Return empty data frame from dapply() for empty input

[GuangchuangYu]

Problem

A layer whose data is entirely removed by scale limits aborts the plot build when the plot has more than one panel.

library(ggplot2)

d1 <- data.frame(panel = "a", x = c(1, 2))
d2 <- data.frame(panel = "b", x = c(10, 20))

ggplot() +
  geom_boxplot(data = d1, aes(x = x)) +
  geom_boxplot(data = d2, aes(x = x)) +
  facet_wrap(~panel) +
  xlim(0, 5)
#> Error in `seq_len()`:
#> ! argument must be coercible to non-negative integer

Full message on ggplot2 4.0.3:

Problem while computing stat.
i Error occurred in the 2nd layer.
Caused by error in `seq_len()`:
! argument must be coercible to non-negative integer

xlim(0, 5) drops both rows of the second layer, so by the time the stat runs
the layer has zero rows and the plot has two panels. With a single panel, or
with a partially-censored layer, everything works — which is why this is easy to
hit by accident and hard to diagnose.

Cause

Stat$compute_layer() finishes with dapply(data, "PANEL", ...). For an empty
data, dapply() still tries to split:

ids <- id(grouping_cols, drop = drop)              # integer(0)
group_rows <- split_with_index(seq_len(nrow(df)), ids)

split_with_index() derives its number of groups from max(f); with no rows
that is max(integer(0)), i.e. -Inf, so seq_len(-Inf) throws.

There is a second, latent problem behind it: even if the split succeeded, the
callback does layout$get_scales(data$PANEL[1]) with an empty data, so it
passes NA. get_scales() then selects every row of a multi-row layout
(self$layout$PANEL == NA is NA for each row), and
self$panel_scales_x[[c(NA, NA)]] is a recursive index, giving
! no such index at level 1.

Fix

Zero rows means zero groups, so fun() must not be called at all. dapply()
now short-circuits to an empty data frame before splitting:

if (nrow(df) == 0L) {
  return(data_frame0())
}

This resolves both symptoms, and also removes a pre-existing wart: the old
single-group shortcut ran apply_fun() on an empty df, which produced a
phantom row with PANEL = NA.

Downstream report

Reported as YuLab-SMU/ggtree#658,
where xlim() combined with geom_facet() annotation panels hit this. The
reprex above reproduces it with ggplot2 alone — no add-on packages involved.

Tests

New tests/testthat/test-compat-plyr.R covers:

• normal split/apply/reassemble behaviour;
• empty input returns an empty data frame without calling fun();
• the single-group shortcut still applies to a non-empty single group, but not
  to an empty one;
• the end-to-end reprex above builds.

Full suite: no new failures. The only failing test,
test-zzz.R (".onAttach does not modify the random stream"), fails identically
with and without this patch — it is an artefact of running the suite through
pkgload::load_all(), since .onAttach() is not attached to the search path
that way.