Description
Currently, a workaround for the length/nrow requirement for the inner computations is to return a single-row data frame with a single list column containing the actual result, then do a time_value-grouped slice and maybe a select to drop the nongrouping key columns on the slide output. This could be slightly less work if one could just say data = list(<real result>) for the computation rather than data = tibble(result = <real result>).
However, we might want to watch out for users trying to achieve something like the %>% mutate(tibble(col1=..., col2=...)) behavior, except with lists instead of tibbles. Perhaps that's enough to dissuade us from implementing this, or maybe there's a way around it; e.g., accepting only unnamed lists, since the names would probably be dropped anyway.
Implementation might require a tweak of as_list_col, not just the validation code.