The "New & Popular" category seems to be calculated like this:
ranking = new_ranking + popular_ranking
The problem is that the "new_ranking" is scaled to match the "popular_ranking" when unfiltered. But when a not so popular filter is applied (for instance "Metroidvania") then the "new"-score is so much higher compared to the "popular"-score, so that the popular score is neglectable. As a result the top 10 in "New & Popular" is identical to the top 10 "Most Recent". This is basically true for all filters except the most popular ones (like "horror").
This is not ideal. I think "New & Popular" ranking is/could be the most interesting ranking. "New" contains to much unfinished & uniteresting things, and "Popular" is very static...
Maybe this could be fixed by:
Multiple "New&Popular" rankings / indexes:
Create n different indexes for new and popular that are calculated like
ranking_n = new_ranking + popular_ranking * 2^n
Then, when filters are applied, choose a proper "n", so that the "new"-part and the "popular"-part are balanced for the filter selection.
New category "Popular&Active":
Rank by the normal "Popular" ranking, but filter out all projects that received no updates for the last 4 weeks.