[This is based on an issue I posted on the Rust development mailing list.]
The documentation of the PCG generators does not state explicit that different seeds generate independent sequences, but the existence of a 128-bit seed implies somehow the idea that the whole seed is meaningful.
The user should be made aware that the second parameter (the constant of the underlying LCG) is almost useless from a mathematical and statistical viewpoint.
Changing constant to an LCG with power-of-two modulus and a constant like the one used in PCG simply adds a constant to the sequence (more precisely, there are two equivalence classes of constants, and in each equivalence class the sequences are identical, except for an additive constant).
The minimal scrambling done by the generator usually cannot cancel this fact, and as a result changing the constant is equivalent to changing the initial state (modulo an additive constant). You can try to run this program:
You can pass any state as first argument, and any two constants as the following two arguments, as long as they are either both even or both odd . The program will set up a second initial state so that the sequences generated by the PRNGs using the two constants as seed are based on almost identical underlying LCG sequences, in spite of having arbitrary, different constants and different initial states. The two streams should be independent, but if you pipe the output in PractRand you'll get immediately
You can also offset one of generator by hundred of iterations, but the sequences are so correlated that the result won't change. If you peek at the state of the two generators you'll see that their difference is constant.
I think the reader should be made aware of the danger. If you start several generators of this kind the state is too small to guarantee that there will be no overlap. Once you get overlap, since there are in practice just two sequences, you will get a lot of unwanted correlation.
There's a reason why nobody in the last decades ever considered creating "streams" using the constant part of an LCG, and it's that people realized very early that it doesn't work (see, e.g., https://ieeexplore.ieee.org/document/718715).