Banks reorder columns between export versions. A reader that finds columns by header name should not care. ai-bookkeeper now tests that: every rectangular CSV fixture is rewritten with its columns reversed and rotated by one, and must read the same dates, amounts and descriptions.
Files whose layout is a known bank's are skipped, since those are matched by the exact column sequence (and some read dates day first because of the bank). For the rest, one fault turned up: when a header had two equally ranked description candidates (Shopify's "Payment Method Name" and "Business Entity Name"), the leftmost won, so shuffling the columns changed the description. Candidates are now ranked by keyword, then by name. Same output on today's files; the choice no longer depends on position.
Limits: the fixtures are the repository's own, not real bank downloads.