Gate out-of-network reach on follower engagement RATE, not raw in-network volume

[aunysillyme]

Problem
Reaching people outside your follower network is extremely difficult for mid-size accounts.

Example account: ~30k followers, ~15k active. Average post: ~1k views, i.e. at most 6.7% of
active followers (views also include non-followers, so true follower reach is lower).

The Aug 26 cold-start change (see #82) made early Home views drive out-of-network retrieval.
That rewards follower-base size rather than how well a post landed with the people who chose
to follow the author.

Contributing mechanics (6bb4594)

• Thunder retains in-network posts for 2 days (post_retention_seconds = 172800)
  https://github.com/xai-org/x-algorithm/blob/6bb4594/thunder/args.rs#L49
• Up to 1,200 in-network candidates compete for 35 feed slots (MAX_POSTS_TO_RETURN, RESULT_SIZE)
  https://github.com/xai-org/x-algorithm/blob/6bb4594/thunder/config.rs#L1
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/params/config.rs#L18
• Author diversity decay 0.5 / floor 0.25 scores an author's 2nd and 3rd posts at 0.625 and 0.4375
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/params/param.rs#L234-L245
• Author cold start only applies under 1,000 followers and 1,000 home views
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/params/param.rs#L660-L686
• OonWeightFactor = 0.75, TopicOonWeightFactor = 0.5
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/params/param.rs#L254-L277
• PostUnexploredWeight = 0.02, gated by PostUnexploredWeightInNetworkOnly = true
  https://github.com/xai-org/x-algorithm/blob/6bb4594/home-mixer/params/param.rs#L384-L406

Proposal

1. Compute an in-network engagement rate per post: engagements from followers / followers who
   were served it.
2. Once a post has been served to 30-40% of the author's active followers (4,500-6,000 for the
   example account) and its engagement rate clears a threshold, relax the OON discount for that
   post and make it eligible for Phoenix/SimClusters retrieval.
3. Keep Phoenix prediction as is; it runs on the follower pool first, then carries into
   out-of-network ranking.

Why

• Followers opted in, so their reaction is the most honest quality signal a post has.
• Rate-normalizing removes the advantage of simply having a larger audience: 500 engagements
  from 5k followers says more than 800 from 185k.
• Non-followers then see posts that arrive with proof of engagement, which gives new viewers a
  credibility signal and should improve follow conversion.

Open questions

• Minimum follower-impression count before the rate is trusted (to avoid small-sample noise)
• Guarding against engagement pods inflating follower rates
• Whether "active follower" should use an existing definition in the codebase or a new one