Stop NaN/Inf Phoenix heads from poisoning For You rank

[Pitchfork-and-Torch]

Problem

Phoenix can emit NaN (logged in model_runner, still returned). RankingScorer::apply did unwrap_or(0.0) * weight, so Some(NaN) / Some(±Inf) entered the sum. Dwell-regret then added that NaN into the slate mean, and every candidate’s score became NaN. TopKScoreSelector kept Some(NaN) (partial_cmp → Equal). DedupConversationFilter used score > best, which is false for NaN, so a NaN first-post locked the conversation.

This is not #142 (all heads None still ranked). This is not #181 (engagement-count store errors zeroing features). A present non-finite feature fail-opened into rank.

Five-line proof

1. Phoenix has_nan increments a metric and still reply_requests the batch (phoenix/xrex/inference/model_runner.py).
2. RankingScorer::apply / dwell-regret means used unwrap_or(0.0) — Some(NaN) stays NaN; one head NaNs mean_favorite.
3. centered_ratio then NaNs every candidate; offset_score writes Some(NaN).
4. TopKScoreSelector / Selector::sort treat NaN as Equal, so poisoned posts occupy For You slots; +Inf wins the slate.
5. After: non-finite heads are missing (0); persisted scores are finite or None; TopK last-places the rest; conversation collapse ignores NaN.

Change

• RankingScorer::finite_head / persistable_score — NaN/Inf heads do not enter the weighted sum or dwell-regret means; non-finite finals are unset.
• PhoenixScoresRankingScorer — same persist rule.
• TopKScoreSelector — non-finite → NEG_INFINITY.
• DedupConversationFilter — non-finite → 0 so a finite sibling can win.

Tests

• nan_favorite_head_does_not_poison_dwell_regret_slate
• inf_favorite_head_does_not_win_weighted_rank
• non_finite_final_score_is_unset_for_topk
• nan_score_does_not_lock_the_conversation
• non_finite_scores_sort_last

cargo test cannot run here: the public dump has no home-mixer crate manifest.

Survey

• Distinct from upstream Stop RankingScorer from ranking posts with no Phoenix heads #142 (missing heads) and Stop engagement-count store errors from zeroing For You rank features #181 (engagement-count zeroing).
• No existing PR on NaN/Inf rank-feature fail-open.
• Rebased onto xai-org/x-algorithm main (902a06f). One commit, four files.

[Pitchfork-and-Torch]

Parking this intern-stack PR per operator GitHub cleanup (2026-09-08). Hunt notes remain local. Not a reject of the class.