Function rerank 

pub fn rerank<'v, F>(
    candidates: Vec<Candidate>,
    query: &[f32],
    metric: DistanceMetric,
    k: usize,
    opts: &VectorAnnOptions,
    sidecar: Option<&CodecSidecar>,
    fetch_vector: F,
) -> Result<Vec<Ranked>, RerankError>
where
    F: FnMut(u32) -> Option<&'v [f32]>,

Shared rerank pipeline. Both Origin and Lite call this after their index-level coarse search.

Callers use opts.oversample to compute fetch_k before pre-fetching from HNSW; this function receives whatever candidates were fetched and reranks by exact distance.

When opts.quantization is None (or VectorQuantization::None), the FP32 path is used: fetch_vector is called once per candidate and must return the stored full-precision vector. Returning None for any id is a hard inconsistency error.

When opts.quantization is Some(_), a CodecSidecar must be provided. The sidecar encodes query and stored vectors; fetch_vector is not called in this path.

When opts.query_dim = Some(d), the FP32 path applies Matryoshka truncated-distance reranking using only the first d components. d must satisfy 0 < d <= query.len(). query_dim combined with quantization is not supported — return BadInput if both set.

target_recall, oversample, and meta_token_budget are accepted via opts but not honored here — callers handle those before calling this function.