I love this project and while I'm not an expert in any of this stuff, you may want to take a cue from nature and design some components to be able to deform (flex) without being damaged.  However, you could take more engineered approach would be to have a stress relief mechanism that would trigger if "too much" force is applied.

It is important to keep in mind that evolution has made leaping quadrupeds utilizing far weaker materials.

The problem with flexible parts in robots is all about control. Yes, it could be beneficial to have flexibles in limbs and joints. But non-rigid mechanical systems (where you can't assume perfect rigidity as a good enough approximation) are much harder to control - these are "higher-order systems" - flexibles add unactuated degrees of freedom to robots. Rigid body control theory is much better understood, more established and much less compute-intensive.

They don't have to be highly flexible, I'm just suggesting it as an alternative to being rigid until it snaps but I understand your point that it increases the system complexity and should be accounted for in the software.  Running a small neural network on bare metal should be able to keep the computing to a minimum while enabling sophisticated control and feedback, more like a biological system.  However, the design doesn't strike me as trying to minimize the computing requirements considering a full-blown SBC running Linux is being utilized.  Not sure about your plans, so perhaps that will change in the future.

No matter your approach, I hope you find a way to remedy Stanley's tendency to break into pieces because I really like this project.