compute groups on ungrouped data.

[romainfrancois]

related to #5473

library(dplyr, warn.conflicts = FALSE)

df <- tibble(x = c(1,2), y = c(3,4), z = c(5,6))

df %>% 
  dplyr::group_by(y) %>%
  dplyr::group_by(across(any_of(c('y','z')))) %>%
  dplyr::group_vars() 
#> [1] "y" "z"

df %>% 
  dplyr::group_by(y) %>%
  dplyr::group_by(y, z) %>%
  dplyr::group_vars() 
#> [1] "y" "z"

^{Created on 2020-11-09 by the reprex package (v0.3.0.9001)}

On master, only the second one incorporates y. Is there any reason why the implicit mutate() step uses the grouped data instead of the ungrouped data, as suggested by this PR ?

library(dplyr, warn.conflicts = FALSE)

df <- tibble(x = c(1,2), y = c(3,4), z = c(5,6))

df %>% 
  dplyr::group_by(y) %>%
  dplyr::group_by(across(any_of(c('y','z')))) %>%
  dplyr::group_vars() 
#> [1] "z"

df %>% 
  dplyr::group_by(y) %>%
  dplyr::group_by(y, z) %>%
  dplyr::group_vars() 
#> [1] "y" "z"

^{Created on 2020-11-09 by the reprex package (v0.3.0.9001)}

I'll add some tests if this is sound.

[hadley]

Oh yeah, this is surprising! We should make sure to make this prominent in the blog post.