fix: search both edge directions during deduplication

[prasmussen15]

Summary

• Edge deduplication now searches both forward (source→target) and inverse (target→source) directions when finding candidate duplicates via get_between_nodes
• Previously, edges with the same endpoints but reversed direction were missed entirely, leading to duplicate edges in the graph
• Bumps graphiti-core version to 0.29.0

Test plan

• Verify existing edge dedup unit tests pass
• Test with edges that have reversed source/target but represent the same relationship
• Confirm no regression in edge resolution performance

🤖 Generated with Claude Code

[claude]

The two semaphore_gather calls are sequential, but forward_edges_list and inverse_edges_list don't depend on each other. Consider combining them into a single parallel call to reduce latency:

Suggested change

	forward_edges_list: list[list[EntityEdge]] = await semaphore_gather(
	*[
	EntityEdge.get_between_nodes(driver, edge.source_node_uuid, edge.target_node_uuid)
	for edge in extracted_edges
	]
	)
	inverse_edges_list: list[list[EntityEdge]] = await semaphore_gather(
	*[
	EntityEdge.get_between_nodes(driver, edge.target_node_uuid, edge.source_node_uuid)
	for edge in extracted_edges
	]
	)
	all_edge_queries = [
	EntityEdge.get_between_nodes(driver, edge.source_node_uuid, edge.target_node_uuid)
	for edge in extracted_edges
	] + [
	EntityEdge.get_between_nodes(driver, edge.target_node_uuid, edge.source_node_uuid)
	for edge in extracted_edges
	]
	all_results: list[list[EntityEdge]] = await semaphore_gather(*all_edge_queries)
	n = len(extracted_edges)
	forward_edges_list = all_results[:n]
	inverse_edges_list = all_results[n:]

[claude]

Performance concern: This change doubles the number of database queries from n to 2n for edge deduplication. Consider modifying get_between_nodes to use a bidirectional match in the Cypher query instead:

MATCH (n:Entity {uuid: $uuid1})-[e:RELATES_TO]-(m:Entity {uuid: $uuid2})

Using - instead of -> would match edges in both directions with a single query, avoiding the performance regression.

If modifying get_between_nodes is not desired (to preserve its current directional semantics), consider adding a new get_between_nodes_bidirectional method instead.

[claude]

The variable seen_uuids shadows the seen dict used earlier in the function (line 256). While not a bug due to scope, consider using a more descriptive name like combined_edge_uuids to avoid confusion during code review.

[claude]

Review Feedback

Missing test coverage: This PR adds important bidirectional edge lookup behavior but doesn't include unit tests to validate it. Consider adding tests that verify:

1. Edges with reversed source/target are correctly found as duplicate candidates
2. The deduplication logic correctly combines forward and inverse results without duplicates
3. No regression in behavior when edges only exist in one direction

This would help ensure the fix works as intended and prevent future regressions.

[claude]

Missing graph_operations_interface check: Unlike the existing get_between_nodes method (lines 405-411), this new method doesn't check for a custom graph_operations_interface implementation before executing the default query. This inconsistency means:

1. Users with custom graph_operations_interface implementations will have their custom logic bypassed during edge deduplication
2. There's no edge_get_between_nodes_bidirectional method defined in GraphOperationsInterface to call even if a check were added

Consider either:

• Adding the interface check with a new edge_get_between_nodes_bidirectional method to GraphOperationsInterface, or
• Documenting that this method intentionally bypasses the interface (if that's the intended behavior)