Persistence: round-trip the graph with embeddings and the retrieval index

[ronaldmannak]

Motivation

GraphRAG.save(toJSON:) persists the knowledge graph, but a built index (chunk/entity embeddings + the hybrid retriever state) is thrown away between runs, so every process start re-embeds the whole corpus — expensive with a remote embedder. Persisting and reloading a built index makes startup cheap and offline-repeatable.

Scope

• Extend save/load to round-trip chunk embeddings (and entity embeddings if present) so a reloaded graph doesn't need re-embedding.
• Add a GraphRAG load path that reconstructs a built instance from a saved graph (rebuild the HybridRetriever index from persisted embeddings without calling the embedder).
• Guard against embedder-dimension mismatch on load (re-embed or error clearly) — consistent with the LightRAG dimension guard already in Searchers.swift.
• Document the save/load lifecycle.

Files (mostly isolated)

Sources/GraphRAG/Graph/KnowledgeGraph.swift (persistence), Sources/GraphRAG/GraphRAG/Engine.swift (load/rebuild), tests. Low overlap with retrieval/community work.

Acceptance criteria

• Save then load reproduces query results without re-invoking the embedder (verifiable with a throwing/failing embedder on the load path).
• Dimension mismatch on load is handled explicitly.
• CI green.

Context: follow-up to PRs #2 / #3.