control outcome transformations

The extract_* functions do different things when there is an outcome transformation in a formula:

library(tidymodels)
library(important)

tidymodels_prefer()
theme_set(theme_bw())
options(pillar.advice = FALSE, pillar.min_title_chars = Inf)

# In-line function in preprocessor formula

ames <- ames %>% select(Sale_Price, Lot_Area, Neighborhood)

lm_wflow <- workflow(log(Sale_Price) ~ . , linear_reg())
lm_fit <- fit(lm_wflow, ames)

# Different variable names and units
important:::extract_data_original(lm_fit, head(ames)) %>% 
    select(contains("Sale"))
#> # A tibble: 6 × 1
#>   Sale_Price
#>        <int>
#> 1     215000
#> 2     105000
#> 3     172000
#> 4     244000
#> 5     189900
#> 6     195500
important:::extract_data_derived(lm_fit, head(ames)) %>% 
    select(contains("Sale"))
#> # A tibble: 6 × 1
#>   `log(Sale_Price)`
#>               <dbl>
#> 1              12.3
#> 2              11.6
#> 3              12.1
#> 4              12.4
#> 5              12.2
#> 6              12.2

# In-line function in extra formula from add_model()

lm_form_wflow <- 
    workflow() %>% 
    add_formula(Sale_Price~ . ) %>% 
    add_model(linear_reg(), formula = log(Sale_Price) ~ .)

lm_form_fit <- fit(lm_form_wflow, ames)

# Same variable names and units
important:::extract_data_original(lm_form_fit, head(ames)) %>% 
    select(contains("Sale"))
#> # A tibble: 6 × 1
#>   Sale_Price
#>        <int>
#> 1     215000
#> 2     105000
#> 3     172000
#> 4     244000
#> 5     189900
#> 6     195500
important:::extract_data_derived(lm_form_fit, head(ames)) %>% 
    select(contains("Sale"))
#> # A tibble: 6 × 1
#>   Sale_Price
#>        <int>
#> 1     215000
#> 2     105000
#> 3     172000
#> 4     244000
#> 5     189900
#> 6     195500

This may be intended/helpful in some cases but definitely not in others. We should perhaps make an option for this.

Start with the extract_data_original() and extract_data_derived() entry points and reproduce the two workflow examples in the issue. Decide how an option should control outcome transformations in each formula arrangement, then verify that the selected behavior is consistent for both workflows.