GRAPH RAG FUSION: name every row in the Data Plane and drop the surrogate marker and the control-plane translator #428

Version / build tested against

origin/main @ d1e957217 plus the branch of PR #426.

Deployment mode

Origin — single node (local), and a three-node cluster

Engine(s) involved

Graph, Vector, Document

Summary

Follow-up to #424 and PR #426. Not a defect.

After #426, GRAPH RAG FUSION has two mechanisms that name a hit in no edge:

1. The owner's Data Plane reads the hit's client identity from its stored row (stored_client_identity) and returns it as key in RagVectorHit and RagTextHit. The cluster coordinator renames the marked rows from it (rag_fusion/naming.rs).
2. The Data Plane sets a surrogate marker on each result that falls back to the surrogate's decimal. A control-plane translator (response_translate/rag_fusion.rs) resolves the marked rows through the answering node's catalog (resolve_surrogate_pk).

The idea: let the Data Plane name every row, then delete the marker, the translator and name_marked_rows. One mechanism, and no deny_unknown_fields coupling between the Data Plane response shape and a control-plane copy of it.

A read of the tree on the branch of PR #426 says this is feasible in part:

The Data Plane never reads the catalog, so it cannot name the catalog-only rows. Net change: about 350 lines out, about 65 in, plus tests.

Proposal

Decide first whether a catalog-only row may answer its surrogate decimal. If yes:

1. Add a single-node three-source test where a text-only hit in no edge names its key, and an in-process RagStage::Local test for two and three sources.
2. Name the output rows late: keep the RRF keys as they are, so scores and ranks stay identical.
3. Read the versioned table in stored_client_identity for a bitemporal collection.
4. Pin the chosen catalog-only behaviour with a test.

Keying each hit by its stored identity before fusion would drop unbound entirely, but it fuses a bound vector hit with its text hit and changes scores. That needs its own change.

Expected behavior

A fusion row is named by one rule, in one place.

Actual behavior

Two mechanisms, and the second needs a marker field and a translator.

What actually happened? (check all that are true)

• Acknowledged/committed data was lost, corrupted, or silently wrong
• The server crashed, hung, or failed to start
• A security or isolation boundary was crossed
• Core functionality is broken with no acceptable workaround
• A workaround exists (rewrite the query, avoid one path, etc.)

None. This is a cleanup proposal.

Proposed severity

SEV-4 — Low: cleanup. The behaviour is correct after #426.

Reproducibility

Not applicable.

Last known-good version / commit (if a regression)

Not applicable.

Environment & logs

Linux x86_64, debug build. No log.

Before submitting

• I searched existing issues and this is not a duplicate. Related: Calvin commit refuses a predicate UPDATE or DELETE with XX000, and GRAPH RAG FUSION names a hit in no edge by its surrogate — transaction blocks and trigger bodies fail, fusion rows do not join to documents #424 (the defect), PR fix(txn,search): route predicate writes outside a block, refuse two writes of one sum balance, refuse the doc_id score alias #426 (the fix), graph: GRAPH RAG FUSION returns a surrogate number as node_id for any hit in no edge — every other search path returns the user PK #420 (closed).
• This is not a security vulnerability.