model.frame creates invalid interlaced vectors

[khusmann]

library(tidyverse)
library(interlacer)

df <- tibble(
  a = c(1,2,3),
  b = interlaced(c(1,-99,3), na=-99),
)

invalid <- model.frame(a ~ b, df)

value_channel(invalid$b)
#> [1] 1 3

# Oops, the na channel hasn't been updated
na_channel(invalid$b)
#> [1]  NA -99  NA

# This will error:
# print(invalid$b)

^{Created on 2024-09-30 with reprex v2.0.2}

[khusmann]

This is because stats::model.frame.default() calls a C function to do the na.action: https://github.com/wch/r-source/blob/b8b2662a848a8eb87d67d66d737b00b6b2b127f2/src/library/stats/R/models.R#L565

This C function drops na values from vectors, but then copies their attributes back on: https://github.com/wch/r-source/blob/b8b2662a848a8eb87d67d66d737b00b6b2b127f2/src/library/stats/src/model.c#L228

Here's where the copying happens: https://github.com/wch/r-source/blob/b8b2662a848a8eb87d67d66d737b00b6b2b127f2/src/main/attrib.c#L302

So this bypasses all of the proper indexing... how annoying! :(

[khusmann]

One way to fix this, could be to detect when the na channel has been corrupted:

na_channel.interlacer_interlaced <- function(x, ...) {
  m <- attr(x, "na_channel_values")
  if (length(m) != length(x)) {
    unspecified(length(x))
  } else {
    m
  }
}

[khusmann]

Actually, what I should do is store a reference to the original data in another attribute. That way, I can detect if the current data changes its reference. If the current data changes its references (i.e. the data has been changed without updating its na_channel), then the na_channel should be invalidated.