plotly failures with dev rlang

The failure is covered by tests/testthat/test-ggplot-legend.R, and guide titles are assembled in ggplotly.R around lines 1071-1072. Start by running the failing test with the dev rlang version and inspect how guide order is determined. Done means the legend-title test passes with deterministic ordering under the updated rlang.

I'm seeing failures for plotly with dev rlang:

  Running ‘testthat.R’
Running the tests in ‘tests/testthat.R’ failed.
Complete output:
  > library("testthat")
  > library("plotly")
  Loading required package: ggplot2
  
  Attaching package: 'plotly'
  
  The following object is masked from 'package:ggplot2':
...
  Expected `legend_title` to match regexp "factor\\(cyl\\)$".
  Actual text:
  ✖ │ factor(cyl)<br />factor(vs)
  
  [ FAIL 2 | WARN 23 | SKIP 63 | PASS 1449 ]
  Deleting unused snapshots: 'ggplot-contour/contour.svg' and
  'ggplot-heatmap/heatmap.svg'
  Error:
  ! Test failures.
  Execution halted

Claude traced this to rlang::hash() producing different hashes in the new version (which was deliberate to fix other stability issues see https://github.com/r-lib/rlang/issues/1681#issuecomment-4267291605):

The chain is:

  1. rlang dev changed hash() values (commit 229db8154 — "Stable hash()"). The NEWS says: "with this version all hash values will now be different". This is intentional for cross-R-version
  stability.
  2. ggplot2's Guides$merge sorts by hash. In Guides$merge(), each guide gets a key {order}_{hash} and then split(pairs, hashes) is called. R's split() returns groups in alphabetical order
  of factor levels. So the final guide order depends on the alphabetical sort of the hash strings.
  3. plotly concatenates guide titles in order. At ggplotly.R:1071-1072:
  legendTitles <- compact(lapply(gdefs, function(g) if (inherits(g, "legend")) g$title else NULL))
  legendTitle <- paste(legendTitles, collapse = br())
  4. The plotly test asserts a specific order (test-ggplot-legend.R:32-33):
  expect_match(legend_title, "^factor\\(vs\\)")   # vs first
  expect_match(legend_title, "factor\\(cyl\\)$")  # cyl last

  With CRAN rlang, the hashes sort as 99_124c... (vs) < 99_4d3... (cyl) → title = factor(vs)<br />factor(cyl).
  With dev rlang, the hashes sort as 99_aefe... (cyl) < 99_ebf7... (vs) → title = factor(cyl)<br />factor(vs).

I've talked to the ggplot2 maintainers and they recommend using the order argument to make guide placement deterministic. Claude also suggests this fix (unverified):

diff --git a/tests/testthat/test-ggplot-legend.R b/tests/testthat/test-ggplot-legend.R
index 9dd92ea..d1d97e5 100644
--- a/tests/testthat/test-ggplot-legend.R
+++ b/tests/testthat/test-ggplot-legend.R
@@ -29,8 +29,8 @@ test_that("Discrete colour and shape get merged into one legend", {
     nms, paste0("(", d$vs, ",", d$cyl, ")")
   )
   legend_title <- info$layout$legend$title$text
-  expect_match(legend_title, "^factor\\(vs\\)")
-  expect_match(legend_title, "factor\\(cyl\\)$")
+  expect_match(legend_title, "factor\\(vs\\)")
+  expect_match(legend_title, "factor\\(cyl\\)")
 })

As a heads up, I'm planning to release this rlang update within the next two weeks.