Randomized is brittle

[wasowski]

Sometimes it produces a singleton, sometimes it produces a list. We did this because otherwise flatMap had a surprising semantics (if would fix the first choice, and process the second list with the first value fixed). But this ends up in frequent mistakes that we think we have a sample, but we have a single value. I slowly start to understand that it would be better to have lazy / infinite lists for all results (and killing repeat?) but a different semantics of flatMap from the one of lazyLists. One should l.flatMap (a => f(a)) should roughly give l.flatMap (a => f(a).head) - the problem is that we do not want to make it deterministic in the second result accidentally. A good test case is that coin(e).flatMap ( c => (c, coin(e))) should produce a mulitvariate sample with statistically independent components.

One requirement that would be good to meet is that:

for 
  x <- Randomized.Gaussian (...) // or any other distribution, not dependent on y
  y <- Randomized.Gaussian (...) // or any other distribution, not dependent on x
yield (x,y)

should have the same (statistical) semantics as:

for  
  y <- Randomized.Gaussian (...) // or any other distribution, not dependent on x
  x <- Randomized.Gaussian (...) // or any other distribution, not dependent on y
yield (x,y)

Right now is not equivalent if the constructors produce infinite lists (the first will lock the value of x, the second will lock the value of y).