Fix str.replace with an empty pattern splitting characters - #8561

str.replace with an empty pattern inserts the replacement at every position in the subject. Wtf8::replace does that through a byte search, and an empty needle matches at every byte, so on anything outside ASCII the replacement lands inside a character.

>>> "á".replace("", "-")
'-ím'            # CPython: '-á-'
>>> "ábç".replace("", "#")
'#ãcb#ãǣ'      # CPython: '#á#b#ç#'
>>> "😀".replace("", "-")
'-𭟭ح-'          # CPython: '-😀-'

The bytes show it. For "á".replace("", "-"):

CPython:    [45, 195, 161, 45]       ->  -  á  -
RustPython: [45, 195, 45, 161, 45]   ->  -  \xC3  -  \xA1  -

The - went between the two bytes of á. Wtf8Buf::from_bytes_unchecked then accepts the result without looking at it, so the string is no longer WTF-8 and every later read of it returns something else:

A pattern that is not empty was already right and stays on the byte search. WTF-8 is self synchronizing, so a sequence never starts inside another one and the search cannot land off a boundary. Only the empty pattern has to walk code points, which is what insert_at_boundaries does. bytes.replace(b"", b"-") is right as it stands, since there the byte is the unit.

Checked against CPython 3.14.7: the empty pattern with and without a count, on ASCII, on Latin-1 range text, on an astral character and on a lone surrogate, plus non-empty patterns to confirm those did not move. The count behaves as a number of insertions, so "abc".replace("", "-", 3) is -a-b-c and the fourth insertion only happens at a count of 4.

Lib/test/string_tests.py exercises the empty pattern only on 'abc', which is why the suite passes either way. The new cases are in extra_tests/snippets/builtin_str.py and in the crate; three of the four unit tests fail without the change, and replace_non_empty_needle_is_unchanged passes both ways on purpose, to catch a fix that goes too far.

test_str, test_bytes, test_string, test_codecs, test_io, test_re and test_unicodedata all pass, 1685 tests.

Summary by CodeRabbit

• Bug Fixes

  • Fixed empty-pattern replacement to insert text only at valid Unicode code-point boundaries.
  • Preserved valid text encoding after replacements, including WTF-8 content.
  • Ensured replacement counts are respected when limiting insertions.
  • Maintained existing behavior for non-empty replacement patterns.

• Tests

  • Added coverage for Unicode boundaries, empty strings, replacement limits, encoding, and post-replacement text usability.