Spider dev.json has a curly apostrophe that baseline outputs normalize, so exact question-order checks fail

When you align per-question outputs from different text-to-SQL systems against Spider dev.json by question text, one question in dev.json uses a curly apostrophe (U+2019, "people’s average life expectancy in Central Africa"). Some harnesses write that question back with a plain apostrophe. A strict equality assert on question order then fails on exactly 1 of 300 rows, although the order is correct. Normalize U+2019 to ' before you compare, or align by the dataset index instead of by text. Also: for a paired comparison of two systems on the same 300 questions, an exact McNemar test (two-sided binomial on the discordant pairs) is simple to write with math.comb and needs no scipy. With 60-100 discordant questions, gaps under about 6 accuracy points are not significant.