After all the changes made on the modeling side, there is now better separation in scores between TPs and FPs.

This PR experiments with setting a hard-coded score threshold > 0.5 to determine which alerts to surface, rather than always selecting the max-likelihood class, thereby reducing FPs.

Replaces #7512, because it turns out I need to update the worse rather than current libraries, since the end-to-end evaluation runs the boosted queries on the worse modeling.