the probe first approach is right, and storing none so hidden likes don’t drag the averages down to zero. the classify gap is real tho. i’ve seen a 401 with require_login set to true that falls through to other_401, so read the body and not only the status code
keep post for the stable stuff like shortcode, timestamp and kind, and put caption and pinned in post_snap. that way hashtag lift reads the caption as it was at snapshot time and a later edit can’t rewrite old results. if the API gives you more_available, next_max_id, since and until, a fixed 30 day window is cleaner than mixing 4.3 and 8.1 day samples, though each extra page costs credits so a slow rolling first page cron might come out cheaper