Range training on factor with NAs

[teunbrand]

This issue is a port of tidyverse/ggplot2#5224.

The following behaviour is unexpected, where the order of items in the range is rearranged.

library(scales)

f <- factor(c("A", "B", "C", NA), levels = c("C", "B", "A"))
range <- DiscreteRange$new()

range$train(na.omit(f))
# This is fine
range$range
#> [1] "C" "B" "A"

range$train(f)
# This is unexpected
range$range
#> [1] "A" "B" "C"

If we flip the order in which the range is trained, everything seems to work as expected:

range$reset()
range$train(f)
range$range
#> [1] "C" "B" "A" NA

range$train(na.omit(f))
range$range
#> [1] "C" "B" "A" NA

^{Created on 2023-03-28 with reprex v2.0.2}