[SanderMuller]

The inbox and the unread badge, which every page shows, both group a user's notifications with window functions. This makes that grouping cheaper in three ways:

• The windows over the whole history move into their own CTE. The per-topic windows after them then share one sort, where they used to sort every row twice.
• The badge counts the distinct groups that have an unread notification, instead of filtering ranked groups on group_rank and unread_count.
• The grouping reads only the seven columns it needs, so its sorts no longer carry whole rows. The inbox and the group popover join the rest of each notification afterwards. The inbox only does that for the rows with group_rank = 1.

What the inbox shows stays the same. On the stress database, the inbox and badge queries return the same rows as on production. The existing tests cover the group popover.

I measured it on the StressSeeder database, for an account with 38,583 notifications (20,250 unread). Each query ran 15 times with EXPLAIN ANALYZE, alternating with the production version:

Query	production median	This PR median	Change, median of the pairs
Inbox page	432 ms	349 ms	−18%
Unread badge	263 ms	226 ms	−15%

The machine was under load during the run, so the absolute times run high. The two versions ran back to back, so the ratio holds.

#407 changes lines next to these, so it may need a small rebase. #308 moves the inbox query into a new method, which would need the same join after a rebase.

🤖 Generated with Claude Code