Instead of tracking the largest-WTN system in each connected component, this PR changes tack:
Hook up three landmarks per component, not one
Try to spread the landmarks around the component's convex hull, by grabbing the systems with the maximum q-axis value, the minimum r-axis value and the minimum s-axis value. Having a landmark roughly behind the target node helps give tighter bounds.
As there's now three landmarks per component, the approx-SP lower bound is now the maximum of those bounds, which most of the time (90-odd %) works out tigether than the static bounds.
I also re-arranged the main processing loop of astar_path_indexes, special-cased neighbours that are the pathfinding target, and ripped out the bits that had become vestigial.
Instead of tracking the largest-WTN system in each connected component, this PR changes tack:
As there's now three landmarks per component, the approx-SP lower bound is now the maximum of those bounds, which most of the time (90-odd %) works out tigether than the static bounds.
I also re-arranged the main processing loop of astar_path_indexes, special-cased neighbours that are the pathfinding target, and ripped out the bits that had become vestigial.