[g4lb]

normalize ran two regex substitutions on every input. The newline regex matched every "\n" and rebuilt the string, also when the input had no "\r". Now normalize runs the newline substitution only when the input contains "\r", and the regex matches only "\r\n?". str.replace now replaces NULL characters, and only when the input contains one. The output does not change.

Benchmark

Interleaved A/B, 20 to 30 rounds per side, Mann-Whitney test, Apple M5:

	CPython 3.11	CPython 3.14	PyPy 3.11
parse of benchmarking/samples/spec.md	-5.7%	-5.8%	-22.1%
render of spec.md	-5.2%	-5.1%	-15.1%
render of one 10 KB paragraph	-45.5%	-45.9%	-24.2%
parse of spec.md with CRLF line endings	-2.4%	-2.4%	-1.7%

I measured no slowdown on 31 benchmarks, including empty input, CR-only and NUL-heavy input, all presets, renderer.render and SyntaxTreeNode. A first version that used only str.replace was 3.7% slower on PyPy for CRLF input, so this version keeps the regex for "\r".

Safety

• Tokens, HTML, env and the syntax tree are identical on 1,003,726 inputs: the CommonMark spec, the port fixtures, the benchmark samples and random documents. I compared the commonmark, default, gfm-like, zero and js-default presets, and setups with typographer, breaks and mdit-py-plugins.
• The full test suite passes on CPython 3.11, 3.14 and PyPy 3.11. pre-commit (ruff, ruff-format, mypy) is clean.
• Added test_normalize for CRLF, CR, mixed line endings and NULL characters.