Add meritocratic author-size IPS to RankingScorer

[Pitchfork-and-Torch]

Summary

Phoenix already estimates P(action | viewer, post). That is the quality signal. Follower count is not a Phoenix feature, but audience size still leaks into For You through Thunder (follow graph), hashed author IDs, SimClusters log-fav retrieval, and the flat 0.75 OON tax.

This PR residualizes ln(1 + followers) inside the scored batch so two posts with the same Phoenix score are not ordered by how many people already follow the author.

p_i   = ln(1 + max(followers_i, 1))
ips_i = 1 / p_i
m_i   = 1 + alpha * (ips_i / mean(ips) - 1)
m_i   = clamp(m_i, 1 / max_boost, max_boost)
score' = score * m    if score > 0
       = score        otherwise

Mean-normalization keeps the batch arithmetic mean at 1. This reallocates score. It does not inflate it. Missing follower counts are identity. Negative scores (predicted block / mute / report) are not lifted.

Quality still wins. Default clamp is 2x, so a 1M-follower post that Phoenix scored 3x higher still ranks above a 100-follower post.

This is individual meritocratic fairness (Singh and Joachims KDD 2018, groups of size one; Biega et al. equity of attention). It is not demographic parity. No identity groups. No impression quotas. Low-quality posts stay low.

Why here, not UserCred / retrieval

• UserCred is an enforcement prestige graph. Ranking does not read it. Prestige-as-immunity is already in Remove credibility-based enforcement bypasses #18 / Make UserCred teleport prior verification-neutral #19.
• SidTail (authors under 1k followers) is the retrieval counterpart. A post that never enters the slate cannot be residualized. This PR does not invent a new cluster.
• Cold start (Smooth cold-start eligibility thresholds #22) lifts one low-impression original into slot ~15. IPS applies to every scored post, continuously.
• Author diversity only decays the 2nd+ post from the same author.

Knobs

Param	Default
rust_home_mixer_enable_author_size_ips	true
rust_home_mixer_author_size_ips_alpha	0.5
rust_home_mixer_author_size_ips_max_boost	2.0

alpha = 0 is a no-op. Applied after the Phoenix weighted sum, before author diversity, the OON tax, and cold start.

Math writeup: docs/MERITOCRATIC_AUTHOR_SIZE_IPS.md

Test plan

• identity when alpha=0, followers missing, or all authors the same size
• smaller author gets a higher multiplier; known-multiplier mean is 1
• clamp respects max_boost; 0 followers treated as 1
• negative / zero scores are not boosted
• equal Phoenix scores: 100-follower ranks above 1M-follower
• 3x Phoenix gap still beats the default 2x clamp
• RankingScorer integration: enable lifts, disable restores equality
• cargo test cannot run in the public dump (no home-mixer manifest)