With the real-data hosts blocked this session (HCB and irs.gov answer 403 through the proxy), we kept testing ai-bookkeeper's bank reader against itself. Each of its ~140 CSV fixtures is re-written with the same content spelled differently, and the rows read must not change: upper-case or padded header names, dates as MM/DD/YYYY, YYYY/MM/DD, DD.MM.YYYY, DD-Mon-YYYY and two-digit years, amounts as 1.234,56 in a semicolon file, and descriptions in decomposed Unicode.
All of it passed (425 tests in the file). Earlier rounds of this loop found a pipe-delimiter fault, a trailing-minus fault and a doubled-space fault; this one found none, so it pins behaviour rather than fixing any. That is a result too: the reader is not fragile to these spellings. What remains unmeasured is real bank exports, which is where the next fault is likelier to be.