? In my opinion, it would only mean, that the society itself has some kind of racial bias and the problem is not the predictor's accuracy, but the environment infected with the racial bias towards the pink people.
One could argue that predictive policing would help to sustain this bias and it's probably true, but it doesn't entail that the policing itself is racist. There are many honest and reasonable actions that may inadvertently keep the bias alive. Unless we create a fourth circular definition, I've seen recently in the internet: "Any action/stance that doesn't actively counteract racism is racist". If that's the case, then sure, racism is everywhere.
Anyhow, predictive policing is still a terrible idea, a perfect dystopian future. It should be illegal even if it wasn't sustaining the racial bias (which it probably does in the current form).
2) your third definition doesn't entail any of the two presented by @maxharris. One can easily believe there are many human races (it's a common error) and still not penalize/marginalize based on it and not believe there is any relation of superiority between those (imaginary) races. And I do not believe, one should call such people racists, as they do not harm anyone, they are just mistaken.