I don't think this is true for ciphers that are anywhere near as complicated as the physical world is

Right -- I tried to make clear that my concern is only about those phenomena that are less complex that the most complex broken cipher. For unknown-physical-law cases (i.e. we don't even know the dynamics of the phenomenon), the comparison is to ciphers that can be broken even if you don't know which cipher or public key is being used; for unknown-constants cases (where we know the form of the equations), that also includes ciphers requiring knowledge of the cipher and public key to break.

For inferring models of limited complexity on limited numbers of variables in constrained form, however, there're some pretty good algorithms in Pearl's book Causality, based on conditional correlation and independence.

True, but there are broken ciphers that are resistant to a cryptanalytic version of this, which suggests that Pearl isn't covering all the ways to find regularity in data.