(a) You don't need to observe confounders to learn structure from data. In fact, sometimes you don't need any standard conditional independence at all. (Luke gave me the impression SI wasn't very interested in that point -- maybe it should be).

(b) Occam's razor / faithfulness gives you enough to learn the structure of statistical models, not causal ones. You need additional assumptions to equate the statistical models you learn with causal models. Bayesian networks are not causal models. Causality is not about conditional independence, it is about counterfactual invariance, that is causality expresses what changes or stays the same after a hypothetical 'wiggle.'

There is no guarantee that even given Occam's razor and faithfulness being true that the graph you obtain is such that if I wiggle a parent, the child will change. To verify your causal assumptions, you have to run an experiment, or no scientist will believe your graph is causal. This is what real causal discovery papers do, for example:

http://www.sciencemag.org/content/308/5721/523.abstract

Here they learned a protein signaling network, but then implemented an experiment where they changed the protein level of a parent via an RNA molecule, and verified that the child changed, but parent of a parent did not change.

I am sure you can set up a Bayesian story for this entire enterprise, if you wanted. But, firstly, this Bayesian story would not be expressed purely in probability theory but in the language that can express counterfactual invariance and talk about experiments (for example language of potential outcomes or do(.)). And secondly, giving something a Bayesian story is sort of equivalent to re-expressing some complicated program as a vi macro. Could be done (vi is turing-complete!) but why? People don't write practical code in vi macros.