Release version 4.0.0: free-threading, CPython json parity, Python 2.7 fixes (#371)

Prepare 4.0.0 release: changelog, version bump, metadata updates

Add comprehensive CHANGES.txt entry for 4.0.0 covering all work since
3.20.2: free-threading support, C extension memory safety fixes,
encoding fast paths, Python 3.14 CI, SPDX license, and RawJSON docs.
Fix “smae” typo in 3.20.2 entry. Bump version to 4.0.0 in setup.py
and init.py. Add GraalPy to setup.py classifiers. Update README
to reflect Python 3.14 testing (was still mentioning 3.8).
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add pyproject.toml, migrate distutils imports, drop Python 3.3-3.5, add PEP 517 CI


Add pyproject.toml with [build-system] for PEP 517 builds. setup.py
is retained for Py2.7 wheel builds and backwards compatibility.
Migrate build_ext import from distutils to setuptools (with distutils
fallback for bare Py2 environments). Error classes stay in
distutils.errors since setuptools vendors distutils on 3.12+.
Remove imp module fallback in compat.py; use importlib.reload
unconditionally on Python 3 (imp was only needed for 3.3, EOL 2017).
Drop 3.3/3.4/3.5 classifiers and add !=3.3.* to python_requires.
Add MANIFEST.in entry for *.toml so pyproject.toml is in sdists.
Add test_pep517_build CI job that builds via python -m build,
installs the resulting wheel, and smoke-tests it. Gated under
gate_ubuntu alongside the other test jobs.
Update CHANGES.txt with all packaging changes.

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add Pyodide (wasm32) wheel builds with C speedups, fix changelog wording

Add build_pyodide_wheel CI job using cibuildwheel’s CIBW_PLATFORM=pyodide
to cross-compile the C extension to WebAssembly via Emscripten. Previously
Pyodide users fell back to the pure-Python none-any wheel; now they get
the compiled C speedups. The new job is gated under gate_ubuntu and
required by both upload_pypi and upload_pypi_test.
Also fix CHANGES.txt: remove “BACKWARDS INCOMPATIBLE” from the heap-types
entry since the Python-level API is unchanged – only the C extension
implementation changed, not its interface.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Skip thread and subprocess tests on Emscripten (Pyodide)

Pyodide/Emscripten doesn’t support threads (RuntimeError: can’t start
new thread) or subprocesses (OSError: emscripten does not support
processes). Add sys.platform == ‘emscripten’ checks in the shared
helper methods (_run_threads, runTool) so all affected tests skip
cleanly instead of erroring.
Verified by simulating sys.platform=’emscripten’: all 7 tests
(4 in test_free_threading, 3 in test_tool) skip with appropriate
messages. Native test suite is unaffected (206 tests, 39 skipped).
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fail hard when C speedups are missing under cibuildwheel

TestMissingSpeedups previously skipped silently when _speedups failed
to load, even in cibuildwheel where a compiled extension is expected.
Now it checks CIBUILDWHEEL=1 (set by cibuildwheel in both build and
test environments) and fails with a clear message instead. This catches
broken Pyodide wasm32 builds, native wheel builds, or any other
cibuildwheel-managed build where the C extension should be present.
Update CHANGES.txt to document the Emscripten test skips and the
speedup verification improvement.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Expand PR #370 changelog entry with specific fixes and optimizations

Replace vague “additional correctness fixes” with the actual changes:
ascii_escape overflow fix, list counter overflow fix, PyDict_Next fast
path, indexed list/tuple fast path, PyUnicodeWriter backend on 3.14+.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Update actions to v6, fix Python version refs, document minimum 3.4


Bump actions/setup-python from v5 to v6 (fixes Node.js 20
deprecation warning in CI).
Change Pyodide job host Python from 3.12 to 3.14 with a comment
explaining it’s just the host for cibuildwheel, not the Pyodide
target.
Update README.rst, index.rst, and CHANGES.txt to say “Python 3.4+”
instead of “Python 3.3+”, matching python_requires and the imp
module removal in compat.py. Consolidate the two CHANGES.txt
entries about 3.3 removal into one clear statement of the new
minimum.

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Narrow support matrix to Python 2.7 and 3.9+

Only these versions have wheels built and tested in CI. Update
python_requires to >=2.7 with 3.0-3.8 excluded, drop classifiers
for 2.5, 2.6, 3.6-3.8, and update README/index.rst/CHANGES.txt
to reflect the actual support matrix.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Remove pre-2.7 and pre-3.9 compat code, add Pyodide classifiers

Now that the support matrix is Python 2.7 + 3.9+, remove dead
compatibility code for 2.5, 2.6, and 3.0-3.8:

decoder.py: remove struct-based float(‘nan’) fallback for <2.6,
drop unused struct import
setup.py: remove Windows 2.6 IOError/ValueError ext_errors,
remove with_statement future import, add Framework :: Pyodide
and Environment :: WebAssembly :: Emscripten classifiers
tool.py: remove with_statement future import
tests/init.py: remove doctest <2.7 cStringIO workaround,
simplify TestMissingSpeedups (skipTest always available)
tests/_helpers.py: remove with_statement import and sys import,
simplify skip_if_speedups_missing (SkipTest always available)
tests/test_tool.py: remove with_statement, replace fragile
test.support/test.test_support imports with inline regex,
simplify open_temp_file (NamedTemporaryFile always available)
tests/test_decode.py: remove assertIs polyfill (available on 2.7+)
tests/test_namedtuple.py: remove namedtuple ImportError fallback
(available on 2.6+), simplify mock SkipTest check
tests/test_free_threading.py, test_subinterpreters.py,
test_speedups.py: remove with_statement future imports

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Remove Framework :: Pyodide classifier

Framework classifiers are for plugins/extensions of a framework, not
for libraries that merely run on the platform. Keep only the
Environment :: WebAssembly :: Emscripten classifier.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Move Python version requirement to top of 4.0.0 changelog

The support matrix change (Python 2.7 + 3.9+ only) is the most
important thing for users upgrading to simplejson 4. Move it to the
first bullet point and make it clear that pip will block installation
on unsupported versions.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Skip cp38 in cibuildwheel to match python_requires 3.9+

cibuildwheel v3.x builds cp38 by default. With python_requires
excluding 3.8, pip refuses to install the built wheel during
cibuildwheel’s test phase. Add CIBW_SKIP: “cp38-*” to the
build_wheels job.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Include Python 3.8 in support matrix (cibuildwheel builds it)

cibuildwheel v3.x builds cp38 wheels by default and the code works
fine on 3.8. Revert the CIBW_SKIP and add 3.8 back to python_requires,
classifiers, README, index.rst, and CHANGES.txt. Support matrix is
now Python 2.7 and Python 3.8+.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fix Python 2 syntax error: replace rb”” with b”” in test_tool.py

Python 2 doesn’t support raw bytes literals (rb”…”). Use regular
bytes literal with escaped backslashes instead.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fix Python 2.7 C extension load failure: clear exception in init_speedups

On Python 2.7, if init_speedups_state() fails (e.g. due to an import
error during module init), it leaves a Python exception set.  CPython
2.7’s import machinery checks PyErr_Occurred() after calling the init
function, and if an error is set, it removes the module from
sys.modules and propagates the error – even though the module object
was successfully created with make_scanner and make_encoder already
installed.
This caused the C speedups to silently fail to load on Python 2.7,
falling back to the pure-Python path.  The TestMissingSpeedups check
we added earlier made this visible.
Fix: clear the exception if init_speedups_state fails, since the
essential types and functions are already installed on the module.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Diagnose Python 2.7 init_speedups_state failure: print error to stderr

Print the actual exception from init_speedups_state to stderr before
clearing it, so the CI log shows what’s really failing on Python 2.7.
This is a diagnostic commit to identify the root cause.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Diagnose Py2.7 speedups import failure; separate Py2.7 CI job


TestMissingSpeedups now captures and reports the actual ImportError
message and lists _speedups files found in the package directory,
so the next CI run will reveal whether the issue is dlopen, missing
.so, or something else entirely.
Replace fprintf/PyErr_Print diagnostic in C init (which proved that
init_speedups_state succeeds) with PyErr_Clear for correctness.
Move Python 2.7 wheel build into its own parallel job
(build_wheels_py27) so results arrive sooner.

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fix Python 2.7 C extension load failure: guard PyDict_SetDefault

PyDict_SetDefault was added in CPython 3.4 and does not exist in the
Python 2.7 C API.  The unguarded call caused an undefined symbol error
at dlopen time, preventing _speedups.so from loading at all on 2.7.
Use PyDict_GetItem + PyDict_SetItem as a fallback for Python < 3.4.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fix Python 2.7 C extension: guard PyDict_GetItemWithError

PyDict_GetItemWithError was added in CPython 3.4.  The unguarded call
in json_PyDict_GetItemRef caused an undefined symbol error at dlopen
on Python 2.7.  Fall back to PyDict_GetItem for Python < 3.4.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add trailing comma detection in JSON decoder

Detect illegal trailing commas in objects and arrays (e.g. {“a”: 1,}
and [1,]) and raise JSONDecodeError with a specific message instead
of the generic “Expecting value” or “Expecting property name” errors.
Both the Python decoder (JSONObject/JSONArray in decoder.py) and the
C scanner (_speedups_scan.h) now report “Illegal trailing comma before
end of object/array”. Matches CPython 3.13+ json module behavior.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add array_hook parameter to decoder (CPython 3.15 feature)

New array_hook parameter for loads(), load(), and JSONDecoder that is
called with each decoded JSON array (as a list).  Its return value
replaces the list, analogous to object_hook for dicts.
Implemented in both the Python decoder (JSONArray in decoder.py) and
the C scanner (_speedups_scan.h).  Threaded through py_make_scanner,
loads(), and load().
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add frozendict encoding support (CPython 3.15 PEP 814)

Extend the Python encoder to treat frozendict (new builtin in CPython
3.15) identically to dict for JSON serialization.  On older Pythons
where frozendict doesn’t exist, behavior is unchanged.
Uses a _dict_types tuple that includes frozendict when available,
checked in _iterencode, _iterencode_list, and _iterencode_dict.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add exc.add_note() context on serialization errors (CPython 3.14)

On Python 3.11+ (PEP 678), serialization errors now get annotated
with notes describing where in the JSON structure the error occurred.
For example, encoding {‘a’: [1, object(), 3]} produces:
TypeError: Object of type object is not JSON serializable
when serializing object object
when serializing list item 1
when serializing dict item 'a'
Notes are added in _iterencode_list (item index), _iterencode_dict
(item key), and _iterencode (default() call).  On Python < 3.11 or
when using the C encoder, behavior is unchanged.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add _iterencode_once cycle breaker (CPython 3.15 optimization)

The mutually recursive closures _iterencode, _iterencode_dict, and
_iterencode_list create reference cycles that delay garbage collection.
Wrap the top-level call in _iterencode_once which deletes the closure
references in a finally block, breaking the cycle after encoding.
Only active on Python 3+ (uses nonlocal).  Python 2.7 returns
_iterencode directly as before.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fix _iterencode_once: move nonlocal to separate module for Python 2

Python 2 raises SyntaxError on nonlocal even inside an if-guarded
block, since the parser processes the entire function body.  Move the
cycle breaker wrapper to simplejson/_cycle_breaker.py which is only
imported on Python 3.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Revert _iterencode_once cycle breaker

The cycle breaker requires nonlocal+del inside _make_iterencode to
actually clear the closure cells.  Moving it to a separate module
only deletes copies of the arguments, not the real inter-closure
references.  Since encoder.py must be parseable by Python 2 (which
lacks nonlocal), this optimization cannot be implemented without
major restructuring.  Drop it — it’s a minor GC improvement from
CPython 3.15, not a correctness fix.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Use plain dict for _dict_types fallback instead of (dict,) tuple

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Update CHANGES.txt with new features and Python 2.7 fixes

Add entries for array_hook, trailing comma detection, frozendict
encoding, exc.add_note() context, Python 2.7 C extension fixes,
and the separated Py2.7 CI job.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Remove internal-only Python 2.7 fix entries from changelog

The PyDict_SetDefault/PyDict_GetItemWithError regressions were
introduced and fixed within the same release cycle, so they are
not user-visible changes worth documenting.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Remove PyErr_Clear() after init_speedups_state on Python 2.7

Silently clearing an exception from init_speedups_state would leave
the module in a broken state (missing string constants, RawJSON type,
etc.).  If state init fails, let the exception propagate so the import
fails loudly.  Also simplify the TestMissingSpeedups diagnostic now
that the Py2.7 undefined symbol issues are fixed.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Strengthen tests to match CPython json test coverage

Trailing comma tests: consolidated into test_trailing_comma_unexpected_data
with 4 cases matching CPython’s test_unexpected_data, verifying exact error
messages (not just substrings).
array_hook tests: added test_array_hook_with_object_hook to verify both
hooks work together on nested structures, matching CPython’s combined
object_hook+array_hook test.
add_note tests: replaced single weak test with three specific tests
matching CPython’s test_recursion.py coverage:

test_add_note_list_recursion: exact notes on circular list
test_add_note_dict_recursion: exact notes on circular dict
test_add_note_nested_error: exact notes list for nested
dict->list->object error path

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add Python 3.15-dev CI, frozendict tests, reject non-ASCII digits


Add 3.15-dev to test_pure_python CI matrix for early detection of
regressions and to validate frozendict encoding.
Add TestFrozenDict test class (skipped on Python < 3.15) covering
toplevel, nested, and mixed dict/frozendict encoding.
Add test_nonascii_digits_rejected: fullwidth digits like U+FF10
must not be accepted as JSON numbers (CPython gh-125687).
Fix NUMBER_RE in scanner.py to use [0-9] instead of \d, which
on Python 3 matches Unicode digits like fullwidth numerals.

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add test_prerelease CI job for Python 3.15-dev with C speedups

Build and test the C extension against Python 3.15-dev to catch
API changes early and validate frozendict encoding support.
Added to the gate_ubuntu dependency list.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Raise refcount leak threshold from 100 to 200

The test_failed_construction_no_leak test fires exception-heavy paths
(BadBool.bool raising ZeroDivisionError) which cause CPython debug
builds to accumulate internal refs (exception traceback interning, code
object specialization, etc.) reaching ~162 in phase2.  This is not a
real leak (a 1-ref-per-call leak would produce phase2 ≈ 2000).
The array_hook field we added to PyScannerObject is not even loaded in
the failure path (the error fires at PyObject_IsTrue before LOAD_ATTR
for array_hook), and scanner_clear properly handles it via the X-macro.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Collect GC every iteration in refcount leak tests

Calling gc.collect() inside the measurement loop instead of only
between phases prevents cyclic garbage from accumulating across GC
generations, which caused noisy refcount deltas (phase2=162) in
exception-heavy tests like test_failed_construction_no_leak.
This lets us restore the tighter threshold of 100 while eliminating
false positives from GC generation promotion timing.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add frozendict support to C encoder

frozendict (CPython 3.15, PEP 814) does not pass PyDict_Check, so
the C encoder was falling through to default() and raising TypeError.

Store FrozenDictType in module state (looked up from builtins at
init; NULL on older Pythons where it doesn’t exist)
Add is_dict_or_frozendict() helper that checks PyDict_Check first,
then falls back to isinstance(obj, FrozenDictType) when available
Use is_dict_or_frozendict in encoder_listencode_obj dispatch and
encoder_steal_encode (_asdict return check)
Use PyObject_Length instead of PyDict_Size for the empty-dict check
since PyDict_Size doesn’t work on frozendict
The slow path (PyMapping_Items) already handles non-dict mappings,
so frozendict iteration works without further changes

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Remove dead Python 3.3/3.4 single-phase init code, update docstring


Remove the #if PY_VERSION_HEX < 0x03050000 fallback in
PyInit__speedups (single-phase init via PyModule_Create) and the
NULL m_slots branch in moduledef. Python 3.3/3.4 are excluded by
python_requires, so multi-phase init (PEP 489) is always used.
Update module docstring to reflect current support matrix (Python
2.7 and 3.8+) instead of claiming Python 2.5 compatibility.

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Add C fast path for encode_basestring (ensure_ascii=False)

Previously, ensure_ascii=False fell back to a pure Python regex-based
encode_basestring.  Add a C implementation (escape_unicode_noascii)
that only escapes control characters (0x00-0x1f), backslash, and
double-quote, passing all other characters through unchanged.
Two code paths:

Python 3.14+: single-pass via PyUnicodeWriter
Pre-3.14 / Python 2.7: two-pass (compute size, then fill)

The C function is exposed as _speedups.encode_basestring and wired
into encoder.py as the default encode_basestring when available.
_toggle_speedups also handles switching between C and Python versions.
Closes simplejson/simplejson#207.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fix Python 2.7 build and -Werror on 3.14t debug builds


Guard module_slots and moduledef behind #if PY_MAJOR_VERSION >= 3
(PyModuleDef_Slot doesn’t exist on Python 2.7)
Guard escape_char_noascii and escape_char_noascii_size behind
if PY_VERSION_HEX < 0x030E0000 (unused on 3.14+ where the
PyUnicodeWriter path writes escapes inline)
Remove redundant nested #if PY_MAJOR_VERSION >= 3

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Update CHANGES.txt with encode_basestring fast path and other additions

Add entries for:

C fast path for encode_basestring (ensure_ascii=False), closes #207
Non-ASCII digit rejection in Python decoder
Dead code removal (Python 3.3/3.4 single-phase init)

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Use skipIf decorator for add_note tests instead of in-body check

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Collect GC during warmup phase in refcount leak tests

https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Fail hard if frozendict is missing on Python 3.15+

On Python >= 3.15, frozendict is a builtin and its absence indicates
something seriously wrong (e.g. MemoryError).  Use a compile-time
version guard in the C extension (#if PY_VERSION_HEX >= 0x030F0000)
and a runtime version check in encoder.py so that errors propagate
instead of being silently swallowed.
On older Pythons, FrozenDictType is simply NULL / _dict_types is dict.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Use PyAnyDict_Check C API for frozendict instead of runtime lookup

CPython 3.15 provides PyAnyDict_Check() (checks both dict and
frozendict) and PyFrozenDict_Check() as compile-time macros in
Include/cpython/dictobject.h.  Use these instead of storing the
frozendict type in module state and doing PyObject_IsInstance at
runtime.
On Python < 3.15, JSON_AnyDict_Check falls back to PyDict_Check.
Removes FrozenDictType from _speedups_state and its init/clear code.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Remove duplicate sys import in encoder.py

sys was already imported at line 8; the redundant import sys as _sys
for the frozendict version check was unnecessary.
https://claude.ai/code/session_01H3C9AwLogP8c74RodYcNZ4

Co-authored-by: Claude noreply@anthropic.com