Graph-interchange restore: import pre-extracted graphs from external systems

[aaronsb]

Idea

Raised while discussing the #505 restore-time durability path. The restore endpoint already imports a kg-backup/2 object (concepts + embeddings, sources, instances, evidence, relationships, ontologies, documents). That makes it, in effect, a bulk importer of graph-shaped data — and integration mode + the external-concept-id handling already exist to attach incoming data to an existing graph across embedding spaces.

The question: can we expose a supported path to import a pre-extracted graph from an external system (federation / interchange), distinct from ingesting raw content?

The boundary (important)

• Restore / interchange = faithful rehydration of an already-extracted graph. No LLM extraction, no concept embedding from scratch — the caller supplies concepts/sources/instances/relationships already in the schema, with compatible embeddings.
• Ingestion = turning raw documents into a graph (chunk → LLM-extract → embed → upsert). This is where the expensive extraction happens.

These must stay decoupled. The #505 durability path deliberately does NOT do extraction — it only reconstructs the derived :DocumentMeta tier from bytes already present. Folding extraction into restore would blur the two and complicate both.

What an interchange feature would need

• A documented external→kg-backup/2 mapping contract (what fields a caller must supply, embedding-profile/dimension compatibility rules).
• Validation at the untrusted-input boundary (the linter lint_backup.py already does referential integrity; extend for interchange-specific cases).
• Probably its own restore mode or a clearly-scoped variant of integration mode, and an ADR documenting the contract + trust model.

Status

Deliberately out of scope for #505. Filed to capture the idea; needs an ADR before any implementation.