Separate transmute() from mutate(.keep = "none")

[DavisVaughan]

In #6035 we reworked .keep to be more consistent for mutate(). I still stand by this decision in the context of mutate(), but this has changed the behavior of transmute(), since that is currently mutate(.keep = "none"). Here is a good example that shows the change in behavior for both group-vars and modified non-group vars.

library(dplyr)

df <- tibble(
  g1 = 1:3,
  g2 = 1:3,
  x = 1:3,
  y = 4:6
)

gdf <- group_by(df, g1, g2)

transmute(gdf, x = x + 1, z = x + 1, y = y + 1, g1 = g1 + 1)

# CRAN:
# - The column order supplied in `...` is kept.
# - Grouping vars not modified by `...` are kept at the front.

#> # A tibble: 3 × 5
#> # Groups:   g1, g2 [3]
#>      g2     x     z     y    g1
#>   <int> <dbl> <dbl> <dbl> <dbl>
#> 1     1     2     3     5     2
#> 2     2     3     4     6     3
#> 3     3     4     5     7     4

# Current dev:
# The column ordering comes from:
# - Modified columns are altered in place
# - New columns are added at the end

#> # A tibble: 3 × 5
#> # Groups:   g1, g2 [3]
#>      g1    g2     x     y     z
#>   <dbl> <int> <dbl> <dbl> <dbl>
#> 1     2     1     2     5     3
#> 2     3     2     3     6     4
#> 3     4     3     4     7     5

It turns out that many people want transmute() to keep the current CRAN behavior because they use it for mixed selection and mutation and expect a specific column ordering from it.

• See Release dplyr 1.0.8 #6080 (comment)
• See update test for dplyr 1.0.8 privefl/bigsnpr#274 (comment)

We previously believed that transmute() had swapped between these two outputs over various dplyr releases, but this was actually not true, as seen in #6080 (comment). The current CRAN behavior has always been how transmute() works. In light of this, we believe we should retain the CRAN behavior of transmute() for the next dplyr release.

That said, mutate(.keep = "none") should retain its current dev behavior. This is an experimental argument, so changing it should not affect too many users. The dev behavior of .keep = "none" is overall more consistent with the rest of the mutate() options, makes it easier to predict the output when combined with .before and .after, and simplifies the implementation because it means that .keep never affects the column ordering, it is mainly about which columns get dropped (#6035 goes into this in great detail).

So, the action items are:

• Fix transmute() to revert to the CRAN behavior, which requires giving it its own implementation separate from mutate()
• Update the NEWS bullet to only mention the change in .keep
• Separate any comparison of transmute() and .keep = "none" in the documentation, making it clear how those are different