Struct GroupingConfig 

pub struct GroupingConfig {
    pub medoid_min_similarity: f64,
    pub min_pairwise_similarity: f64,
    pub sampling_threshold: usize,
    pub sample_size: usize,
    pub max_component: usize,
}

Tuning for grouping. Similarities are in [0, 1]; the defaults are provisional and calibrated against the chain corpus.

Fields

medoid_min_similarity: f64

Smallest similarity a member may have to the medoid and stay in the group; members below it are ejected and regrouped.

min_pairwise_similarity: f64

Complete-linkage floor: the smallest similarity any pair inside a reported group may have. Below it the group is split.

sampling_threshold: usize

Component size above which medoid selection samples candidates rather than scoring every member, to avoid quadratic blow-up on huge components. The sample is deterministic and key-diverse, so repeated content cannot occupy every medoid candidate.

sample_size: usize

Number of candidate medoids scored when a component exceeds Self::sampling_threshold.

max_component: usize

Largest component refined as one piece. A component above this is cut into key-ordered pieces, each refined on its own. Content-identical units are never separated: one equivalence class can therefore exceed this limit by itself.

Refinement materializes the component’s pairwise similarity matrix and orders it once, so it costs O(k² log k) time and O(k²) memory. A codebase of thousands of structurally interchangeable units — generated code, or a repository built to make the scan expensive — still produces exactly that component, which is why the ceiling exists at all (AGENTS.md §2-10, §7).

Cutting costs recall, never soundness: each piece is refined by the same medoid and complete-linkage rules, so every reported group is still cohesive. What is lost is the chance that two members landing in different pieces would have grouped. The cut is by stable key, so it is deterministic, and the count of components it fired on is reported. Keeping equal-key units together prevents independently cut pieces from minting the same content-derived group and finding identifiers.

Which members the cut put apart is reported too, through GroupingSet::severed_by_the_ceiling. A caller that carries out the verified relations no group expresses needs it: a relation across the cut is not one refinement weighed and declined, and carrying it out would restate the set once per crossing — at the size that makes this ceiling fire, that is the whole report.

Trait Implementations

impl Clone for GroupingConfig

fn clone(&self) -> GroupingConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for GroupingConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for GroupingConfig

fn default() -> Self

Returns the “default value” for a type.

impl PartialEq for GroupingConfig

fn eq(&self, other: &GroupingConfig) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for GroupingConfig