Struct FrozenFeatureHost 

pub struct FrozenFeatureHost {
    pub e_feat: DMatrix<f32>,
    pub b_feat: Vec<f32>,
    pub keep_target_indices: Vec<usize>,
    pub keep_src_indices: Vec<usize>,
    pub src_e_feat: DMatrix<f32>,
    pub src_names: Vec<Box<str>>,
    pub n_src: usize,
    pub h: usize,
}

Loaded + aligned frozen feature side ready to hand off to a candle engine (graph-embedding-util or one of the topic-model encoders).

Rows are reordered to follow the target axis. keep_target_indices records which positions in the caller’s target feature axis survived the intersection — the caller MUST restrict its data (triplets, encoder D, decoder β) to these indices, otherwise the row order disagrees with the embedding rows.

Fields

e_feat: DMatrix<f32>

[|keep|, H], rows in the same order as keep_target_indices.

b_feat: Vec<f32>

[|keep|]. Zeros when no bias_path was given.

keep_target_indices: Vec<usize>

Indices into the target feature axis that matched a source row. Length equals e_feat.nrows().

keep_src_indices: Vec<usize>

The source (dictionary) row each kept target row came from, parallel to keep_target_indices — what a caller needs to look up any other table keyed on the dictionary’s rows (a module membership, say).

src_e_feat: DMatrix<f32>

The whole source table [n_src, H] and its row names, as read — so a caller that needs the unmatched rows too (to place a gene by its neighbours’ rows) does not decode the file a second time.

src_names: Vec<Box<str>>

n_src: usize

Rows in the dictionary file, i.e. how many features the MODEL has.

The only field that survives the intersection unfiltered, and the reason it exists: e_feat and keep_target_indices are both already restricted to the matched features, so a coverage fraction built from them is identically 1 and tells a caller nothing. This is the denominator.

h: usize