Repro: sqlparse.parse(b"SELECT '\x41', '\n' \xff") where the input is bytes that aren't valid UTF-8 (a single stray \xff is enough to take the fallback branch).
Cause: the non-UTF-8 fallback in Lexer.get_tokens decodes via unicode-escape, which evaluates backslash escape sequences in the SQL bytes (\x41 becomes A, \n becomes a newline, plus octal and \u…). The parsed token stream then no longer matches the raw bytes the database receives, so anything inspecting or sanitising the SQL bytes sees a different statement from the one that runs.
Fix: decode the fallback as latin-1, which maps all 256 byte values one-to-one without evaluating escapes. For input without backslashes the result is byte-identical to today; only the escape reinterpretation is dropped.

• ran the tests (pytest)
• all style issues addressed (ruff)
• your changes are covered by tests
• your changes are documented, if needed