The shuf implementation refers to the https://en.wikipedia.org/wiki/Fisher–Yates_shuffle algorithm, however it does not seem to actually follow it. This leads to peculiar behaviour. For example, given input file: foo bar baz after shuffling the input file, foo will never end up back on the first line. This came to light when I ran into a use-case where someone was selecting a random line from a file using shuf | head -n 1, and the results on busybox were showing a statistical anomaly (as in, the first line would never ever be picked) vs the same process running on environments that had gnu coreutils installed. On line https://git.busybox.net/busybox/tree/coreutils/shuf.c#n56 it uses r %= i, which will result in 0 <= r < i, while the algorithm specifies 0 <= r <= i.
I did a quick test with shuf.c line 56 changed to: r %= i+1; and it seems to solve the problem.
Fixed in git, thanks!