The "all curve" sampling is implemented as two functions internally. The first finds which curve each "global" sample should be on. Then the second is the regular evaluation and sampling in that curve. The first operations creates lengths, but they were processed as factors when passed to the second function.