SIGNIFICANT USER-VISIBLE CHANGES:

• The R citation now provides a DOI (10.32614/R.manuals): please update
  your bibliographies accordingly.

• The default C++ standard has been changed to C++20 where available
  (which it is on all known platforms from 2021 on): if not C++17 is used
  if available otherwise C++ is not supported (as before). (GCC 16 has also
  switched to C++20 as its default.) Packages can request C++17 if
  essential.

• The Sweave driver Rtangle no longer puts empty lines at the end of
  tangled scripts.

NEW FEATURES:

• The default AIC and BIC methods now check whether the methods namespace
  is loaded rather than stats4. They still use stats4::logLik (and
  stats4::nobs), but packages no longer need to load stats4 explicitly.

• confint.default() now also works on S4 objects, as long as suitable
  coef() and vcov() methods exist.

• binomial(identity) and quasibinomial(identity) now work without
  having to quote the argument.

• str() now shows more via format() for external pointers.

• terms(<formula>, specials = *) now treats non-syntactic specials more
  gracefully, thanks to Mikael Jagan's PR#18568.

• help.start() gets a lib.loc option and its default for
  make.packages.html() is enhanced to allow, e.g., fast access to
  .Library when that contains only base and recommended packages.

• In *.Rd help files, the \linkS4class macro now also works for links to
  other packages via \linkS4class[<pkg>]{<myClass>}, much nicer than the
  “workaround” \link[<pkg>:<myClass>-class]{<myClass>} necessary in earlier
  versions of R. Packages with the new syntax need to formally depend on R >= 4.6.0.

• tools::analyze_license() now also computes SPDX license identifiers,
  thanks to Thierry Onkelinx and LLuís Revilla.

• read.dcf() now recognizes lines starting with # as comment lines. By
  Dirk Eddelbuettel, Laurent Gatto and Hugo Gruson.

• tcltk::tclVersion() now only optionally includes the ‘patch level’, as
  Tcl full versions can be of the form 8.7a5 or 9.0b1 without three clear
  components.

• untar(tar = "internal") supports extras = "-P" to use unchanged the
  recorded file paths (as many external tar programs do).

• extSoftVersion()[["zstd"]] now reports the version of the zstd
  compression library if available, thanks to Trevor Davis' patch proposal
  in PR#18914.

• New logical option() quiet, defaulting to false, set to TRUE by R's
  command-line option --quiet (and similar), now can be switched during
  an R session. Prompted by Dirk Eddelbuettel's proposal in PR#18913.

• Credentials for basic HTTP authentication with download method "libcurl"
  can now be provided in a netrc file. The path to the file can be
  specified in options("netrc").

• tools::deparseLatex() gains a math argument, now used by the default
  bibstyle to format bibentry() fields with $...$ syntax via Rd \eqn
  markup.

• gzcon() can now decompress a concatenation of compressed streams PR#18887.

• Bibliographic citations and references in Rd files can now be
  auto-generated from bibentries in bibliographic databases in R or BibTeX
  formats. See RShowDoc("R-exts#Bibliographic-citations-and-references")
  for more information.

• Arguments to loadNamespace() which will be ignored because the
  namespace is already loaded are now noted.

• chkDots()' optional argument allowed has been implemented thanks to
  Trevor Davis' proposal in PR#18936.

• New function messageCondition() to create structured "message" objects
  analogous to warningCondition() and errorCondition() (PR#18940, by
  Trevor Davis).

• wilcox.test() can now perform exact (conditional) inference in case of
  ties. Based on contributions by Torsten Hothorn.

• wilcox.test() can now optionally compute improved asymptotic p-values
  by including up to 3 correction terms of the Edgeworth series for the
  normal approximation.

• barplot(.., orderH = *) for stacked bar plots allows to sort the stacks
  each according to size.

• New argument panel.first for hist() (via plot.histogram()),
  barplot(), bxp(), and hence implicitly boxplot(). All except the
  former also gain a new panel.last. Suggested by Marieke Stolte and Julian
  Welz.

• plot.default() gets a new switch lim2 to indicate that the default xlim
  and ylim range computations only use jointly finite x- and y-values;
  indirectly suggested in PR#7986 by Jincai Jiang.

• Rman.css, the stylesheet for the texi2any-generated HTML versions of
  the R manuals, was improved following joint efforts at the ‘R Dev Days’
  of useR! 2024 and RSECon25.

• New system Rd macro \manual{<name>}{<node>} to refer to a specific
  section etc (possibly empty, for Top) of one of the R manuals (e.g.,
  R-exts).

• list.files() now has optional argument fixed to allow literal pattern
  matching. Suggested and implemented by Duncan Murdoch in PR#18925.

• substring(txt, i1, i2), substr(..) and their replacement versions now
  accept i2 = NULL to mean “to the end_of_string”, thus fulfilling Kevin
  Ushey (and others') suggestions in PR#18851.

• The default method for pretty() gets a new switch bounds for
  completeness, corresponding to .pretty(). Also, it now catches typos
  such as eps.corrected = 2, noted in PR#18521 by Mikko Korpela.

• Plain-text help now renders Rd \describe lists using line breaks after
  item labels. Set tools::Rd2txt_options(descStyle=NULL) to get the old
  behaviour with a simple space; a "colon" variant is also available.

• x %notin% table newly in base is an idiom for !(x %in% table) and
  provided almost entirely for convenience and code readability, from an
  R-devel suggestion, after many years of private definitions mostly hidden
  in packages, including in R's tools package.

• The default method of sequence() gets a new option recycle to swap
  between current (not fully recycling) and future behaviour where all
  three of (nvec, from, by) are jointly recycled.

• ftable()s default method and the as.table() method for "ftable"s get
  a perm argument with back compatible default. The as.table() method
  also gets a named.dim = FALSE switch, no longer creating a named
  dim(<result>) by default.

• match.arg() gets a new option several.ok = "all", fulfilling the wish
  of Jim Hester in PR#16659](https://bugs.r-project.org/show_bug.cgi?id=PR#16659).

• provideDimnames() gets new switch use.names, as proposed including
  patch by Mikael Jagan, in PR#18975.

• The default summary() is more informative for character vectors,
  optionally resorting to the factor method via the new argument
  character.method = "factor", e.g., when summarizing a data frame
  (PR#16750). It also provides more useful summaries for complex vectors, with polar
  determining if polar or cartesian coordinates are more relevant; mostly
  based on a proposal by Mikael Jagan.

• Error messages from failed Rd \Sexpr code now include the call from the
  original condition (if available). Warning messages now include Rd source
  locations.

• When the environment variable R_WARN_BOUNDS_OPT is set to "yes",
  options(warn = val) will check if val is inside -1:2; this may happen
  unconditionally in the future.

• weighted.residuals() now returns weighted working residuals for "glm"
  objects. It used to return deviance residuals, but these do not give the
  one-step approximated dfbeta values, as pointed out by Ravi Varadhan.
  influence.glm() has been amended so that it now returns an object with
  both a "wt.res" and a "dev.res" element.

• Several influence measures for "glm" objects have been updated to use
  Pearson residuals rather than deviance residuals (the latter are not
  unbiased) and to not use leave-one-out estimates of dispersion in models
  with fixed dispersion. Thanks to Ravi Varadhan for discussions and
  testing of these updates.

• R now supports custom binary types which are of the form
  "<system>.binary.<build>" where <system> is lower-case name (letters
  only) of the system and <build> is the name of the build (alphanumeric
  and dashes). The corresponding paths returned by contrib.url (which
  govern the repository layout) with the above type= have the form
  bin/<system>/<build>/contrib/<x.y> where <x.y> is the R version without
  patch level. This is a generalisation of the binary types previously
  supported on macOS. In addition, binary packages can have the extensions
  .tar.bz2, .tar.xz, .tar.zst and .tar.zstd with the corresponding
  compression support.

• The value of .Platform$pkgType which defines the native binary type for
  the current build of R can be overridden with R_PLATFORM_PKGTYPE
  environment variable (previously this was only set by the macOS builds of
  R). This should be typically only done by the author of a binary R
  distribution to define the type of binaries supported by that build of R.

• If R uses custom binary type in its .Platform$pkgType then
  R CMD INSTALL --build will produce
  <package>_<ver>_R_<system>-<build>.tar.xz binary files. Please note
  that future versions may use .zstd compression by default so package
  tooling should support any of the above mentioned extensions.

• New facilities have been introduced for package integrity verification
  and package signing. The optional file SHA256 in a package contains
  SHA-256 hashes, lengths and paths of files provided in the package
  allowing for more reliable integrity verification (relative to MD5). This
  file is now generated by default in R CMD INSTALL --build and
  optionally via R CMD build --sha256. A package can be signed by
  creating a detached GnuPG signature file SHA256.sig signing the
  entirety of the SHA256 file. Upon installation, the signature is verified
  using GnuPG, thus ensuring the integrity of the SHA256 file and by
  induction all files in the package. Both source and binary packages can
  be signed. Currently, this only applies to tar-based packages, and the
  command-line GnuPG gpg tool (or equivalent tool named in the GPG
  environment variable) has to be available for signing and verification to
  work. R CMD INSTALL --sign can be used to sign binary packages,
  including its use with --build. System-wide GnuPG keyrings for the R
  installation can be placed in <R_HOME>/etc/keyrings as files with
  .gpg extensions and will be checked first before consulting user's
  GnuPG keyring (if present).

• Experimentally, several math library (math.h) functions, such as
  exp(.), log1p(), sin(), atanh(), when called from R, are now
  ensured to give fully accurate results in a small neighbourhood of
  special values, such as exp(0) == 1, log1p(0) == 0, etc. This makes R
  slightly less platform dependent for these functions, notably on
  platforms with less accurate system math libraries, see also
  https://blog.r-project.org/2026/01/30/debugging-sensitivity-to-c-math-library-on-linux/.

• Function free1way() implementing semiparametrically efficient inference
  procedures (likelihood- and permutation-based) in distribution-free
  stratified K-sample one-way layouts was added to the stats package.
  Approximate (via power.free1way.test()) and simulation-based (utilising
  rfree1way() for random number generation) power-assessment and
  sample-size planning as well as model diagnostics (available from a plot
  method and probability-probability plots in ppplot()) complement the
  functionality. Provided by Torsten Hothorn.

• plot.data.frame() now supports a formula argument.

• New Rd macro \linkS4methods{}, e.g., for easier linking to method docs in
  other packages, proposed and coded by Mikael Jagan in PR#18998.

• Class "difftime" is now also among the S4 "OldClasses".

• New function bitstring() as a user-friendly wrapper around
  intToBits(), etc.

• plot.lm() now omits the smoother in the residuals vs leverage plot.
  This never made much sense, but if you really want it, you can force the
  old behaviour by setting panel.raw=panel.smooth.

• hatvalues() now has a method for nls (nonlinear model) results,
  prompted by the proposal of Patrice Kiener and Heather Turner in PR#18897.

• The Sweave driver RweaveLatex now allows the value of logical and
  numerical chunk options (only) to be the name of an object defined in
  earlier, evaluated code chunks. This is useful to pass computed values to
  options. For options set this way for weaving purposes, the Rtangle
  driver rather uses their default values, with a warning. By Vincent
  Goulet.

• The Sweave driver RweaveLatex gains an option to include unparsable R
  code or arbitrary text in code chunks. Chunks marked as ignore.on.weave =
  TRUE are completely ignored on weaving, i.e., not even parsed. The
  chunks however remain written out verbatim on tangling. Similarly, the
  driver Rtangle gains an option to control the extraction of code chunks.
  Chunks marked as ignore.on.tangle=TRUE are processed as usual on
  weaving, but left out of the tangled scripts (or thrown away when
  split=TRUE). The drivers also support the shorter aliases weave for
  !ignore.on.weave, and tangle for !ignore.on.tangle. A further option
  ignore, available for both drivers, sets ignore.on.weave and
  ignore.on.tangle at once. By Vincent Goulet.

• The Sweave driver Rtangle now provides an option chunk.sep to control the
  separator between code chunks in tangled scripts (with an unchanged
  default of \n\n). Using check.sep=FALSE omits the separator (between
  two chunks when used as a chunk option). By Vincent Goulet.

• The Sweave driver Rtangle now provides an option extension to specify the
  extension, without the leading dot, for the file name of a tangled code
  chunk when splitting is selected (the option has no effect with
  split=FALSE). If TRUE, the default extension of Rtangle (usually
  .R) is used; if FALSE, no extension is added to the file name. By
  Vincent Goulet.

• New function SweaveGetSourceName() in package utils to obtain the name of
  the file being woven by an Sweave process launched from the command line.
  By Vincent Goulet.

GRAPHICS:

• The graphics engine version, R_GE_version, has been bumped to 17 and so
  packages that provide graphics devices should be reinstalled. The
  glyphFont() function gains a variations argument, which allows variable
  font axes to be specified, e.g., c(wght = 100). Only the quartz()
  device and Cairo-based graphics devices (but not cairo_pdf()) support
  variable font axes so far.

INSTALLATION ON A UNIX-ALIKE:

• configure appends bsdtar to the list of command names it tries when
  finding a default for TAR.

• Failures in building the documentation for the base packages, including
  from parsing their Rd files, are no longer ignored by make docs.

• Building the R manuals now requires Texinfo 6.8 or later.

• A new LIBR_LDFLAGS defaulting to LDFLAGS allows more flexibility, e.g.,
  for distributions such as Gentoo, Debian or Ubuntu, fixing PR#18992;
  additionally a make failure to create libR.pc should no longer be
  ignored, fixing PR#18993; both thanks to Benjamin Drung.

CHECKING AN INSTALLATION:

• There is a new directory tests/C++Standards which contains small packages
  which can be test-installed to check support for various C++ standards.
  This is principally for use on Windows as on a Unix-alike this partially
  duplicates tests done by configure - although it can also be used to test
  the behaviour of R CMD INSTALL on unsupported standards.

PACKAGE INSTALLATION:

• Specifying C++11 or C++14 in src/Makevars has long given a note: in
  99% of CRAN packages the specification was unnecessary as the default
  C++17 sufficed (in a few cases with a deprecation warning from some
  compilers). Support for these standards has been removed: the default C++
  standard will be used. R CMD config variables CXX11, CXX14 and their
  associated CXXxxFLAGS, CXXxxPICFLAGS CXXxxSTD, SHLIBCXXxxLD and
  SHLIBCXXxxLDFLAGS variables are no longer supported and reported as
  ‘defunct’.

• When installing binary packages the time stamp in the "Built" field is
  considered even if the package version is identical to allow binary
  package repositories to update binaries without the need to increase the
  package version.

• The "Built" column returned by installed.packages() and
  old.packages() now contains the full value of the "Built" entry from
  the (installed) DESCRIPTION file. Previously, only the R version from the
  "Built" entry was returned.

• If the timestamp passed as --built-timestamp to R CMD INSTALL is in
  RFC-2822 form (with time-zone offset) it will be converted to ISO UTC
  form to conform to the format used by R for purposes described above.
  (PR#19037, thanks to Dirk Eddelbuettel)

C-LEVEL FACILITIES:

• New functions GEcreateDD() and GEfreeDD() for allocating (and
  initialising) DevDesc structures. Can be used by external graphics
  devices. Satisfies PR#18292.

• New function R_eval_with_gd() designed to “lock” a graphics device
  during call to eval() within device drawing routines. External
  on-screen graphics devices should consider using this function to avoid
  crashes from the device being closed by the user (e.g., a mouse click)
  during drawing.

• Declarations for non-API functions LOGICAL0, INTEGER0, REAL0,
  COMPLEX0, and RAW0 have been removed from installed header files.

• Non-API entry points Rf_acopy_string and Rf_lazy_duplicate have been
  added to those reported by R CMD check.

• Packages using the non-API functions ATTRIB and SET_ATTRIB will now
  receive check NOTEs. See ‘Writing R Extensions’ for alternatives to use.

• R CMD check NOTEs on the use of these non-API entry points have been
  upgraded to WARNINGs in preparation for removing declarations and, where
  possible, hiding these entry points: R_nchar, R_tryWrap,
  Rf_GetOption, R_lsInternal, BODY, FORMALS, CLOENV,
  SET_TYPEOF, STRING_PTR, R_duplicate_attr, getConnection,
  R_data_class, STRING_PTR, SET_OBJECT, ATTRIB, SET_ATTRIB,
  Rf_findVarInFrame3.

• The deprecated macros CHARACTER_DATA and CHARACTER_POINTER now return
  const pointers since using writable pointers in packages is not safe.

• New function DATAPTR_RW() for use in implementing ALTREP Dataptr
  methods. This function should not be used in any other contexts.

• New function R_class() as the C equivalent of class() in R.

• New function Rf_isScalarString.

• The R_ObjectTable type definitions formerly in R_ext/Callbacks.h are
  now available in the new header file R_ext/ObjectTable.h. This new
  header file should be used in packages instead of R_ext/Callbacks.h`.

• Task callback support declared in R_ext/Callbacks.h has been enabled
  again. It was disabled for R 4.5.0.

• The non-API header file R_ext/PrtUtil.h is no longer copied to the
  installed include directory.

• New function R_mapAttrib for iterating over the attributes of an
  object.

• New functions R_altrep_class_name and R_altrep_class_package for
  retrieving the class name symbol and package name symbol of an ALTREP
  object. These should be used instead of ALTREP_CLASS.

• New function R_getAttributes returns the same result as the R function
  attributes. New functions R_getAttribCount, R_getAttribNames,
  R_hasAttrib, R_nrow, and R_ncol.

• The function Rf_isFrame has been removed; use Rf_isDataFrame instead.

• The function VECTOR_PTR has been removed. The non-API function
  DATAPTR is no longer declared in an installed header file.

• These entry points are now marked as hidden and no longer have
  declarations in installed header files: ENVFLAGS, EXTPTR_PROT, FRAME,
  ENCLOS, EXTPTR_PTR, EXTPTR_TAG, HASHTAB, IS_S4_OBJECT,
  LEVELS, NAMED, OBJECT, R_shallow_duplicate_attr,
  Rf_isValidString, Rf_lazy_duplicate, Rf_NonNullStringMatch,
  SETLENGTH, SETLEVELS, SET_BODY, SET_CLOENV, SET_ENCLOS,
  SET_FORMALS, SET_ENVFLAGS, SET_FRAME, SET_GROWABLE_BIT,
  SET_HASHTAB, SET_NAMED, SET_S4_OBJECT, SET_TRUELENGTH,
  STDVEC_DATAPTR, TRUELENGTH, UNSET_S4_OBJECT, XTRUELENGTH.

• New experimental API for working with variable bindings in environments.
  The function R_GetBindingType identifies the binding type without
  forcing a delayed or active binding. New bindings care created by
  R_MakeDelayedBinding, R_MakeForcedBinding and R_MakeMissingBinding.
  Delayed and forced bindings can be inspected with
  R_ForcedBindingExpression, R_DelayedBindingExpression, and
  R_DelayedBindingEnvironment. Bindings for ... in function call
  environments can be examined with R_findDotsEnv, R_DotsExist,
  R_DotsLength, R_DotsNames, R_DotsElt, R_GetDotType,
  R_DotDelayedExpression, R_DotDelayedEnvironment, and
  R_DotForcedExpression. Contributed by Lionel Henry, Davis Vaughan, and
  Charlie Gao in PR#18928.

• New function R_envSymbols returns a vector of the symbols for which an
  environment contains bindings (PR#18928).

• R CMD check NOTEs on the use of the non-API entry points PRCODE,
  SET_PRCODE, PRENV, SET_PRENV, PRVALUE, SET_PRVALUE,
  R_PromiseExpr, and Rf_allocSExp have been upgraded to WARNINGs in
  preparation for removing declarations and, where possible, hiding these
  entry points.

• New functions R_getVar and R_getVarEx. In most cases these can be
  used instead of the non-API functions Rf_findVar and
  Rf_findVarInFrame. Other cases can be handled by the new API for
  working with variable bindings.

• Packages using any non-API variables will now receive check NOTEs.

• New function R_getRegisteredNamespace finds and returns the namespace
  environment registered for a name, or R_NilValue if none is found.

UTILITIES:

• R CMD check now reports further clang warnings including -Wkeyword-macro.
  This is most commonly seen for the C23 keywords bool, true and false, but
  is also seen for nullptr (C23 and C++11). Masking these by macros may
  generate correct code (but not always) and always results in
  confusing-to-read code.

• The cascaded stylesheet R.css now adds less vertical white space when
  rendering Rd files' \itemize, \enumerate, and \describe list \items to
  HTML. Where R is used via extensive IDEs such as RStudio, their
  maintainers may need to adapt R.css files accordingly.

• R CMD check when passed a tarball looks at its contents and so no longer
  derives the package name from the tarball name (which can now be
  arbitrary).

• Package README.md files are now installed and featured in HTML help.

• package.skeleton() now invisibly returns the path of the created
  directory, similar to prompt() et al., and allows creating blank
  skeletons without any R objects.

DEPRECATED AND DEFUNCT:

• structure(NULL, <name> = <val>) is now defunct.

• Three tk*.slaves() functions were deprecated in favor of tk*.child()
  which do need Tcl/Tk version 8.6 or higher.

BUG FIXES:

• chol(x, pivot=TRUE) now zeroes the trailing submatrix when the
  algorithm stops early. Following the help page instructions on how to
  reconstruct the matrix in positive semidefinite cases gave incorrect
  results if the rank was 2 or more less than the dimension (PR#15853).

• Profiling "mle" objects (stats4 package) could fail because profile MLE
  used starting value from full MLE. Now starts from profile MLE from
  previous step.

• Setting attributes on primitive functions is now an error; previously it
  modified without copying (as noticed by Henrik Bengtsson on the R-devel
  mailing list). Relatedly, changing the environment() of a primitive
  function does no longer happen and gives a warning (to become an error,
  later).

• Dataset LifeCycleSavings had a misspelled row name: "Guatamala" was
  corrected to "Guatemala".

• The Pair(x,y) construction in the formula interface to paired tests did
  not work with the subset argument, due to the absence of a [.Pair
  method, which has now been added.

• Subsetting "table" objects keeps S3 classes more consistently, fixing PR#18845.

• When browser() is active or when otherwise debugging, R code such as
  if(TRUE) now longer triggers a new browser level, fixing PR#18885
  (differently than the fix for PR#15770), with thanks to Ivan Krylov.

• text() now truncates labels to maximum length of x and y (if it is
  longer), fixing PR#7084. Thanks to Heather Turner, Ella Kaye, and Philippe Grosjean.

• <Date> %in% set is again as fast as it was before R 4.3.0, via a new
  S3 method mtfrm.Date. <character> %in% <Date> and vice versa are now
  documented to work in concordance with == and as an exception to the
  typical match() behaviour which relies on “univariate” mtfrm() alone.

• It is now possible to convert to "snpc" units via, e.g.,
  grid::convertWidth(), fixing PR#18915. Thanks to Trevor Davis.

• The precip dataset had typos in the names "Bismarck" and "Pittsburgh" (PR#18895).

• tools::checkReplaceFuns() now deals better with replacement methods
  not available as regular functions in the namespace.

• length(xpr) <- n now also works for expression vectors, the same as for
  lists, i.e., padding with NULL entries if the length is increased.
  Thanks to Mikael Jagan's PR#18917.

• <complex>[i] <- NA now only sets the real part to NA, consistently
  with the as.complex(NA) behaviour since R 4.4.0; thanks to Mikael Jagan's PR#18918.

• format(<list>) calling format.default() now calls generic format()
  on each list element (rather than format.default()).

• tools::Rd2HTML() detects paragraphs in more cases of \Sexpr-generated
  Rd code, fixing a regression in R 3.6.0 originally reported by Georgi
  Boshnakov and later also in PR#18470. An inactive #ifdef block no
  longer breaks the paragraph.

• Parsing a multi-line string from Rd \Sexpr code no longer doubles the
  newlines.

• requireNamespace(.., versionCheck = *) now shows an error and returns
  FALSE when the package version check is not fulfilled, thanks to Iris
  Simmons and Trevor Davis's PR#18255.
  Additionally, the versionCheck argument is checked and when invalid, an
  error is signalled to avoid subsequent confusing error messages.

• requireNamespace() now throws an error even if the name space is
  already loaded if it is called with an unmatched argument such as
  quitely = TRUE or quiet = TRUE (argument quietly cannot be abbreviated
  as it follows ...).

• rep() signals more specific error messages, notably when using both
  times and each optional arguments. Fixes PR#18926, thanks to Ella Kaye and collaborators
  at useR! 2025's ‘R Dev Day’.

• jitter(x, amount, factor) becomes more robust for non-finite x or also
  negative amount, factor. help(jitter) is easier to read, fixing
  PR#17279, thanks to its contributors.

• Plain-text help (Rd2txt) lacked markup for \cited titles, which are now
  wrapped in single quotes.

• Default column labels are now correctly aligned when printing a matrix
  with right alignment (PR#18943, thanks to Mikael Jagan).

• poly(fct, n) signals an error again, for a non-ordered factor fct,
  thanks to Roland Fuss' notice on R-devel.

• terms(<formula>, simplify=TRUE) now works correctly also in the cases
  where the formula has more than one term with NAs such as f(u, b=NA) or
  g(NA), finally fixing PR#15275 thanks to Mikael Jagan's patch.

• The help page example(matplot) now correctly labels Sepal vs Petal
  dimensions in its first iris data plot, thanks to Jeff Dick's post on
  R-devel.

• diff(m, lag, dif) for matrix m now still returns matrices, also when
  lag * dif > nrow(m) (PR#18972, thanks to Mikael Jagan and Suharto Anggono).
  diff(<ts-matrix>) remains matrix, even when it has length zero.

• Values returned by active binding functions are now marked as not mutable
  to prevent unintended mutation in complex assignment operations.

• The default method of sequence() now allows to fully recycle all three
  of (nvec, from, by) as always documented; with thanks to Mikael Jagan's
  report and patch in PR#18304. Its behaviour remains back
  compatible for now, but can be switched to future behaviour via environment
  variable R_sequence_recycle, or explicit recycle = TRUE.

• Subassignment, x[ind] <- val, to 1-dimensional arrays no longer “drops”
  them to simple vectors when ind is a name, fixing PR#18973, reported
  by Thomas Soeiro.

• The implicitGeneric definition for toeplitz() produced a default method
  that failed when toeplitz(x, r) was called with more than the first
  argument. Simple fix from Mikael Jagan.

• The error message for array(dim = c()) now uses "'dim'", thanks to Hugo
  Gruson.

• Some legal but strange nested complex assignments are handled a bit more
  robustly in interpreted code.

• cacheMetaData() does not to mangle primitive generics.

• The S4 method norm(<ANY>, type = <missing>) for the implicit generic
  norm() now uses type = "O" as has been the intention for more than ten
  years, fixing PR#18979 thanks to Mikael Jagan.

• On Windows terminals Ctrl-C now breaks out of waiting for keyboard input,
  fixing PR#18982.

• The default summary() and its factor method now label missing values
  NAs rather than NA's, fixing a decades-old grammatical faux pas
  inherited from S (PR#16348 and PR#18948).

• Asymmetric toeplitz(x, r) is type consistent now, even when r of length
  0 or 1 is not used directly. Fixes PR#18996, thanks to Mikael Jagan's patch and test.

• format(r) now keeps names(r) also for raw objects r.

• rep.int() and rep_len() again drop names also for factors, fixing a
  regression in R 4.0.0. (Thanks to a report by Rui Barradas and PR#18999 by Suharto Anggono.)

• str(x) with a list-like x of a class having an str() S3 method using
  NextMethod(), now no longer outputs "List of <...>" in the default
  method, thanks to Ian Farm's report and fix proposal in PR#19001.

• as.POSIXlt(ch), as.Date(ch) etc now also work correctly when ch
  contains "Inf", via strptime(.), fixing PR#19006.

• c(<POSIXlt>, ..) now fully preserves sub-second accuracy, fulfilling a
  part of Suharto Anggono's wishlist PR#18989.

• When subassigning "POSIXlt", e.g., <dat>[i] <- v, v is now converted to
  the time zone of dat, before being assigned, see PR#18919.

• The "tzcode" attribute of "POSIXlt" objects should no longer depend on
  earlier calls, fixing PR#19005 thanks to Ivan Krylov, a patch by Suharto Anggono
  and reports of “Non-deterministic POSIXlt tzone in Europe/London ..” on R-devel.

• In intervals with very low density cdplot() was unreliable, returning
  probabilities that are greater than one or NaN (PR#19011). Fix
  from Christoph Dalitz and Achim Zeileis.

• R CMD build now installs packages providing their own vignette engine
  (PR#18191). Patch by Paul Smith.

• ar.burg(), ar.yw() and ar.ols() deal better with non finite AIC
  values thanks to Simone Giannerini's remarks on R-devel.

• Parsing deeply nested expressions no longer should segfault in certain
  circumstances, fixing five different checks on context stack size, thanks
  to Kevin Ushey's report and Michael Chirico's patch in PR#18994.

• The internal cwilcox function used in computations for the Wilcoxon
  distribution is now based on an implementation of a formula by Andreas
  Löffler provided by Andreas Löffler, Aidan Lakshman, and Ivan Krylov,
  substantially reducing the memory complexity of the original
  implementation (PR#18655).

• When Ops group methods for matrix-"ts" return a zero-extent matrix, its
  type is correct now. E.g., + now returns type "double", see
  also PR#18972.

• intToUtf8(<negative>) now returns NA and no longer accesses invalid
  memory, thanks to Kevin Ushey's report with patch in PR#19052.

• When file.choose() is terminated with an EOF, it now signals an error
  rather than returning garbage, thanks to Kevin Ushey's report and patch
  in PR#19053,