The "Probablistic Functional Programming" paper was really fun, and it builds on some decades-old observations in mathematics.
You can easily extend the PFP framework to handle Bayes rule. I experimented with this a few years back, and was pleasantly surprised how natural it was:
http://www.randomhacks.net/articles/2007/02/22/bayes-rule-an...
(Or, for those who love details, a PDF: http://www.randomhacks.net/darcs/probability-monads/probabil... )
A Haskell embedding seems easier than adding new syntax to R, though you don't get all the nice algorithms that are already implemented in R.