I like this idea, particularly bringing more of the exploratory discussion back to the mailing list.
There are probably quite a few things that currently appear first as PRs that would be more interesting to discuss here as ideas. One example I've been experimenting with is relaxing the N -> 1 merge assumption and allowing merges to produce multiple, partitioned segments: https://github.com/apache/lucene/pull/16571 The original motivation was range routing, but it made me wonder whether the same mechanism could enable more unusual things, like partitioning segments according to vector space. I don't know if any of this is ultimately a good idea, but perhaps that's exactly the kind of thing we could brainstorm here before worrying too much about the implementation.
