Module fingerprint 

Architecture fingerprints — a stable identity for a graph’s shape.

Two questions the experiment pool has to answer are different enough to need two answers:

• “Is this the exact same architecture I already ran?” → ArchitectureFingerprint::digest, an exact hash over node ids, node kinds and edges. Sensitive to renaming, which is what makes it usable as a dedup key.
• “What have I run that looks like this?” → ArchitectureFingerprint::node_tokens / edge_tokens, bags of type tokens that carry no node ids at all, compared with structural_similarity. Renaming scaler to norm leaves them untouched.

Both are derived from the same canonical form: nodes sorted by id, edges sorted, with edge.id, node.label and node.target excluded — they are cosmetic or deployment detail, not architecture. SubGraph nodes recurse by digest, so nesting terminates and the result stays independent of traversal order.

Structs

ArchitectureFingerprint

Structural identity of a graph, exact and fuzzy.

EdgeRef

One edge, by node id — the diffable form.

Functions

pipeline_summary

One-line human description of a graph’s topology, for the pipeline_summary field of an experiment record.

structural_similarity

Structural resemblance of two fingerprints in [0, 1].