graph-node 2.1.0: createEdge/batchInsert silently discard metadata, blocking lossless recovery

Reproduced on the current published native binding

@ruvector/graph-node@2.1.0 (npm latest checked 2026-09-13), Node 24.14.1, macOS arm64. Resolved from @claude-flow/cli@3.41.2.

Both createEdge() and batchInsert() accept the documented metadata: Record<string,string>, return successful edge IDs, and retain endpoints/connectivity after a fresh-process reopen. However, all supplied edge metadata is absent from query results before and after reopen. Node properties round-trip in the same fixture.

Source inspection confirms this is not merely a result projection: create_edge and batch_insert build persisted edge properties from only __confidence and __embedding, never copying edge.metadata. run_query returns non-internal properties, so public metadata would be observable if stored.

Minimal disposable reproduction

Use a new temporary directory, never a live managed graph. Run writer and reader in separate processes against that temporary path.

const db = new GraphDatabase({
  storagePath: temporaryGraphPath, dimensions: 8, distanceMetric: 'Cosine'
});
const embedding = new Float32Array([1,0,0,0,0,0,0,0]);
for (const id of ['mem:a', 'mem:b', 'mem:c'])
  await db.createNode({ id, embedding, properties: { diagnostic: 'true' } });
const edge = {
  from: 'mem:a', to: 'mem:b', description: 'supports', embedding,
  confidence: 0.75,
  metadata: {
    sourceEdgeId: 'fixture-source-edge',
    weight: '0.8',
    metadata: JSON.stringify({ witness: 'fixture-witness' })
  }
};
await db.createEdge(edge);
await db.batchInsert({ nodes: [], edges: [{ ...edge, to: 'mem:c' }] });
console.log(await db.query('MATCH (a)-[r]->(b) RETURN a,r,b'));
// nodes: diagnostic property retained
// both edges: properties: {}

A fresh reader process with the same options returns the same endpoints/types and empty edge properties. Counts remain 3 nodes / 2 edges. This issue does not dispute the connectivity/persistence fixes in #809/#879/#938.

Impact and bounded proposed fix

This blocks lossless reconstruction of a native graph from retained managed SQL relationships: weights carried as public metadata, witnesses, original SQL edge identity and other metadata cannot be verified or retained by this API. Retrying causal-edge writes in Ruflo is not an acceptable workaround because that handler also writes SQL.

• Copy public JsEdge.metadata into stored edge properties in both single and batch insertion, with an explicit collision policy for reserved internal keys.
• Preserve native confidence/embedding handling; do not repurpose relation type or node IDs to smuggle metadata.
• Test exact public metadata before and after a separate-process reopen, including Unicode, JSON strings and reserved-key behavior.
• Audit createHyperedge() separately: source appears to ignore its metadata too, but the executed reproduction above covers ordinary edges only.
• Keep missing metadata observable; do not return successful import/recovery receipts without fidelity verification.

Related downstream persistence adapter fix: https://github.com/ruvnet/ruflo/issues/3313 . The constructor patch enables durable future graph writes but cannot repair this compiled native metadata omission. No production databases were changed during these tests.

Start in crates/ruvector-graph-node/src/lib.rs at the create_edge and batch_insert implementation, then run the disposable two-process JavaScript reproduction against a temporary graph path. Done means public edge metadata, including Unicode and JSON strings, is returned before and after reopen, while confidence and embedding handling and reserved-key behavior remain correct.