The short answer is it's very difficult. The long answer is 3 parts:
1) With collisions of n bodies, you have n*(n-1)/2 unique pairs to check. That is quadratic complexity, O(n^2). It quickly becomes too many checks, then you need spatial partitioning to reduce checks to only those that could collide. Even those few collision checks can be expensive, depending on the complexity of the shapes. Also a fast moving object may jump over a collider from frame to frame, determining that it should actually have collided is hard.
2) Next the collisions themselves are complex. When 2 objects intersect, you need to compute how they should respond. They need to not penetrate somehow and have appropriate forces applied. The order you do this is tricky! Consider a bunch of stacked boxes: when one is adjusted, it could then intersect others.
3) Lastly, doing it robustly is very hard. Physics systems are error prone. They can gain energy, jitter, tunnel, sink, stick, explode. The more stable and accurate you want it, the slower it gets.
We tried doing simpler versions, hoping they would be good enough for simple needs. We had some limited success, but it was complex and most importantly, while it could solve some simple collision needs, it did not solve the much more common secondary motion needs. Eventually we found that the physics as we have it today is much more generally useful and is very efficient.
There are many libraries that do full physics well: box2d, bullet, chipmunk, jolt, etc. We have a whole animation tool to build and support, so the only reasonable way for us to do physics with colliders well is to leverage libraries like these. They are built by super smart people and are very complex.
Integrating a physics library into our runtimes would add bloat and complexity and tie us strongly to that library. Like particle systems, we feel full physics is better as a game toolkit feature. We don't want to make Spine into its own game toolkit. It's better for Spine to be a library that integrates with game toolkits.