The boot-time install check reports how many nodes a package declares and which of them are missing from the mesh. It never said which install record it read those numbers from — and the record is the half of the comparison that can be out of date.

That gap cost a day. A check reported 2 of 201 declared node(s) are ABSENT and named two files; the install record, read afterwards, declared 193 files and neither of those names. The number was right about something, and there was no way to tell what: a count taken over the current record and a count taken over a stale one looked exactly the same.

Every verdict now names the record it was taken over — where it lives, which version of it was read, when that version was written, and the stamps identifying the package snapshot its file map came from — and prints the map's own size beside the count it produced, so a record that has gone backwards shows up in the line itself.

The verdicts that are not a shortfall carry it too: "the mesh could not be read" is as unactionable as a missing node when nobody can say which record went unverified. And when a caller supplies no record, the line says so in words rather than leaving a blank that reads like an answer.