I am excited to see that it can get customized over several dialects.

In the Japanese 5th generation project they thought they could parallelize Prolog but found out early on that Prolog could not be parallelized so they came up with KL1 which could be parallelized but is not as nice as Prolog.

I'd love to have a "language construction set" where I could trade off expressiveness and efficiency and such.

My take on facts and rules is that they are somewhat portable between different regimes. For instance the same set of rules can work well in a forward chained mode as in a RETE rules engine or in a backward chained mode using Prolog or even in an SMT solver for consistency checking.

I call it "rules and schemes" where you reuse the same rules with different execution strategies to solve different inference problems. In fact you want it to be easy to move work between build and run time.