As per Discord discussion the current autorouter seems to wire multiple "elements" to the same location. you can see it here below. A probably valid DRC board with overlapping paths going same direction.

This will overcomplicate wiring wasting copper :) For example we might end up in a situation like the following (bottom layer not considered for clarity)

What you usually would like to obtain is something like this:

It will allow for much simpler solutions to high density boards....

An heuristic in pseudocode to create something similar to the desired state would be:

let unsolvedNodes = nodes.sortBy("numOfEdges");
while (unsolvedNodes.length >0) { 
  let curNode = unsolvedNodes.pop();
  let nodesToConnect = curNode.neighbors();
  let all = [curNode, ...nodesToConnect]
  let L = all.length;
  // remove all neighbors and assume solvability
  for ( const n of nodesToConnect) { 
    unsolvedNodes.remove(n)
  }
  // path find matrix calculation
  let paths = Matrix(L * L)
  for (let i=0; i<L; i++) {
    for (let j=i+1; j<L; j++) {
      paths[i][j] = pathFind(all[i], all[j]);
    }
  }
  // calculate MST ( minimum spanning tree )
  res = kruskalMST(paths)
  if ( res is false ) { throw Error }
}

This algorithm as it is should be O(N^2 * O(pathFind)) where N are the connection points/smd pads or nodes in the algo before.

IT should already produce a good pcb ready to be optimized later in the pipeline.

Further optimizations:

• Selecting the pathFind step is crucial but I wouldnt be surprised if a simple A* produces a good starting point. furthermore we could potentially cache weights from previous steps as all[i] is being used as target for all the subsequent nodes.
• instead of calculating full matrix we could heuristically filter out too far (air/linear distance) nodes.

Notes:
The algorithm assumes that each component can be "solved" independently, which is common in certain graph partition or network design problems. I believe PCBs fall into these subsets.
Also A PCB is composed by pads (nodes) and wires (edges) that form a particular case of a forest in graph theory where each tree is non disconnected