In its normal mode as I've established our hypothetical Arithmetic Core there's only a couple KiBs of RAM, shared with code. Whereas its harvard-mode would have a full 128KiB (64Ki-words), which could be excessive for our needs!
At which point we may add a hashtable to the top of data-RAM to allow us to update existing entries in our min-heap once we've found a faster way to reach that intersection, instead of growing this min-heap.
Still... We can do better than Dijkstra's!
4/4 Fin today!