[jwalith4418-Mj]

Case-insensitive attribute selectors currently apply Python's Unicode .lower() to both operands. Consequently, [data-value="ÉA" i] incorrectly matches an attribute containing éa, and [data-value="k" i] matches the Kelvin sign K.

Selectors Level 4, section 6.3 restricts the i flag to ASCII case-insensitive matching, independently of the document language. Use the existing webencodings.ascii_lower helper for the selector constant and the generated attribute-value expression. That helper is already imported and available to compiled selectors. ASCII case differences still match; non-ASCII characters retain their identity.

The regression cases exercise all six attribute-value operators in both HTML and XML, distinguish accented uppercase/lowercase values, cover Unicode characters whose lowercasing introduces ASCII characters, and retain unflagged/s/S and i/I ASCII controls.

Validation:

• Before the compiler change: 18 regressions fail, 10 controls pass.
• After the change: all 28 focused cases pass.
• Full Windows CPython 3.14.3 suite: 475 passed, 9 xfailed, 88% overall coverage. Coverage emits one configuration warning because --source takes precedence over the project's include setting.
• python -m ruff check, Sphinx HTML, wheel/sdist and git diff --check: passed.
• A separate Linux CPython 3.12 environment installed the wheel and passed public-API checks across all six operators and both document wrappers; the installed import path was verified.

The full CI interpreter/platform matrix was not run locally. This was found during source inspection and reproduced through the public API. OpenAI Codex assisted with investigation, implementation, validation and this description.

[grewn0uille]

Please follow the contribution guidelines.