GeezerVison

v7.17.2

Compare Source

• Fix: two different scenarios failed with "attempt to write a readonly
  database". These were regressions in 7.17.1. They are fixed, closing issue 2356_:

  • When using a version of SQLite older than 3.34.0, such as on RHEL 8 or
    Ubuntu 20.04, coverage combine (and pytest-cov, which combines
    automatically) always failed.

  • coverage combine and coverage debug data failed when a data file
    existed but had no coverage schema, for example an empty file left behind
    by a worker that died. A file we can't read is now reported like any other
    unusable data file, and is left in place.

.. _issue 2356: #​2356

.. _changes_7-17-1:

v7.17.1

Compare Source

• Fix: a one-line class or function body immediately followed by another
  statement (such as class Foo: ... on its own line) was incorrectly
  reported as a missing branch when using the sys.monitoring core. This is now
  fixed, closing issue 2167_ and issue 2337_.

• Fix: the HTML report only rewrites a file's page when the file or its data
  has changed, but each page also links to the previous and next files. When
  the set of reported files changed, unchanged pages kept their old links. For
  example, with skip_covered, a file that became 100% covered had its page
  deleted, but the pages next to it still linked to it. Now a page is also
  rewritten when its neighbors change.

• Fix: --fail-under now uses at least the decimal precision of the
  threshold for both the comparison and failure message. With precision=0,
  a total of 0.4 now passes --fail-under=0.3 and fails --fail-under=0.5
  with the message "total of 0.4 is less than fail-under=0.5". Previously, the
  comparison rounded the total to zero and the message could claim "1 is less
  than 0". Closes issue 2295_.

• Fix: the HTML report escaped file names twice, so a name with &, <,
  or quotes in it was shown with entities like &. Now they are properly
  escaped only once.

• Fix: --source=/ measured no files. A file system root can't be a source
  directory, so now it is an error. Closes issue 2291_.

• In an abundance of caution, prevent forged SQLite journal files from deleting
  files they shouldn't.

.. _issue 2167: #​2167
.. _issue 2291: #​2291
.. _issue 2295: #​2295
.. _issue 2337: #​2337

.. _changes_7-17-0:

v7.17.0

Compare Source

• Dropped support for Python 3.10.

• Binary wheels are now provided for iOS and Android. This allows the use of
  the optimized C tracer on mobile platforms.

• Fix: three or more levels of nested with statements could make a with
  line look like a branch with an exit that can never happen. Using the
  sys.monitoring core, these phantom branches were reported as partial. Other
  cores hid the problem but still counted the phantom branches in the total.
  Closes issue 1987_.

• Fix: a comment or string that merely mentioned if True: or
  while True: could mark its line as a branch that is known to be partial,
  so a branch that never ran was reported as taken. These built-in patterns
  now only match at the start of the statement. Closes issue 2314_.

• Fix: region detection missed functions defined in unusual places, such as
  else and except clauses. This is now fixed.

• Fix: recent versions of the lcov tool complain if the same function name
  appears more than once in a report. Python can have multiple same-named
  functions in a few ways, for example, getters/setters or conditionally
  defined functions. These are now disambiguated by adding a @<line number>
  suffix to the function name. Fixes issue 2335.

• Fix: The "Can't combine statement coverage data with branch data" error now
  names the two data files it couldn't combine. The :ref:error description <error_cant_combine> in the docs now lists common causes and how to find the
  process that wrote the file.

• Fix: at coverage totals close to 0 or 100, the XML report could get incorrect
  line-rate values like 8.333e-05 or 1. This is now fixed, closing
  issue 2293_ and issue 2294_.

• Fix: a parallel data file saved more than once kept the hash of its first
  save in its file name. Combining could then skip it as a duplicate of another
  file that had the same data at that point, and delete it, losing the data
  recorded after the first save. Each save now updates the hash. This had been
  a problem since 7.13.1. Fixes issue 2338_.

• Fixes for obscure but legal file names:

  • File names (and plugin names) can contain control characters, which are
    legal in POSIX paths but unsafe in reports: in XML 1.0 they make the
    document non-well-formed, and in a terminal they act as escape sequences
    (colors, title/clipboard sequences, cursor moves that forge output). Now
    the XML report, the text and Markdown summary reports, and coverage debug data replace non-printable characters in names. Printable names are
    unchanged.

  • A file pattern (from include, omit, or a [paths] alias)
    containing a newline sent the glob-to-regex translation into an infinite
    loop, hanging the process. Newlines in patterns are now handled like any
    other character.

• Fixes in the implementation of the C tracer:

  • If a plugin's dynamic_source_filename method raised an exception, lines
    in that function could be recorded as belonging to a different file. Also
    fixed some minor memory leaks in the C tracer.

  • The refcount of None was mismanaged during function calls, including
    comprehensions. This was unlikely to cause an actual problem, but is now
    fixed. Thanks, Andrew Dalke.

  • The C tracer held borrowed references to dictionary values, which was
    unsafe on free-threaded Python. Now it takes its own references.

  • An internal error while recording a new file could be reported as a
    confusing SystemError instead of the actual exception.

• Performance: the text summary report (coverage report) held on to every
  file's parsed source until it finished, so its memory use grew with the size
  of the project: ten times or more what 7.14 used on large projects. Now it
  keeps only what it needs to print. Fixes issue 2277_.

• Performance: HTML reports should be about 5% faster to produce.

• Kept the political banner in the README. Closes issue 2344_.

.. _issue 1987: #​1987
.. _issue 2277: #​2277
.. _issue 2293: #​2293
.. _issue 2294: #​2294
.. _issue 2314: #​2314
.. _issue 2335: #​2335
.. _issue 2338: #​2338
.. _issue 2344: #​2344

.. _changes_7-16-2:

v7.16.2

Compare Source

• Fix: on Python 3.14 and later, a for loop completing immediately before a
  function return could mistakenly report an uncovered branch. This is now
  fixed, closing issue 2168_.

• Fix: on Python 3.14 and later, the else clause of a try whose body is
  a with statement could incorrectly be reported as covered when the
  with raised. This is now fixed, closing issue 2289_.

• Fix: with dynamic_context = test_function, test methods written as
  @staticmethod or @classmethod were not given a context of their own.
  Now they are, on Python 3.11 and later. Closes issue 1923_.

.. _issue 1923: #​1923
.. _issue 2168: #​2168
.. _issue 2289: #​2289

.. _changes_7-16-1:

v7.16.1

Compare Source

• Fix: when the body of an irrefutable case (like case _:) is entirely
  excluded, the case line is now excluded too, just as an excluded
  else: body removes the else: line. Previously the case line was
  left behind and reported as missing. Closes issue 1563_ with pull 2269_.

• Fix: using :meth:.CoverageData.update twice on an in-memory database would
  fail, as described in issue 2279_. This is now fixed.

• Fix: the Total row in the footer of the HTML report's index page was
  recomputed in JavaScript with rounding that could disagree with the total
  the server rendered in the heading, including showing 100% when coverage
  was not complete. The footer now clamps values the way the rest of the
  report does, and when nothing is filtered out it is shown exactly as the
  server rendered it, so the two can't disagree. That also keeps rows the
  report omits (--skip-covered drops fully covered files) in the total.
  Fixes issue 2256_.

.. _issue 1563: #​1563
.. _issue 2256: #​2256
.. _pull 2269: #​2269
.. _issue 2279: #​2279

.. _changes_7-16-0:

v7.16.0

Compare Source

• When combining files, now path separator slashes will automatically be
  converted to the local file system style. This makes it less necessary to
  define [paths] configuration to combine data across operating systems.
  Fixes issue 2266_.

• The :meth:.Coverage.switch_context method now returns the previous context.

• Fix: previously, a [paths] pattern would be replaced everywhere in a file
  path when it was only meant to be replaced once, in the leading portion of
  the path. This is now fixed, in pull 2268_.

• Fixes to validation of options and configuration settings:

  • Negative precision settings now always cause useful error messages (pull 2261_).

  • An invalid regex in the --contexts option (or the [report] contexts setting) reported a confusing "Couldn't use data file ...:
    user-defined function raised exception" error. Now it raises a proper
    configuration error naming the bad regex, like other regex settings do
    (pull 2262_).

  • Non-string values in TOML configuration settings now produce a helpful
    error message instead of a traceback. This affects list settings whose
    elements aren't strings (like omit, exclude_lines, or a [paths]
    entry), file settings like data_file, and any wrong-typed value in the
    [paths] section (pull 2263_).

  • coverage run refuses run-affecting command-line options like
    --branch alongside --concurrency=multiprocessing, since they can't
    reach the subprocesses. The check only recognized multiprocessing as
    the entire option value, so --concurrency=multiprocessing,thread
    slipped through and failed later with "Can't combine statement coverage
    data with branch data". Each named concurrency library is now properly
    considered (pull 2270_).

• Fix: coverage annotate -d DIR raised an AssertionError if any
  measured file had an extension other than .py, such as a .pyw file on
  Windows. The original extension is now restored on the annotated copy (pull 2265_).

.. _pull 2261: #​2261
.. _pull 2262: #​2262
.. _pull 2263: #​2263
.. _pull 2265: #​2265
.. _issue 2266: #​2266
.. _pull 2268: #​2268
.. _pull 2270: #​2270

.. _changes_7-15-4:

v7.15.4

Compare Source

• Fix: in the HTML report, a source file name containing a double quote (legal
  on POSIX) wasn't escaped where it's dropped into the href of the index
  and prev/next links, so it could close the attribute early and inject markup.
  Page URLs are now escaped. Thanks, Rajath Mohare <pull 2227_>_.

• Fix: the LCOV report wrote file names and other fields into its
  line-oriented records without neutralizing control characters. A measured
  file whose name contained a newline (legal on POSIX) could forge extra
  records, inflating the coverage seen by tools that read the report. Control
  characters in a field are now replaced. Thanks, Rajath Mohare <pull 2226_>_.

• Wheels are now provided for Python 3.15.

.. _pull 2226: #​2226
.. _pull 2227: #​2227

.. _changes_7-15-3:

v7.15.3

Compare Source

• Fix: the sysmon core is incompatible with dynamic contexts. Previously, the
  combination would be prevented when read from the coverage.py configuration.
  But using the context API as pytest-cov does, contexts would be silently
  dropped. Now a warning is issued, thanks to Jisang Han <pull 2234_>.
  Closes issue 2200.

• A performance improvement in the low-level line number bookkeeping when
  combining data files, thanks to Kevin Turcios <pull 2239_>_.

• Performance improvement in HTML reporting by reducing the number of times
  files have to be parsed, thanks to Kevin Turcios <pull 2240_>_.

.. _issue 2200: #​2200
.. _pull 2234: #​2234
.. _pull 2239: #​2239
.. _pull 2240: #​2240

.. _changes_7-15-2:

v7.15.2

Compare Source

• Fix: one of the performance improvements in 7.15.1 (pull 2215) dramatically
  increased memory use during reporting for large projects. Now we use a
  different approach that is both faster and slimmer than 7.15.0. Fixes issue 2229_.

.. _issue 2229: #​2229

.. _changes_7-15-1:

v7.15.1

Compare Source

• Fix: in the HTML report with show_contexts enabled, a context label
  containing </script> (for example a parametrized pytest node id) could
  close the inline <script> element in a file page early, injecting markup.
  Context labels are now fully escaped. Thanks, Rajath Mohare <pull 2224_>_.

• A number of performance improvements thanks to Paul Kehrer, in pull requests
  2213 <pull 2213_>, 2214 <pull 2214_>, 2215 <pull 2215_>, 2216 <pull 2216_>, 2218 <pull 2218_>, 2220 <pull 2220_>, and 2221 <pull 2221_>_.

.. _pull 2213: #​2213
.. _pull 2214: #​2214
.. _pull 2215: #​2215
.. _pull 2216: #​2216
.. _pull 2218: #​2218
.. _pull 2220: #​2220
.. _pull 2221: #​2221
.. _pull 2224: #​2224

.. _changes_7-15-0:

v7.15.0

Compare Source

• Since 7.14.0, reporting commands implicitly combine parallel data files. Now
  those commands have a new option --keep-combined to retain the data files
  after combining them instead of the default, which is to delete them.
  Finishes issue 2198_.

• Fix: the LCOV report would incorrectly count excluded functions as uncovered,
  as described in issue 2205. This is now fixed thanks to Martin Kuntz Jacobsen <pull 2206_>.

• When running your program, coverage now correctly sets
  yourmodule.__spec__.loader as strongly recommended <--loader--_>,
  avoiding the deprecation warning described in issue 2208. Thanks, A5rocks <pull 2209_>_.

• Fix: with Python 3.10, running with the -I (isolated mode) option didn't
  correctly omit the current directory from the module search path, as
  described in issue 2103. That is now fixed thanks to Ilia Sorokin <pull 2211_>.

.. --loader--: https://docs.python.org/3/reference/datamodel.html#module.__loader_
.. _issue 2103: #​2103
.. _issue 2198: #​2198
.. _issue 2205: #​2205
.. _pull 2206: #​2206
.. _issue 2208: #​2208
.. _pull 2209: #​2209
.. _pull 2211: #​2211

.. _changes_7-14-3:

v7.14.3

Compare Source

• Fix: the default ... exclusion rule now also matches function bodies
  whose closing return-type bracket is on its own line (for example, after a
  long -> dict[ ... ] annotation that a formatter has split over multiple
  lines). Closes issue 2185, thanks Mengjia Shang <pull 2196_>.

• Fix: On 3.13t, we incorrectly issued Couldn't import C tracer errors.
  We can't import the C tracer because in 7.14.2 we stopped shipping compiled
  wheels for 3.13t. Thanks, Hugo van Kemenade <pull 2203_>_.

.. _issue 2185: #​2185
.. _pull 2196: #​2196
.. _pull 2203: #​2203

.. _changes_7-14-2:

v7.14.2

Compare Source

• Fix: some messages were being written to stdout, making coverage json -o - useless for capturing JSON output. Now messages are written to stderr,
  fixing issue 2197_.

• Fix: CoverageData kept one SQLite connection per thread that recorded
  coverage, but never closed them when those threads terminated. On long runs
  with many short-lived threads this leaked one file descriptor per dead
  thread, eventually failing with OSError: [Errno 24] Too many open files.
  Connections belonging to terminated threads are now closed and dropped.
  Fixes issue 2192. Thanks, Matthew Lloyd <pull 2193_>.

• Fix: when using sys.monitoring, we were assuming we could use the
  COVERAGE_ID tool id. But other tools might also assume they could use
  that id. Pre-allocated ids don't really make sense, so now we search for a
  usable one instead. Fixes issue 2187_.

• Following the advice of cibuildwheel <no-13t_>_, we no longer distribute
  wheels for Python 3.13 free-threaded.

.. _issue 2187: #​2187
.. _issue 2192: #​2192
.. _pull 2193: #​2193
.. _issue 2197: #​2197
.. _no-13t: https://py-free-threading.github.io/ci/#building-free-threaded-wheels-with-cibuildwheel

.. _changes_7-14-1:

v7.14.1

Compare Source

• Fix: the HTML report used typographic niceties to make file paths more
  readable by adding a small amount of space around slashes. Those spaces
  interfered with searching the page for file paths of interest. Now the report
  uses CSS to accomplish the same visual tweak so that searches with slashes
  work correctly. Closes issue 2170_.

• Add a 3.16 PyPI classifier <hugo-316_>_ since we test on the 3.16 main
  branch.

.. _issue 2170: #​2170
.. _hugo-316: https://mastodon.social/@hugovk/116588523571204490

.. _changes_7-14-0:

v7.14.0

Compare Source

• Feature: now when running one of the reporting commands, if there are
  parallel data files that need combining, they will be implicitly combined
  before creating the report. There is no option to avoid the combination; let
  us know if you have a use case that requires it. Thanks, Tim Hatch <pull 2162_>. Closes issue 1781.

• Fix: the output from combine was too verbose, listing each file
  considered. Now it shows a single line with the counts of files combined,
  files skipped, and files with errors. The -q flag suppresses this line.
  The old detailed lines are available with the new --debug=combine option.

• Fix: running a Python file through a symlink now sets the sys.path correctly,
  matching regular Python behavior. Fixes issue 2157_.

• Fix: Collector.flush_data could fail with "RuntimeError: Set changed
  size during iteration" when a tracer in another thread added a line to the
  per-file set that add_lines (or add_arcs) was iterating. The values
  passed to CoverageData are now snapshotted via dict.copy() and
  set.copy(), which are atomic under the GIL. Thanks, Alex Vandiver <pull 2165_>_.

• Fix: the soft keyword lazy is now bolded in HTML reports.

• We are no longer testing eventlet support. Eventlet started issuing stern
  deprecation warnings that break our tests. Our support code is still there.

.. _issue 1781: #​1781
.. _issue 2157: #​2157
.. _pull 2162: #​2162
.. _pull 2165: #​2165

.. _changes_7-13-5:

v7.13.5

Compare Source

• Fix: issue 2138_ describes a memory leak that happened when repeatedly
  using the Coverage API with in-memory data. This is now fixed.

• Fix: the markdown-formatted coverage report didn't fully escape special
  characters in file paths (issue 2141). This would be very unlikely to
  cause a problem, but now it's done properly, thanks to Ellie Ayla <pull 2142_>.

• Fix: the C extension wouldn't build on VS2019, but now it does (issue 2145_).

.. _issue 2138: #​2138
.. _issue 2141: #​2141
.. _pull 2142: #​2142
.. _issue 2145: #​2145

.. _changes_7-13-4:

v7.13.4

Compare Source

• Fix: the third-party code fix in 7.13.3 required examining the parent
  directories where coverage was run. In the unusual situation that one of the
  parent directories is unreadable, a PermissionError would occur, as
  described in issue 2129_. This is now fixed.

• Fix: in test suites that change sys.path, coverage.py could fail with
  "RuntimeError: Set changed size during iteration" as described and fixed in
  pull 2130_. Thanks, Noah Fatsi.

• We now publish ppc64le wheels, thanks to Pankhudi Jain <pull 2121_>_.

.. _pull 2121: #​2121
.. _issue 2129: #​2129
.. _pull 2130: #​2130

.. _changes_7-13-3:

v7.13.3

Compare Source

• Fix: in some situations, third-party code was measured when it shouldn't have
  been, slowing down test execution. This happened with layered virtual
  environments such as uv sometimes makes. The problem is fixed, closing issue 2082_. Now any directory on sys.path that is inside a virtualenv is
  considered third-party code.

.. _issue 2082: #​2082

.. _changes_7-13-2:

v7.13.2

Compare Source

• Fix: when Python is installed via symlinks, for example with Homebrew, the
  standard library files could be incorrectly included in coverage reports.
  This is now fixed, closing issue 2115_.

• Fix: if a data file is created with no read permissions, the combine step
  would fail completely. Now a warning is issued and the file is skipped.
  Closes issue 2117_.

.. _issue 2115: #​2115
.. _issue 2117: #​2117

.. _changes_7-13-1:

v7.13.1

Compare Source

• Added: the JSON report now includes a "start_line" key for function and
  class regions, indicating the first line of the region in the source. Closes
  issue 2110_.

• Added: The debug data command now takes file names as arguments on the
  command line, so you can inspect specific data files without needing to set
  the COVERAGE_FILE environment variable.

• Fix: the JSON report used to report module docstrings as executed lines,
  which no other report did, as described in issue 2105_. This is now fixed,
  thanks to Jianrong Zhao.

• Fix: coverage.py uses a more disciplined approach to detecting where
  third-party code is installed, and avoids measuring it. This shouldn't change
  any behavior. If you find that it does, please get in touch.

• Performance: data files that will be combined now record their hash as part
  of the file name. This lets us skip duplicate data more quickly, speeding the
  combining step.

• Docs: added a section explaining more about what is considered a missing
  branch and how it is reported: :ref:branch_explain, as requested in issue 1597. Thanks to Ayisha Mohammed <pull 2092_>.

• Tests: the test suite misunderstood what core was being tested if
  COVERAGE_CORE wasn't set on 3.14+. This is now fixed, closing issue 2109_.

.. _issue 1597: #​1597
.. _pull 2092: #​2092
.. _issue 2105: #​2105
.. _issue 2109: #​2109
.. _issue 2110: #​2110

.. _changes_7-13-0:

v7.13.0

Compare Source

• Feature: coverage.py now supports :file:.coveragerc.toml configuration
  files. These files use TOML syntax and take priority over
  :file:pyproject.toml but lower priority than :file:.coveragerc files.
  Closes issue 1643_ thanks to Olena Yefymenko <pull 1952_>_.

• Fix: we now include a permanent .pth file which is installed with the code,
  fixing issue 2084. In 7.12.1b1 this was done incorrectly: it didn't work
  when using the source wheel (py3-none-any). This is now fixed. Thanks,
  Henry Schreiner <pull 2100_>.

• Deprecated: when coverage.py is installed, it creates three command entry
  points: coverage, coverage3, and coverage-3.10 (if installed for
  Python 3.10). The second and third of these are not needed and will
  eventually be removed. They still work for now, but print a message about
  their deprecation.

.. _issue 1643: #​1643
.. _pull 1952: #​1952
.. _pull 2100: #​2100

.. _changes_7-12-1b1:

v4.26.0

Compare Source

=======

• Decrease import time by delaying importing of urllib.request (#​1416).

v1.9.0

Compare Source

Added

• Add support for Python 3.15 by @​AlexWaygood and @​thatch in #​1454 and #​1456
• Add CodemodCommand helpers for adding and removing imports by @​frvnkliu in #​1432

Fixed

• Recognize Python 3.14 parser configurations by @​bowiechen in #​1452
• Allow a trailing comma after **rest in class patterns by @​itamaro in #​1438

Updated

• Document the new CodemodCommand import helpers by @​frvnkliu in #​1437
• Fix the tokenize function's docstring grammar by @​galenseilis in #​1434
• Remove the retired macOS 13 runner from CI by @​drinkmorewaterr in #​1433

New Contributors

• @​AlexWaygood made their first contribution in #​1454
• @​bowiechen made their first contribution in #​1452
• @​galenseilis made their first contribution in #​1434

Full Changelog: Instagram/LibCST@v1.8.6...v1.9.0

v6.1.3

Compare Source

==================

Bugs fixed

• LP#2165901: External parameter entity parsing was allowed by default
  (with resolve_entities="internal").
  Issue found by Tomer Fichman.

v6.1.2

Compare Source

==================

• GH#526: Some build files were missing in the sdist.
  Patch by Nicola Soranzo.

• Some minor corrections for error handling cases.

Other changes

• Built with Cython 3.2.9.

v6.1.1

Compare Source

==================

Bugs fixed

• The known link attributes in lxml.html.defs.link_attrs were missing xlink:href,
  which can be used for URL bypass attacks in embedded SVG/MathML/etc. content.
  GHSA-4jhm-jv67-739f

• The Linux wheels use a patched libxslt 1.1.43, fixing CVE-2025-7424 and CVE-2025-11731.

• The Windows wheels use libxslt 1.1.45, fixing CVE-2025-7424 and CVE-2025-11731.

v6.1.0

Compare Source

==================

This release fixes a possible external entity injection (XXE) vulnerability in
iterparse() and the ETCompatXMLParser.

Features added

• GH#486: The HTML ARIA accessibility attributes were added to the set of safe attributes
  in lxml.html.defs. This allows lxml_html_clean to pass them through.
  Patch by oomsveta.

• The default chunk size for reading from file-likes in iterparse() is now configurable
  with a new chunk_size argument.

Bugs fixed

• LP#2146291: The resolve_entities option was still set to True for
  iterparse and ETCompatXMLParser, allowing for external entity injection (XXE)
  when using these parsers without setting this option explicitly.
  The default was now changed to 'internal' only (as for the normal XML and HTML parsers
  since lxml 5.0).
  Issue found by Sihao Qiu as CVE-2026-41066.

v2.5.4: (Oct 10, 2026)

Compare Source

NumPy 2.5.4 Release Notes

Released on 2026-10-10.

NumPy 2.5.4 is a patch release that fixes bugs discovered after the 2.5.3
release. Highlights are:

• Many typing fixes and improvements,
• A workaround for the MSVC 19.51 bug, see below.

This release supports Python versions 3.12-3.15.

Improvements

• numpy.linalg.svd and numpy.linalg.lstsq now raise for inputs containing
  infinities instead of returning garbage results.

• Fixed hangs in numpy.linalg.svd and numpy.linalg.lstsq for certain inputs
  containing infinities for recent versions of some LAPACK implementations.

• numpy.linalg.cond now handles non-finite inputs more consistently.

(gh-32593)

StringDType hashing with NaN sentinels

StringDType instances with equivalent floating-point NaN sentinels now
have equal hashes. Previously, distinct NaN objects could produce equal
dtype instances with different hashes, causing dictionary lookups and set
deduplication to fail.

(gh-32563)

Changes

MaskedArray resets a fill_value that cannot be represented in a new dtype

A fill_value copied from a source array is now reset to the default fill_value
for the new dtype when it cannot be represented in that dtype. Previously the
copied value could be stale: ufuncs that change dtype left the result holding a
fill_value typed for the old dtype, which raised a TypeError only when
something later validated it (such as .view()), and a floating point
fill_value that overflows an integer dtype was kept as an out-of-range value
together with a RuntimeWarning. Both now fall back to the default,
including when a masked array is passed to MaskedArray with a different
dtype. A fill_value given explicitly by the user is still validated and
raises as before. This can now raise a ComplexWarning if the fill_value is
complex and the new dtype is real.

(gh-32508)
(gh-32562)

Contributors

A total of 9 people contributed to this release. People with a "+" by their
names contributed a patch for the first time.

• Aniket Singh Yadav
• Bhuvansh Kataria +
• Charles Harris
• Iason Krommydas
• Joren Hammudoglu
• Lazizbek Ergashev +
• Nathan Goldbaum
• Sebastian Berg
• Yeonho Kim

Pull requests merged

A total of 17 pull requests were merged for this release.

• #​32520: MAINT: Prepare 2.5.x for further development
• #​32557: BUG: Reset an unrepresentable fill_value on a dtype change (#​32508)
• #​32565: BLD: work around MSVC 19.51 bug (#​32561)
• #​32600: BUG: fix np.strings.replace truncating old/new to a's
• #​32707: BUG: fix [ureduce]{#ureduce} reshape when a kept axis has size 0 (#​32536)
• #​32709: BUG: fix two possible crashes in the ufunc.accumulate implementation...
• #​32710: BUG: Reset an inherited fill_value in MaskedArray.__new__ as...
• #​32711: ENH: Improvements and fixes for stringdtype hashing (#​32563)
• [#​3

[!IMPORTANT]
✂ PR body was truncated to here.