This is a great write up of the problem space. I've also tried an approach here that converted the tile dependencies into a Boolean constraint satisfaction problem and then used the Open Source clasp (https://potassco.org/clasp/) answer set solver to return valid tilings. The inspiration was from the paper "Answer Set Programming for Procedural Content Generation: A Design Space Approach" (https://adamsmith.as/papers/tciaig-asp4pcg.pdf) which is also a good read.
Using an answer set solver was nice because it was so easy, I just had to encode the tile constraints and then the solver did all the work, backtracking, etc, but it could be slow, and it could also fail to return (infinite loop). I gave up because it seemed like WFC and similar could return results fast enough for "online" generation, like generating chunks in Minecraft and also because it seemed hard to encode tile probabilities (from some initial example map) like WFC does.
Using an answer set solver was nice because it was so easy, I just had to encode the tile constraints and then the solver did all the work, backtracking, etc, but it could be slow, and it could also fail to return (infinite loop). I gave up because it seemed like WFC and similar could return results fast enough for "online" generation, like generating chunks in Minecraft and also because it seemed hard to encode tile probabilities (from some initial example map) like WFC does.