Reset the random streams before every replayed record

[LuShadowX]

random.seed(0) ran once per replay process, so the records shared one stream.
A candidate that changes how many values one function draws shifts every
record after it, and a function nobody touched is reported as different.

Reproduced with two functions, one drawing randoms and one merely formatting
one: adding a single extra draw to the first made the second report 2 of 2
inputs differ. It was never edited.

That is a false positive, which is the one error this tool cannot afford, and
it was reachable by any refactor touching random use. Reseeding per record
costs nothing and hides nothing: a function that stops using randomness still
returns something other than the seeded value, so real changes are still
caught. The regression test is the reproduction.