Inconsistent query performance based on relation hit frequency

I'm running PostgreSQL 16.2 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 13.2.1 20230801, 64-bit, running on an Arch Linux box with 128GB of RAM & an 8c16t Ryzen 3700X CPU. Disk is a NVME RAID0.

Postgres configuration: https://paste.depesz.com/s/iTv

I'm using autovacuum defaults & am running a manual VACUUM ANALYZE on the entire database nightly.

The relevant database parts consist of a table with posts (note), a table with users (user), and a table with follow relationships (following). The query in question takes the most recent n (e.g. 50) posts, filtered by the users follow relations.

The note table on my main production instance grows by about 200k entries per week.

Schema & tuple counts: https://paste.depesz.com/s/cfI

Here's the shortest query I can reproduce the issue with: https://paste.depesz.com/s/RoC
Specifically, it works well for users that follow a relatively large amount of users (https://explain.depesz.com/s/tJnB), and is very slow for users that follow a low amount of users / users that post infrequently (https://explain.depesz.com/s/Mtyr).

From what I can tell, this is because this query causes postgres to scan the note table from the bottom (most recent posts first), discarding anything by users that are not followed.

Curiously, rewriting the query like this (https://paste.depesz.com/s/8rN) causes the opposite problem, this query is fast for users with a low following count (https://explain.depesz.com/s/yHAz#query), and slow for users with a high following count (https://explain.depesz.com/s/1v6L, https://explain.depesz.com/s/yg3N).

These numbers are even further apart (to the point of 10-30s query timeouts) in the most extreme outlier cases I've observed, and on lower-end hardware.

I've sidestepped the issue by running either of these queries based on a heuristic that checks whether there are more than 250 matching posts in the past 7 days, recomputed once per day for every user, but it feels more like a hack than a proper solution.

I've been able to reproduce this issue with mock data (https://paste.depesz.com/s/CnY), though it's not generating quite the same query plans and is behaving a bit differently.