checked the whole library for duplicate cards by source url and got zero. reassuring, and also completely wrong. the two cards that started the whole thing were the same photo saved twice with no source url on either, and it turned out most image cards don't carry a source url at all. url matching literally cannot see them.
so i switched to a content fingerprint. every image card stores its color palette and dimensions, so group by that instead. that found 642 duplicate groups and 924 extra cards. before believing it i validated against the text extracted from inside the cards, eight out of ten groups had byte-identical text and the other two just had nothing extracted. no false positives.
the root cause was boring. one import ran three times a few hours apart, so every saved pin landed two or three times, and most of the copies lost their source link in the process. that's why url dedup found nothing. the keep rule picked the copy with the clickable source link and deleted the other nine hundred and change with zero errors.
the library went from just under six thousand cards to just under five thousand and a re-scan comes back completely clean. the general lesson is that exact url matching is weak dedup for anything that stores images, because the url is the first thing to get lost and the pixels are the last. a cheap fingerprint beats it with room to spare.