vctrs can coerce c(NA) logical to character, but not logical(0) to character. Why?

[Ljupch0]

vctrs::vec_c("a", NA) # WORKS
vctrs::vec_c("a", logical(0)) # FAILS

This issue has broken my R prod code more than any other. I would like to understand why.

Usually I have bind_rows combine some dataframes. Sometimes the dataframes have all NA columns that are read-in as logical.
When I bind these dataframes with others that have values in these columns, type conversion works and I don't have a problem. But if it happens that all dataframe rows are deleted before merging, like in a shiny reactive df for instance, then bind_rows fails with message Can't combine ..1<character> and..2 <logical>.

Example:

df_with_na <- tibble(
  column = c(NA, NA, NA)
)

# Column is logical
# class(df_with_na$column)))

print(
  bind_rows(
    df_with_na,
    tibble(
      column = c("a", "b", "c")
    )
  )
)

# Binding works, logical made of NAs is combined with character.

# Rows are removed from the df.
df_without_rows <- df_with_na[0,]


print(class(df_without_rows$column))

bind_rows(
  df_without_rows,
  tibble(
    column = c("a", "b", "c")
  )
)

Knowing that it happens, I can be proactive and drop all NA columns, or skip binding dataframes with 0 rows. But I'm curious why this behavior happens.

Is it that an all NA vector is a "fake" logical, while logical(0) could be a real logical vector comprised of booleans?