d27d182 sought to DRY .writeheaders a bit, but introduced a bug. No diff is too simple for a test, apparently.
The built-in csv.reader only yields numbers if the dialect specifies csv.QUOTE_NONNUMERIC. By default (dialect=excel), this isn’t enabled, and avoiding this isinstance() call makes unicodecsv parsing about 2x as fast for large files.
…e runtime reading a large file with 45 columns.