Author diversity resets on every For You request, so the same author repeats across consecutive loads

Problem
The same author can fill For You across consecutive refreshes and scrolls, because author diversity is computed per request and starts from zero each time.

Current behavior (6bb4594)

• author_pool_counts counts an author's posts only among the current request's candidates:
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/scorers/ranking_scorer.rs#L629
• AuthorDiversityDecay = 0.5, AuthorDiversityFloor = 0.25 (2nd post from an author in a request scored at 0.625, 3rd at 0.4375):
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/params/param.rs#L229-L245
• The query carries served_ids (post IDs only), and PreviouslyServedPostsFilter uses them to drop exact posts already served, not authors:
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/models/query.rs#L49
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/filters/previously_served_posts_filter.rs#L26-L30

Result: an author's first post in every new load gets a full 1.0 multiplier, even if the viewer was shown that author several times in the previous load seconds earlier.

Proposal

1. Carry recently served author IDs (bounded, e.g. last N requests or last few minutes) with the query.
2. Seed author_pool_counts with those counts, time-decayed, so diversity continues across loads instead of resetting.
3. Ship behind a param, off by default, for controlled evaluation.

Related

• #37: hard author cap within a single slate (off by default)
• #83: cross-request semantic fatigue (topic-keyed, not author-keyed)
• #172 (closed): diversity keyed on the retweeter

Open questions

• Window size and decay rate
• Whether mutual follows or authors the viewer engages with heavily should get a softer penalty

Start with home-mixer/scorers/ranking_scorer.rs and inspect how author_pool_counts is built, then trace served_ids through home-mixer/models/query.rs and previously_served_posts_filter.rs. Review the parameters in home-mixer/params/param.rs and resolve the proposed window, decay, and rollout choices. Done means bounded recent author history can seed diversity across loads behind an off-by-default parameter.