Graph-based crossword representation

[jqdq]

Is your feature request related to a problem? Please describe.

• As the optimization algorithm I would like to know what words conflict with each other.
• As the optimization algorithm I would like to know what words work well together.
• As the optimization algorithm I would like to know the history of specific words
  This would allow us to use SAT solvers (or other knapsack-like solvers) when generating the crossword.

Describe the solution you'd like
Representing a single word as an object with a list of all other words that interact with it (maybe via a field class?). Depending on the type of interaction, these might be considered a good thing or a bad thing.
This would require rewriting the crossword libs.

Describe alternatives you've considered
Random choice, wrappers, simple backtracking with copying, adding a child class.

Additional context
I thought about it when thinking about action representation. It's somewhat similar to how Larch backtracking worked.

[jqdq]

CNF-based SAT solvers could be quite cool here, as during the solving process we can automatically disregard every word that fits into all versions of the crosswods.
The actual process can be mixed -- sometimes it will "crystalize" the crossword by leaving only the best version, and all the other times it will just add more of them.