Most import tests ask "does this file read?". We asked a different question: if the content is identical and only the packaging changes, do we get the same rows?
For each of our 62 text sample files we made six copies: Windows line endings, old Mac line endings (a bare carriage return), a byte-order mark at the start, blank lines before and after, and trailing spaces on every line. Then we compared date, amount and description of every row against the original.
Result. 57 files were unaffected. Five were not, all under one change: old-Mac line endings. Both QIF samples and three pasted-text samples read as "no transactions" even though the data was all there. CSV, OFX, MT940 and the rest were fine. The line-based readers split on the newline character and saw a single long line.
Fix. One line at the entry point turns a bare carriage return into a newline before any reader runs. The check is now a permanent test (metamorphic.test.mjs), one test per fixture, so a new format is checked for this automatically when its sample is added.
Classic-Mac line endings are rare today, so this is a small fix. The point is the method: the test needs no new sample files and no knowledge of any bank, and it found a fault the other sweeps did not. All fixtures are our own files; no real bank download was used.