This is the "works in a textbook" take.

Being able to plot a series of inputs that can more efficiently use available traction than a human doesn't prevent you from blundering your way into a dumb situation where the laws of physics dictate that the only possible outcomes are various flavors of bad ones.

It's not clear how often the software will chose poorly and need to brute force its way out with traction/handling. The fact that they seem to be hedging against this by putting the hardware on particularly performant cars indicates it must happen enough to matter or be rare but bad enough to matter when it does happen.

Waymo will probably also rack up a ton of technically not at fault accidents by being obtuse in traffic since there's when there's snow there's a lot less margin for the "two people trying to pass each other in a hallway" type missteps that behavior tends to create.