You can mimic the owl's animation mixing using preview. For example, set idle to play on track 0, on track 1 enable additive (it's important you do this before setting the animation), set left to play on track 1, and then set the alpha on track 1 between 0 and 1 for how much of left gets applied. The alpha slider is often small, but you can drag left/right or up/down on the text box instead. Note that none of this solves your problem, I just wanted to mention it as an easier/faster way of exploring AnimationState tracks, mixing animations, additive blending, etc.
The behavior you see occurs because of the way the sign of the scale is mixed between animations. Mixing animations, whether between different tracks using alpha or between animations on the same track using mix duration, is a different situation than the interpolation between keys.
Between keys the interpolation between 1 and -1 scale would happen as you expect, the value changing smoothly from 1 to 0 to -1. If that is not desired, you can set the interpolation to stepped so the scale of 1 is held until the -1 key is reached.
When mixing animations and the sign of the scale differs between the animations, we've found it's rare to want the scale to be interpolated smoothly through zero. Doing that causes everything attached to the bone to get very thin as it approaches zero. Usually negative scale is used to have a bone axis point in the opposite direction for the pose in the animation. When mixing poses of two animations, usually you care about which direction the bone is pointed in each pose and don't want the bones flipping over like a sheet of paper.
Unlike when keying, when mixing you don't have the possibility to choose stepped, linear, or Bezier for each keyed property. We don't currently have a way to customize the behavior for the scale sign. The sign is always treated as an instant transition, like a draw order key or slot attachment visibility key.
The above makes sense for typical animation mixing between animations on the same track or across tracks, where the poses from the animations are the final poses you want when fully mixed (when alpha is 1 or the mix duration is reached). However, the owl uses additive mixing. Why would you use negative scale when mixing additively? Is the goal to have a bone axis point the opposite direction or to flip it over like a sheet of paper? I have a feeling it could make sense to interpolate the scale through zero when mixing additively. The relevant code starts here.
There are some other interesting challenges when mixing animations. For example, say the pose from each animation specifies the rotation for a bone and we want a rotation 50% of the way between those. There are two possible answers: the short way between the rotations or the long way between them. AnimationState chooses the short way, however the animations can change over time. The short way now may become the long way in a few frames! Once AnimationState has chosen the short way for a mix between animations, it continues using that same direction for the rest of the mix.
We've updated all 4.0 runtimes so when scale timelines are applied, the sign of the scale is not treated as an instant transition. The owl will now flip over as you expect. This change is also in the Preview view starting in editor version 4.0.32.
