The problem is that the biggest N0f8 not larger than 0.33 is 0.329, whereas the biggest one not larger than 0.3 is 0.298. If you floor with only one decimal digit of precision, the algorithm rightly chooses ever smaller numbers until it reaches a number with one digit that is exactly representable as x/256. (Remember, N0f8 is the type representing all fractions of that form!). In this case, that’s 0.2.
Note the documentation of floor:
floor(x) returns the nearest integral value of the same type as x that is less than or equal to x.