I wonder if any games use a landmark approach to bound their A* search.

The basic idea of landmarks is that you choose a small number of landmarks (e.g. four corners of a map) and for every landmark, you pre-compute the distances to every node. You can then use this information to get a lower-bound for the distance to the target via the triangle inequality: d(p,t) >= |d(p,LM) - d(t,LM)|.

This is obviously weaker than the goal bounding explained in the video, but the precomputing only takes O(n). We used it quite successfully in a global router (for EDA), where the heuristic lower-bound is often much better than a standard A-star heuristic based on geometric distances, because edges could have wildly different costs.