`CsvConfig.quote` was only half wired up. The closing-quote scan and the unescaping consulted it, but the set of bytes that ends an unquoted field, the test for whether a field is quoted at all, and the check for a doubled quote all compared against a literal `"`. Any other setting therefore did not work: with `.quote = '\''` the input `'a,b',c` came apart into three fields rather than two, because the comma inside the quotes was still treated as a separator, and `a"b` failed outright with ShortBuffer, because the `"` still opened a quoted field that then had no `'` to close it. Take all five from the configuration. A configured quote now behaves as the default one does, including rejecting text after a closing quote, and the default quote becomes ordinary field data once it is not the quote. This costs about 8% on the quoted benchmark cases, which is the price of comparing against a value rather than a constant. The unquoted cases are unaffected, and an option that did not work is not worth the difference. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NGYDXszLdpBxhkfqqHp17B