fix(graph-node): wire NAPI binding to persistent storage (hydrate on open, edge/hyperedge write-through, delete cascade)

[Bolivar1234]

Fixes #809.

What

Wires the ruvector-graph-node NAPI binding to the persistence the ruvector-graph crate already provides, so storagePath graphs survive process restarts.

• Hydration: new()/open() now replay storage into both the hypergraph (traversal index) and the property graph — nodes, binary edges, and multi-node hyperedges. __embedding/__confidence are round-tripped as FloatArray properties so embeddings survive too.
• Write-through: createEdge, createHyperedge, and batchInsert edges now persist via storage.insert_edge/insert_hyperedge and mirror into the property graph (previously they only touched the in-memory hypergraph).
• Deletes: deleteEdge also cleans the hypergraph traversal index; deleteNode cascades through persisted edges/hyperedges touching the node, so a later hydrate never replays dangling references.
• Tolerant hydrate: dangling records (e.g. written by a pre-fix binary) are skipped and counted with a stderr report instead of failing open().

One file changed: crates/ruvector-graph-node/src/lib.rs (+227/−9). No crate/API changes; existing in-memory (no storagePath) behavior is untouched.

Verification

12-gate suite, each phase in its own process because redb is single-writer:

Gate	Result
create nodes+edge / multi-node hyperedge / batchInsert / tx / Cypher label scan	PASS
fresh-process reopen (exact stats)	PASS
reopen kHop over binary, hyper and batch edges	PASS
deleteEdge / deleteNode(cascade) persist across reopen	PASS
in-memory mode unaffected	PASS
single-writer lock still enforced on concurrent open	PASS

Scale check: a 154,711-node / 353,496-edge code graph (12 repos) reopens with exact counts after restart (hydrate ~15s on a shared host). Before this patch the same reopen returned {totalNodes: 0, totalEdges: 0}.

Happy to move hydration into the crate instead of the binding if you prefer that layering, or to add the test suite as a __test__ file in the package.

[Bolivar1234]

Pushed a follow-up commit (bbceebb) fixing a bug in my original patch: PropertyValue::from(Vec<f32>) resolves to the generic From<Vec<T>> impl and produces Array(Float(f64)...), never FloatArray — so hydrate (which matched only FloatArray) replayed every embedding as empty, and searchHyperedges returned f32::MAX sentinel scores for everything after a restart. Traversal was unaffected, which is why the kHop gates missed it.

Fix: persist sites write PropertyValue::FloatArray explicitly, and hydrate uses a tolerant extractor that also reads the boxed Array/List form, so databases written by the earlier revision stay readable. Verified with a new 'vector search survives reopen' gate (13/13 passing) plus a 16K-node real workload returning identical scores pre/post restart.