This might be a won't fix, but I thought I'd float the balloon anyway.

Consider:

library(dplyr)
data_frame(
  id = c(1,2,3,4,5,6),
  value = c(1L,2L,3L,NA,NA,NA),
  group = c("A","A","A","B","B","B")
) %>%
  group_by(group) %>%
  mutate(value = ifelse(is.na(value),0,value))

Which fails.

Versus:

data_frame(
  id = c(1,2,3,4,5,6),
  value = c(1L,2L,3L,4,NA,NA),
  group = c("A","A","A","B","B","B")
) %>%
  mutate(value = ifelse(is.na(value),0,value)) %>%
  group_by(group)

Which does not fail.

As a matter of linguistic parsing one might not expect different outcomes based on ordering. Is there a reason dplyr can't coerce the returns from the group_by operation to share in the most general common data type (as done elsewhere in R) so that both functional orders yield successful results?