Most of our bank-file tests are fixtures we wrote ourselves, which can only find the mistakes we already imagined. So we ran the reader over someone else's: the 24 OFX files in ofxparse's test suite (MIT, from real Fidelity, Vanguard, Suncorp and ANZ layouts), and compared rows read against <STMTTRN> entries present.
Twenty-two matched. One is brokerage lines we read on purpose. The last, a file built to have bad dates, turned up a real bug: DTPOSTED of 20120231 went into the ledger as 2012-02-31. CSV dates were already calendar-checked; the OFX path only matched digits. A date like that would sort strangely, fall in no month's report, and break a period lock. It is now skipped with a "unreadable date" line, like any other bad date, and a leap day (20120229) still reads.
The same sweep is worth repeating on any format with its own date reader.