Module node_catalog 

One registry for every node in a graph — implementations, metadata, and trained states.

NodeCatalog holds both kinds of node. Filters and steps used to live in separate registries joined by a borrow-pair adapter, which meant three ways to answer “what is this node”, and callers that reached for the filter half alone silently skipped half the schema validation. There is one place to ask now.

The compiler reads metadata through NodeRegistry; the executor reads implementations and states directly. No intermediate conversion.

States live in a pluggable StateStore — by default MemoryStateStore, but users can inject a disk- or S3-backed store for pipelines whose trained states don’t fit comfortably in RAM. Steps hold no trained state: their history is the journal.

Structs

NodeCatalog

Every node a graph can execute, plus the states its filters have learned.

Enums

NodeImpl

What sits behind a node id.