Search procedure node materialization

Search procedures return logical nodes: ordinary stored properties, labels and identity remain readable, including a vector stored as a user property. Managed chunk and named embedding payloads remain internal and are not loaded for results. Neo4j 5.26.30 returns the full stored node from both vector and fulltext index procedures; reading node.embedding therefore returns a stored property when one exists, and null when there is only an internal managed embedding.

Owner and consumers

pkg/search/search_result_materialization.go owns the conversion from a stored hit to a search result and from that result to a typed node. The result retains the original read-time node metadata. SearchResult.MaterializedNode makes one detached output body with the result's selected properties and labels. This keeps creation/update times available to later property-visibility checks. Rebuilding a node from only its ID, labels and properties loses that information and can cause fresh properties to be treated as old under an enabled decay policy.

The consumers are:

• db.retrieve and db.rretrieve: regular rows, continuation page rows and grouped passages in pkg/cypher/call_rag.go.
• db.retrieveGraph and db.rretrieveGraph: ordinary rows and terminal graph pages through graphSearchEntity in pkg/cypher/call_rag_graph.go.
• Vector-only and fulltext-only hit enrichment through enrichIndexResults, and hybrid hit enrichment through enrichResults. Both use hitNodes and searchResultFromNode.
• Ranked, catalog and grouped continuations: compact retained rows discard the source node and property maps. Page hydration reads embedding-free bodies and uses the same result owner before the procedure adapter materializes nodes.

db.index.vector.queryNodes already reads complete storage nodes after ranking. db.index.fulltext.queryNodes already returns complete nodes from its ranked and wildcard paths. Neither rebuilds a metadata-free node. Relationship search and relationship index procedures return relationship values; retrieval graph and relationship procedures obtain their start/end nodes from storage directly.

TestSearchProcedureNodeMaterializationOwner scans production Cypher/search ASTs and rejects a second node projection in search adapters or a second search-result constructor carrying labels and properties. The repeatable shared regression contract covers the public node-returning procedures, ordinary and graph rows, continuation pages and grouped passages through Execute, Bolt and HTTP.

Retention and visibility

Final hit reads retain the existing BatchGetNodesWithoutEmbeddings / GetNodeWithoutEmbeddings contract. An embedding-bearing fallback read or an explicit property projection retains only a metadata copy, with managed vectors and excluded properties removed. The result cache accounts for the retained node body and embedding metadata. Compact continuations retain no source node body.

This preserves property suppression: the result carries the selected visible property map, and later normal-query reads still use the canonical visibility owner with the original timestamps. It does not bypass decay checks or fetch a newer node after ranking.