Struct PitFusionScoring 

pub struct PitFusionScoring {
    pub ego_depth: usize,
    pub blend: f64,
    pub agreement_bonus: f64,
    pub agreement_top_k: usize,
}

Percentile fusion: the same two signals as RRF, blended on their empirical distribution position rather than on rank alone.

RRF converts each component to a pure rank, which throws away the magnitude that says “this scored near zero”. Measured on the oracle corpus that costs 97 cases against EGO and wins 18, all on precision: BM25 gives any generic-token match a small positive score, and 1/(k + rank) promotes that noise to real fused mass (#125).

The probability-integral transform keeps the position. A fragment in the 5th percentile of a component contributes 0.05 from it, not 1/(k + 12). Two signals that disagree therefore cannot manufacture a strong candidate out of two weak opinions, which is precisely what the rank form allowed.

score = blend * PIT(ego) + (1 - blend) * PIT(bm25) + bonus * [both in top-k]

The agreement term is what fusion is actually for — a fragment both signals rank highly is more trustworthy than either alone — and it is additive and small so it breaks ties rather than deciding the ranking.

Fields

ego_depth: usize

blend: f64

agreement_bonus: f64

agreement_top_k: usize

Implementations

impl PitFusionScoring

pub fn new(ego_depth: usize) -> Self

Trait Implementations

impl ScoringStrategy for PitFusionScoring

fn score_and_filter( &self, all_fragments: &[Fragment], core_ids: &FxHashSet<FragmentId>, hunks: &[DiffHunk], repo_root: Option<&Path>, seed_weights: Option<&FxHashMap<FragmentId, f64>>, discovered_paths: Option<&FxHashSet<Arc<str>>>, deadline: Deadline, ) -> ScoringResult