jakeberv · Oct 17, 2026 · Oct 17, 2026 · Oct 17, 2026 · Oct 17, 2026 · Oct 17, 2026
diff --git a/cran-comments.md b/cran-comments.md
@@ -1,100 +1,40 @@
-# CRAN submission preparation: bifrost 0.2.2
+# bifrost 0.2.2
 
-Draft for an update from CRAN version 0.2.0 to 0.2.2. Version 0.2.1 was
-published on GitHub. Version 0.2.2 has not yet been submitted or published.
+This updates CRAN version 0.2.0; version 0.2.1 was released on GitHub only.
 
-## Release changes
+## Changes
 
-Searches now default to data-dependent BM/BMM starting values adapted from
-the mvMORPH 1.2.1 development initializer used in earlier analyses. The
-compatibility helper passes these values through `start` to the installed
-mvMORPH; it does not pin or modify that package. Native initialization remains
-available through `start_strategy = "native"` and is required for EmpBayes.
-The adapted code credits Julien Clavel in its source and in `Authors@R`.
-`corpcor` and `RhpcBLASctl` are now explicit imports.
+This release restores the data-dependent starting-value strategy used in
+earlier analyses, while retaining an option for native mvMORPH initialization.
+It also fixes progress-related random-state and numerical-thread handling.
+The adapted starting-value code credits Julien Clavel. New imports are
+`corpcor` for initialization and `RhpcBLASctl` for worker thread control.
+Further changes and reproducibility limitations are described in NEWS.md.
 
-This update also corrects progress-related random-stream and numerical-thread
-handling and refines the console displays. It includes the citation and source
-archive updates from the GitHub-only 0.2.1 release. NEWS.md describes the
-changed defaults and reproducibility limitations.
+## Checks
 
-## Package contents and network access
-
-Worked articles and empirical data remain on the package website and are
-excluded from the source package. `bifrost_example_file()` downloads data
-only on explicit request, verifies SHA-256 and byte size, and reports download
-failures. Installation, attachment, and examples do not require these downloads.
-
-## Verified CI results
-
-The following results apply to main commit
-`30e58e82af46f6f0ef35ebf69f3a563f8b459774` (2026-10-15 UTC):
-
-- [R CMD check](https://github.com/jakeberv/bifrost/actions/runs/38688925126)
-  passed on macOS and Windows R release, and Ubuntu R release, devel, and
-  oldrel-1, using `--as-cran --run-donttest --no-manual`.
-- [Coverage and parallel smoke tests](https://github.com/jakeberv/bifrost/actions/runs/38688925143)
-  passed. Whole-package R coverage was exactly 100%, with zero uncovered rows.
-  The coverage run recorded 4,947 passing assertions, zero failures or warnings,
-  and 43 expected skips for coverage-specific cases or development files absent
-  from the built package. Separate Linux and Windows parallel smoke jobs passed.
-- [The pkgdown build and deployment](https://github.com/jakeberv/bifrost/actions/runs/38688925130)
-  passed.
-- [The advisory audit](https://github.com/jakeberv/bifrost/actions/runs/38688925163)
-  completed. It still reports print/cat usage and documentation/style findings;
-  a successful advisory job does not mean every advisory finding is cleared.
-  URL-liveness and advisory spelling checks were skipped.
-
-## Final source-package check
-
-On 2026-10-15, `bifrost_0.2.2.tar.gz` built from the same commit passed
-`R CMD check --as-cran --run-donttest --library=<isolated library>` with:
+`R CMD check --as-cran --run-donttest`, including PDF and HTML manuals,
+completed on macOS arm64 (R 4.4.2) on 2026-10-17 UTC with:
 
 **0 errors | 0 warnings | 1 note**
 
-The sole note was `unable to verify current time`, from the local environment's
-timestamp check. Installation, loading, examples (including `\donttest`),
-tests, and both PDF and HTML manual checks passed. The test suite recorded
-2,969 passing assertions, zero failures or warnings, and 138 expected skips
-(including CRAN-mode skips and development files absent from the tarball).
-The 101-page PDF manual also passed a separate `qpdf --check`.
-
-Environment: macOS Sequoia 15.7.9, arm64, R 4.4.2, mvMORPH 1.2.3, and OpenBLAS
-0.3.34. Bifrost was installed in an isolated library; other dependencies were
-accessed read-only. Numerical libraries used one thread per worker, with at
-most eight workers. Archive contents were inspected to confirm that local
-outputs, caches, and development tooling were excluded.
-
-Archive SHA-256:
-
-```text
-153dbcf17195ad2c0e727fda3acad1e47c91a4ae67f32f5a609b525b0dbfe92b
-```
-
-The first attempt passed tests and the PDF manual but crashed inside the
-existing V8 8.2.0 installation during the HTML manual check. A standalone V8
-call reproduced the crash without loading bifrost. A private build of V8 8.2.0
-with bundled engine 14.6.202.26 resolved it. The complete check then passed on
-the identical archive, without disabling checks or changing bifrost source.
-Existing installations were unchanged during this validation.
-
-## Known limitations retained in this update
-
-- High-dimensional searches remain sensitive to initialization and random
-  state, and some intermediate fits can fail to converge. The compatibility
-  policy restores a reproducible starting strategy, not guaranteed optimality.
-  Follow-up: https://github.com/jakeberv/bifrost/issues/242.
-- The existing Gumbel KL-divergence bootstrap can discard valid large draws
-  because of overflow/underflow and understate uncertainty. Its correction is
-  deferred and is not part of 0.2.2 preparation.
-  Follow-up: https://github.com/jakeberv/bifrost/issues/231.
-
-## Pending before submission
-
-Merge the check-evidence PR through normal review and confirm CI on its merge
-commit. Select the release commit and verify that its package contents still
-match the checked source; rerun the final check if package contents change.
-These maintainer notes and `tools/releases/0.2.2.md` are excluded from source
-archives. Keep the checksum above tied to the actual checked archive, not an
-unverified rebuild. Remove draft/pending wording only when submission is
-authorized. No tag, GitHub release, or CRAN submission has been made for 0.2.2.
+The sole note was `unable to verify current time`, an environment-related
+timestamp check. Installation, examples and tests passed, with 2,980 passing
+test assertions and no test warnings or failures. The RNG chunk-invariance
+test now exercises two and three chunks using at most two worker processes.
+
+[CI](https://github.com/jakeberv/bifrost/pull/252/checks)
+also checks macOS and Windows R release, and Ubuntu R release, devel and
+oldrel-1 (`--as-cran --run-donttest --no-manual`). A Windows temporary-file
+note was traced to the worker-cancellation test; its worker files now use a
+test-owned temporary directory with scoped cleanup.
+
+## Additional notes
+
+- Worked articles and empirical data are hosted separately. Data downloads
+  occur only on explicit request, not during installation, attachment or checks.
+- High-dimensional fits remain sensitive to initialization and random state;
+  some intermediate fits can fail to converge (issue #242).
+- The existing Gumbel KL-divergence bootstrap can understate uncertainty because
+  of overflow/underflow. Its correction remains deferred to a later release:
+  https://github.com/jakeberv/bifrost/issues/231.
diff --git a/inst/WORDLIST b/inst/WORDLIST
@@ -20,7 +20,6 @@ Caetano
 Carvajal
 Cheng
 Claramunt
-Clavel
 Codecov
 Colab
 Collyer
@@ -29,9 +28,9 @@ Cov
 Cracraft
 Cui
 DF
-DOI
 Dipnoi
 EB
+EmpBayes
 Escarguel
 Evaluable
 Felsenstein
@@ -54,7 +53,6 @@ Interspecific
 Iteratively
 JS
 Jolliffe
-Julien
 KDE
 Kaliontzopoulou
 Kozak
@@ -103,6 +101,7 @@ Pennell
 Phylogenies
 Phytologist
 Precompute
+Preprint
 Probst
 Purvis
 Qiao
@@ -149,12 +148,12 @@ Welch
 XQuartz
 Xie
 Zenodo
+Zenodo's
 Zhaoqi
 Zhou
 Zhu
 Ziebell
 abline
-accuracies
 acknowledgements
 adjustcolor
 adp
@@ -223,6 +222,7 @@ gic
 gmShiny
 grDevices
 grey
+gz
 heterochronous
 hindlimb
 hjust
@@ -231,6 +231,8 @@ htmltools
 ic
 idx
 ij
+initializer
+initializer's
 interpretable
 intersp
 intra
@@ -260,7 +262,10 @@ morphometric
 mtext
 mult
 multisession
+multithreading
+mustWork
 mvMORPH
+mvMORPH's
 mvPGLS
 mvgls
 mvmorph
@@ -269,6 +274,7 @@ ncol
 nodelabels
 nodepath
 nonmonophyletic
+normalizePath
 nph
 nrow
 oma
@@ -277,9 +283,7 @@ overrepresented
 oversubscription
 pGLS
 parallelization
-parallelize
 parallelizes
-parameterizing
 passerine
 pca
 pch
@@ -312,7 +316,6 @@ reclusters
 rect
 renderer
 renormalizes
-replicate's
 reproducibility
 rescale
 rescaled
@@ -345,18 +348,19 @@ tunable
 ultrametric
 ultrametricity
 uncached
-unevaluable
 unflagged
 unname
 unrepresentable
 unscaled
+unshifted
 vapply
 vdots
 vectorize
 vectorized
 vectorizes
 vectorizing
 viridis
+winslash
 xend
 xlab
 xlim

diff --git a/tests/testthat/test-search-progress.R b/tests/testthat/test-search-progress.R
@@ -489,6 +489,11 @@ test_that("search progress stage preserves failure lifecycle and rethrows errors
 })
 
 test_that("animated Future failures cancel remaining work and restore settings", {
+  # Killing a Windows Rscript worker bypasses its normal temporary-file cleanup.
+  # Keep this fault-injection test's worker files in a directory it owns.
+  worker_tmp <- withr::local_tempdir(pattern = "bifrost-cancel-")
+  withr::local_envvar(c(TMPDIR = worker_tmp, TMP = worker_tmp, TEMP = worker_tmp))
+
   thread_vars <- c(
     "OMP_NUM_THREADS",
     "OPENBLAS_NUM_THREADS",
@@ -557,6 +562,21 @@ test_that("animated Future failures cancel remaining work and restore settings",
 })
 
 test_that("animated Future RNG matches ordinary work and preserves parallel chunk invariance", {
+  # Compare two and three real Future chunks on at most two worker processes.
+  # Only cap the backend; retain production chunking and per-item RNG streams.
+  with_future_plan <- .bifrost_search_with_future_plan
+  testthat::local_mocked_bindings(
+    .bifrost_search_with_future_plan = function(workers, work, ...) {
+      with_future_plan(
+        workers = min(workers, 2L),
+        work = function() {
+          testthat::expect_lte(future::nbrOfWorkers(), 2L)
+          work()
+        },
+        ...
+      )
+    }
+  )
   run_random_work <- function(workers) {
     set.seed(42)
     .bifrost_search_future_lapply(
@@ -592,7 +612,10 @@ test_that("animated Future RNG matches ordinary work and preserves parallel chun
   set.seed(42)
   serial_reference <- as.list(stats::runif(6L))
   testthat::expect_identical(run_random_work(1L), serial_reference)
-  testthat::expect_equal(run_random_work(2L), run_random_work(3L))
+  two_chunks <- run_random_work(2L)
+  seed_after <- .Random.seed
+  testthat::expect_identical(run_random_work(3L), two_chunks)
+  testthat::expect_identical(.Random.seed, seed_after)
   testthat::expect_error(
     run_random_work(NA_integer_),
     "`workers` must be a single finite number",

diff --git a/tools/releases/0.2.2.md b/tools/releases/0.2.2.md
@@ -54,9 +54,13 @@ for details. GitHub publication does not update CRAN automatically.
   `\donttest` examples and PDF/HTML manuals: **0 errors, 0 warnings, 1 note**
   (`unable to verify current time`). See `cran-comments.md` for the environment,
   archive checksum, and resolution of the local V8 crash using a private build.
-- [ ] Merge the check-evidence PR through normal review and confirm its CI.
-  Select the exact release commit: **pending**. Confirm package contents match
-  the checked source; rerun the final check if package contents change.
+- [x] Check-evidence PR #251 merged as `e1d2fe3`; its post-merge CI passed.
+- [x] Rebuild with the PR #252 dictionary update and test worker-limit fix;
+  complete the package and manual check with zero test warnings or failures.
+- [x] Preserve the two-versus-three-chunk RNG comparison using at most two real
+  workers; also verify identical final caller RNG state.
+- [ ] Merge PR #252 through normal review and confirm its CI. Select the exact
+  release commit and use the verified submission archive recorded below.
 - [ ] Retain the documented deferral of #231; this preparation does not fix it.
   Keep the initialization/convergence follow-up (#242) and RRphylo dependency
   work (#247) separate from this release.
@@ -70,7 +74,63 @@ for details. GitHub publication does not update CRAN automatically.
 - [ ] Apply the publication-time documentation edits below through a PR.
   Update CRAN availability separately, only after CRAN actually publishes it.
 
-## Checked archive and local evidence
+## Latest check: PR #252 Windows cancellation-test cleanup
+
+Source commit: `87b6d31444424a90f823aebcbcbd291bbf80b5d6`; checked on 2026-10-17 UTC.
+Archive SHA-256: `4a817a3ab4e3f6c98efc01342ebaf8fb0b47b61287a05294d943c596a78a00dc`.
+
+The Windows NOTE came from a test that deliberately cancels an active Future
+worker. A Windows diagnostic reproduced leftover `Rscript…` files after forced
+cancellation and none after cooperative shutdown. Windows R removes those
+scripts during normal exit, which forced termination bypasses. The test now
+launches its workers with temporary paths in a test-owned directory and cleans
+that directory on exit. It still forcibly cancels a real worker and retains all
+eight assertions. Production behavior and the RNG tests are unchanged.
+Temporary CI diagnostic code was removed after gathering this evidence.
+
+`R CMD check --as-cran --run-donttest`, including both manuals, returned
+**0 errors, 0 warnings, 1 note** (`unable to verify current time`). All **2,980
+assertions passed**, with zero test warnings/failures and 138 expected skips.
+The PDF passed `qpdf --check`. The isolated library used existing dependencies
+read-only, two advertised cores, and one numerical thread per worker;
+dependency DESCRIPTION hashes remained unchanged.
+
+Evidence: `output/release-0.2.2-pr252-check/run-yr9i7roc/` and
+`output/windows-temp-investigation/`. The Windows diagnostic is retained in
+[its CI log](https://github.com/jakeberv/bifrost/actions/runs/38998764680/job/117297030452).
+Check the latest [PR #252 CI](https://github.com/jakeberv/bifrost/pull/252/checks)
+before merging; the older archives below are superseded.
+
+## Earlier check: PR #252 dictionary and test worker-limit updates
+
+Source commit: `49e8b20faef72532fcee6a127d625e5da32e963b`; checked on 2026-10-17 UTC.
+Archive SHA-256: `4ad44c0dde81d0ae924c4ef1b524c536ad12cdce7d9e44c142a0a29d0df0beea`.
+
+`R CMD check --as-cran --run-donttest` with both manuals returned **0 errors,
+0 warnings, 1 note** (`unable to verify current time`). Testthat recorded
+**2,980 passing assertions, zero failures, zero warnings**, and 138 expected
+skips. The 101-page PDF also passed `qpdf --check`.
+
+The RNG chunk-invariance test retains two and three real Future chunks, but
+caps its backend at two worker processes. It asserts the worker limit and
+identical results and final caller RNG state. Production fitting code and RNG
+handling are unchanged; no test skips, warning suppression, or coverage
+exclusions were added. The original targeted test failed on the worker warning;
+the fixed test passed. A diagnostic mutation bypassing per-item seeding made
+the fixed test fail as expected, confirming it still detects that RNG regression.
+
+Validation used a fresh isolated library with existing dependencies accessed
+read-only, one numerical thread per worker, and two advertised cores. Existing
+dependency DESCRIPTION hashes were unchanged. Relative to the previous PR #252
+archive, only the corrected test and generated packaging timestamp differ.
+
+Evidence: `output/release-0.2.2-pr252-check/run-hivedoav/`, including the archive,
+`SHA256SUMS`, `final-summary.txt`, `bifrost.Rcheck/00check.log`, and the manual.
+Targeted regression logs: `output/rng-worker-limit-regression/`.
+The earlier warning and its diagnostic evidence remain preserved under
+`output/release-0.2.2-pr252-check/run-jx3msm29/`; that archive is superseded.
+
+## Earlier checked archive and local evidence
 
 Checked archive: `bifrost_0.2.2.tar.gz`, built from `30e58e8` on 2026-10-15.
 SHA-256: `153dbcf17195ad2c0e727fda3acad1e47c91a4ae67f32f5a609b525b0dbfe92b`.