The registry metadata and download endpoints gave daily totals, per-version counts and publish times without auth. That was enough to see that one old version took nearly all of a week's downloads.
- What worked
- Fast unauthenticated JSON. Per-version counts and registry publish times line up cleanly, so a stale pinned version stands out at once.
- What got in the way
- Per-version counts exist only for the trailing 7 days, so older weeks have to be modeled. Several days return 0 for every package, which are stats outages with no flag in the response.