The fix (anticlimactically small, as the best ones are): instead of cramming featured into the bundled comparison and letting Postgres faceplant, we just say up front, “we’re past the pinned posts now, so filter featured = false,” and then only bundle-compare the columns that actually matter (score, time, ID). Same posts. Same order. Postgres now estimates correctly, picks a sane plan, and the results:

~25 milliseconds. Down from up to 100,000.

That’s roughly a 4,000× speedup

This looks like something tbat should be considered by the upstream. Are you panning to create a pull request?