Modernize cChardet for Python 3.11–3.14 and stable 2.2.0

[PyYoshi]

Summary

• support CPython 3.11 through 3.14 and standardize local development, testing, and packaging on uv
• update the bundled uchardet to the maintained PyYoshi fork without changing cChardet's expected API
• release the GIL during native detection, avoid input copies, support bounded and streaming detection, and safely handle inputs beyond uchardet's internal 32-bit length
• return encoding, confidence, and language results; expose ranked candidates with detect_all() and optional ISO 639-1 language downweights
• add PEP 561 type information and bytes, bytearray, and memoryview support
• add streaming, JSON, byte-limit, and language-weight options to cchardetect
• publish reproducible performance and accuracy comparisons against optimized chardet and charset-normalizer, including their native accelerators and Pure Python fallbacks
• strengthen native, arbitrary-byte, wheel, and release validation and document the external test-data and release policies

Benchmark highlights

CPython 3.14.2, AMD Ryzen 7 8845HS, identical preloaded 166-file / 641,202-byte corpus:

Detector	Optimized serial	Optimized 4-thread	Pure fallback serial
cChardet 2.2.0a3	92.0 ms	42.1 ms	n/a (native)
chardet 7.6.0	129 ms (mypyc)	209 ms	549 ms
charset-normalizer 3.5.1	88.7 ms (Cython)	110 ms	493 ms

Optimized charset-normalizer is competitive with cChardet for serial detection. cChardet retains a material advantage for concurrent independent buffers after releasing the GIL.

Accuracy on all 158 files in the bundled uchardet corpus:

Detector	Encoding compatible	Language	Encoding + language
cChardet	94.30%	95.57%	91.14%
chardet	85.44%	87.34%	76.58%
charset-normalizer	72.15%	56.33%	45.57%

The analysis also reports exact and decode-equivalent results and keeps the independently maintained chardet and charset-normalizer corpora separate. See docs/performance-analysis.md for methodology, pinned revisions, limitations, and reproduction commands.

uchardet changes

• use the maintained PyYoshi/uchardet cchardet branch rather than changing to an API-incompatible upstream revision
• improve UTF-8 validation and GB18030 ranking
• initialize multibyte prober state to remove a compiler warning found with warning-as-error builds
• expose detector completion state through an additive C API
• preserve the existing default ranking; language weights are strictly opt-in
• pass all 153 enabled standalone uchardet C tests

Release preparation

• prepare the unreleased stable 2.2.0 changelog while keeping the package version at 2.2.0a3 until the release commit
• build wheels for CPython 3.11–3.14 on Linux, macOS, and Windows, including ARM64 targets
• smoke-test installed wheels, including type-marker files and CLI behavior
• build an sdist and publish tag builds through PyPI Trusted Publishing with attestations
• document the required protected GitHub pypi environment and PyPI publisher setup
• remove obsolete cibuildwheel.sh and manual GitHub release-asset download scripts

Verification

• make check in the locked uv environment
• ruff and strict mypy
• 28 Python tests on CPython 3.11, 3.12, 3.13, and 3.14
• 153/153 standalone uchardet C tests, including a warning-as-error build
• wheel and sdist builds plus twine check
• isolated installed-wheel smoke test
• cibuildwheel build-identifier validation for the supported Linux targets
• native ASan/UBSan job added to CI

Scope notes

• free-threaded Python is intentionally not supported in this release
• unlicensed external corpora and issue attachments are not downloaded or vendored; detector changes must be evaluated across complete, separately reported corpora