Rhys: Afr0 showed me your issue. Both your renderer and the Niotso renderer have the issue with the legs with a2o-soc-death-possess-sima.anim. It's not an issue with quaternion interpolation, since yours is using slerp and mine is using n-lerp, and we both get the exact same rendering glitch. I thought that it could be that we're handling the blend vertices correctly.
But we noticed that a2o-soc-death-possess-sima.anim (and probably all the other animations as well) changed slightly all throughout the file between Pre-Alpha and Play Test, where the Play Test version was unchanged into New & Improved. (bsdiff produces a 35.7kB difference from Play Test; bps after compression 55.3kB; xdelta 73.0kB.) The Pre-Alpha version renders without any glitches. Additionally, if you replace a2o-standing-loop.anim in Pre-Alpha with the Play Test version of a2o-soc-death-possess-sima.anim, it renders with the leg glitches there too. (If I can enable the simulator in Play Test in NoNetwork mode, I can probably confirm that the animation is glitched in Play Test as well.)
So I believe a programmer or an artist decided to run an optimizer on all the animations (I looked at a few and they had different sha256 sums), and verified on several of them that it didn't break or reduce the quality of the animations. And then waaaaaay later, somebody noticed "hey guys, this animation is broken", but nobody had the original animation anymore and they either didn't want to fix it (e.g. maybe their only option was to remake the whole animation from scratch and they didn't want to), or didn't know how to anymore (e.g. all the artists were gone).