A small tweak to this approach would be to precompute and cache the entire chunk-graph distance matrix, so the chunk-distance between any two chunks can be computed by a lookup, instead of searching. Use a bit of memory and make the time to generate the world a bit slower, but could be viable if the chunk-graph pathfinding calls are still slowing things down a bit at runtime.
This might make sense if a bunch of assumptions hold: the chunk-graph does not change during gameplay, map sizes aren't much bigger than the one shared in the post so the number of vertices in the chunk graph is pretty small, land units share the same costs for traversing terrain (i.e. same chunk graph with same edge cost shared by all units).
Rough back of the envelope: number of vertices, V = 20 chunks high * 20 chunks wide = 400. Chunk graph is sparse, each chunk connected to 8 neighbours by undirected edges, so number edges, E = (1/2) * 8 * V .
Memory required to store chunk-graph distance matrix: V^2 * sizeof(distance) = 400^2 * sizeof(float) = 640,000 bytes. Could use +inf for disconnected chunks due to islands.
Time to precompute full chunk-graph distance matrix by repeatedly dijkstra-ing from each chunk to all chunks in the chunk graph: V * T(dijkstra(E, V)) = V * O(E + V log V) = O(V E + V^2 log(V)) = O(V^2 + V^2 log(V)) = O(V^2 log(V)) , which won't scale due to the quadratic-ish V^2 log(V) term, but probably fine if V <= 400 .
This might make sense if a bunch of assumptions hold: the chunk-graph does not change during gameplay, map sizes aren't much bigger than the one shared in the post so the number of vertices in the chunk graph is pretty small, land units share the same costs for traversing terrain (i.e. same chunk graph with same edge cost shared by all units).
Rough back of the envelope: number of vertices, V = 20 chunks high * 20 chunks wide = 400. Chunk graph is sparse, each chunk connected to 8 neighbours by undirected edges, so number edges, E = (1/2) * 8 * V .
Memory required to store chunk-graph distance matrix: V^2 * sizeof(distance) = 400^2 * sizeof(float) = 640,000 bytes. Could use +inf for disconnected chunks due to islands.
Time to precompute full chunk-graph distance matrix by repeatedly dijkstra-ing from each chunk to all chunks in the chunk graph: V * T(dijkstra(E, V)) = V * O(E + V log V) = O(V E + V^2 log(V)) = O(V^2 + V^2 log(V)) = O(V^2 log(V)) , which won't scale due to the quadratic-ish V^2 log(V) term, but probably fine if V <= 400 .