which is a bummer... a toposort is cheap, but we need it incrementally here because as new nodes are introduced, we need to compute their inputs on the fly