LazyString is corrupted by copy, deepcopy and pickle on Python 3.11+

[DerevenetsArtyom]

Affected versions

• transifex-python 3.7.0 (latest on PyPI) and current devel
• Python 3.11 and 3.12 (works correctly on 3.10 and below)
• Independent of the Django version

Steps to reproduce

import copy, pickle, sys
from transifex.common.strings import LazyString

def render():
    return "hello"

v = f"{sys.version_info.major}.{sys.version_info.minor}"
s = LazyString(render)
print(f"py{v}  str(original)      -> {str(s)!r}")
for name, fn in (("copy.copy", copy.copy), ("copy.deepcopy", copy.deepcopy),
                 ("pickle round-trip", lambda x: pickle.loads(pickle.dumps(x)))):
    try:
        print(f"py{v}  {name:18} -> {str(fn(s))!r}")
    except Exception as e:
        print(f"py{v}  {name:18} -> {type(e).__name__}: {e}")

A module-level function is used rather than a lambda so the pickle case is meaningful on every version.

uv run --no-project --python 3.10 --with 'transifex-python==3.7.0' python repro.py
uv run --no-project --python 3.11 --with 'transifex-python==3.7.0' python repro.py
uv run --no-project --python 3.12 --with 'transifex-python==3.7.0' python repro.py

Observed behavior

py3.10  str(original)      -> 'hello'
py3.10  copy.copy          -> 'hello'
py3.10  copy.deepcopy      -> 'hello'
py3.10  pickle round-trip  -> 'hello'

py3.11  str(original)      -> 'hello'
py3.11  copy.copy          -> TypeError: 'str' object is not callable
py3.11  copy.deepcopy      -> TypeError: 'str' object is not callable
py3.11  pickle round-trip  -> TypeError: 'str' object is not callable

py3.12  (same as 3.11)

The original object is fine — only copies are corrupted, and the exception is raised later, when the copy is evaluated.

Expected behavior

A copied or unpickled LazyString should still evaluate to the same text, as it does on Python 3.10.

Why this is a problem

It is easy to hit through Django without doing anything unusual. django/forms/forms.py does self.fields = copy.deepcopy(self.base_fields) on every form instantiation, and ChoiceField.__deepcopy__ deepcopies its choices. So a lazy string used as a choice label — for example on a TextChoices — is corrupted on every request that builds such a form.

It is also hard to trace back to this library. The traceback surfaces inside copy and django.forms, and the only frame belonging to transifex is a generated wrapper named b, with no mention of __getstate__:

django/forms/forms.py:107        self.fields = copy.deepcopy(self.base_fields)
django/forms/fields.py:867       result._choices = copy.deepcopy(self._choices, memo)
copy.py:151                      rv = reductor(4)
transifex/common/strings.py:136  return func(str(self), *args, **kwargs)

Cause

Python 3.11 added object.__getstate__ (gh-70766 / bpo-26579), so '__getstate__' in dir(str) is False on 3.10 and True from 3.11 on.

lazy_str_meta rewrites every name in set(dir(str)) minus an exclusion set. That set deliberately protects the serialization protocol — it lists __reduce__, __reduce_ex__ and __getnewargs__ — but it predates 3.11 and therefore does not list __getstate__.

As a result LazyString.__getstate__ returns the evaluated text as the object's state rather than _func / _args / _kwargs. copy and pickle then apply that string as the new object's state, overwriting _func, and the next evaluation runs return self._func(*self._args, **self._kwargs) with self._func being a str.

object.__reduce_ex__ consults __getstate__ from C, which is why no intermediate Python frame appears in the traceback.

Suggested fix

Add "__getstate__" to the exclusion set in lazy_str_meta, alongside the other serialization-protocol entries.

__getstate__ is the only name added to dir(str) between 3.10 and 3.12, so this single entry covers the regression rather than being the first of several:

added to dir(str) in 3.12 vs 3.10: ['__getstate__']
removed: []

Note on CI

The test matrix already covers py3.11-dj4.1 and py3.12-dj4.2 and is green, because nothing in the current suite copies or pickles a LazyString.

I have a patch ready with the one-line change plus a regression test in tests/common/test_strings.py, verified to fail without the fix and to pass with it, and green on your py3.12-dj4.2 and py3.9-dj3.2 images. Happy to open the PR if that is welcome.