Module features 

Structural-mode candidate-extraction features over the Syntax IR.

Structural mode never compares whole units pairwise — that is quadratic in corpus size. Instead every unit (function, method, closure) is reduced to a set of cheap per-unit features, and candidate pairs are proposed only where features collide or lie close. One pass over a SyntaxIrFile extracts four feature families per unit:

• statement windows (WindowFeature): hashes of fixed-length runs of adjacent statements, the fragment-level candidate signal;
• subtree fingerprints (SubtreeFeature): Merkle hashes over the IR tree, the exact structural-match signal;
• a characteristic vector (CharacteristicVector): shape-tag counts used as a cheap candidate filter;
• an approximate control-flow profile (CfgFeature) and an API-call profile (ApiCallFeature).

Rename invariance

Candidate extraction must survive Type-2 edits, so no identifier text and no literal text enters any hash, with one deliberate exception: API-call names. Lexical signal comes exclusively from token kind tags (TokenKind::tag) and shape tags (Shape::tag). API-call names are exempt because external API names are normalization-exempt, matching the Fast engine’s treatment of external names.

The control-flow profile is syntactic

CfgFeature is a syntactic approximation built from AST control shapes, not a real control-flow graph: it linearises loop, branch and match nesting plus control statements in source order. A compiler-provided CFG can replace it behind the same feature interface in a later phase; doing so changes feature derivation and therefore bumps FEATURE_SCHEMA_VERSION.

Determinism

Every output is derived from source order alone; no hash-map iteration order reaches any feature. Extracting twice from the same IR yields identical results.

Structs

ApiCallFeature

The API-call profile of one unit.

CfgFeature

The approximate control-flow profile of one unit.

CharacteristicVector

Shape-tag counts plus tree size and depth: a candidate filter.

FeatureHash

A 128-bit feature hash.

FileFeatures

The features of every unit in one file, in pre-order source order.

SubtreeFeature

One subtree fingerprint: a Merkle hash over an IR subtree.

UnitFeatures

The candidate-extraction features of one unit.

UnitRef

A reference to one unit inside a slice of FileFeatures.

WindowFeature

One statement window: a fixed-length run of adjacent statements inside one block, hashed from per-statement summaries.

Enums

FeatureKind

The kind of a persisted feature hash.

Constants

FEATURE_SCHEMA_VERSION

Version of the feature-derivation recipe.

MIN_SUBTREE_NODES

Minimum subtree size, in nodes (the subtree root included), for a SubtreeFeature to be emitted. Smaller subtrees are ubiquitous and would only inflate the candidate index.

SHAPE_TAG_SLOTS

Number of slots in CharacteristicVector::counts: one per Shape tag, with slot 0 unused because tags start at 1.

WINDOW_LENGTHS

Statement-window lengths, in statements. Windows slide with stride 1 over each block’s statement sequence; a block shorter than a length yields no window of that length.

Functions

extract

Extract the candidate features of every unit in file.