melt(, na.rm=T) doesn't handle NA-columns properly

[mplatzer]

The code

dt <- data.table(id=1:3, x=NA_character_, y=c('a', NA_character_, 'c'))
melt(dt, id.var='id', na.rm=TRUE)

produces in data.table v1.9.4

   id variable value
1:  1        y     a
2:  3        y     c

but in data.table v.1.9.6 incorrectly returns

   id variable value
1:  1       NA     a
2:  3       NA     c

It seems the bug only occurs when melt is called with na.rm=TRUE on a data.table with a column that purely consists of NA-values. Also note, that the bug does not occur, when the NA-column is the last column of the data.table.

[oseiskar]

There is also something seriously wrong with the resulting NA values:

library(data.table)
dt <- melt(data.table(x='a',y=NA_character_,z='c'), id.var=c('x'), na.rm=T)

melt_na <- dt$variable
regular_na <- as.factor(c(NA_character_,'z'))[1]

Both melt_na and regular_na show as

[1] <NA>
Levels: z

but write.table(melt_na) crashes with the error message

Error in .External2(C_writetable, x, file, nrow(x), p, rnames, sep, eol,  : 
  index out of range