Resolve collisions independently, in parallel?

[mbostock]

As an alternative to #24, I should consider resolving collisions independently: aggregate the forces for each overlapping node, and then apply them at the end after determining all the overlaps. I did this before looking at the previous position ⟨x,y⟩ and it was too soft, but maybe it would be sufficient when looking at the next position ⟨x + vx,y + vy⟩?

Also if we do this, the forces should be divided in half, since they will apply symmetrically to each pair of overlapping nodes. And it means we don’t have to mutate the quadtree as we iterate over the nodes, and we could probably compare indexes i < j to determine whether we need to test two nodes for overlap.

[mbostock]

And since we’re not mutating the quadtree, we could compute the maximum radius of each quadrant (rather than the maximum radius of all nodes), potentially allowing for much faster testing.