found out my visual bookmark vault was quietly missing four hundred fourteen saved posts. the source dump had four thousand nine hundred eighty eight cards and the vault only carried four thousand five hundred eighty six items, and of those only four thousand five hundred seventy four could even trace back to a source id. a dry run of the importer confirmed the exact gap before anything ran. the missing posts weren't a clean tail either — they spanned a couple of years across reddit, x, screenshots, and instagram. the row id gap matched, which means they were removed in some past cleanup, not never imported in the first place.
the import that fixed it was comfortably boring: idempotent, self-healing, thirty one seconds, reusing cached previews and embedding vectors instead of re-downloading or re-analyzing anything. no new api calls, no explainer generation, just the missing rows coming back. when a tool is built to be re-run safely you get to trust it enough to run it blind, which is the entire reason the dry run exists.
then a second, smaller pass caught the genuinely new stuff. the source dump itself was a couple of weeks old and my new bookmarks since then weren't in it, so a fresh dump found exactly thirteen posts saved after the cutoff — camera gear, a sunscreen, a glucose biosensor, a tiny red string game thing. those imported with their previews and vectors reused from the earlier embedding pass. the vault landed at five thousand thirteen items and the search index picked it all up without a restart. the lesson is that library integrity is a routine, not an event: verify the count against the source, keep the import idempotent, and re-run it whenever the source changes.