fix: 6 bugs found running the ArcadeDB driver end-to-end (cross-tenant leak, MVCC retry, DDL syntax, edge save, temporal filter)

[agc-63]

Follow-up to getzep#1310 — two driver-side bug fixes found while running this driver end-to-end against a real ArcadeDB instance.

Based on feat/arcadedb-backend @ e9e4b47, targeting that branch (not main) so it merges cleanly into getzep#1310.

What's fixed

1. Cross-tenant data leak in execute_query()'s database routing — database_ was set inside the Cypher query params dict instead of passed as a real kwarg to the underlying neo4j driver call, so per-tenant database routing was silently ignored.
2. build_indices_and_constraints() DDL syntax invalid over the Cypher/Bolt channel — get_range_indices()/get_fulltext_indices() generated ArcadeDB SQL DDL sent over the Cypher/Bolt channel, which rejects it entirely. Rewritten with standard Cypher CREATE INDEX syntax; fulltext indexes dropped (not supported by ArcadeDB via Bolt).

What changed since the original 6-bug version

This PR originally included 4 more fixes (MVCC retry, single edge-save UNWIND workaround, datetime(...) wrap in retrieve_episodes(), plus a documented dehydrate_datetime app-level workaround). @lvca confirmed all 4 are now fixed upstream in ArcadeDB itself (arcadedb#4906, #4910, #4921, #4923 — see the comment on #1310), so those workarounds were reverted from this PR — no longer needed on the driver side. Verified end-to-end against a live arcadedata/arcadedb:26.8.1-SNAPSHOT container that the 2 remaining fixes are sufficient without them.

Testing

• 21/21 driver unit tests pass (tests/driver/test_arcadedb_driver.py), trimmed to cover only the 2 fixes above.
• 339 non-integration tests pass across the rest of the suite (same env vars/ignores as the CI unit-tests job).
• ruff check and ruff format --check clean.
• End-to-end verified against a real ArcadeDB instance: cross-tenant isolation (30 tenants × 15 rounds, 450 checks, 0 leaks) and build_indices_and_constraints() completing with zero syntax errors.

[gemini-code-assist]

Warning

Gemini encountered an error creating the review. You can try again by commenting /gemini review.

[lvca]

@agc-63 All six of your findings are now resolved upstream in ArcadeDB - four fixed and merged, two kept as genuine graphiti-side fixes. Thank you again; the repros and fail-before/pass-after tests made every one easy to act on, and you're credited on each issue.

Fixed upstream (merged to main):

• Update Maintenance LLM Queries and Partial Schema Retrieval getzep/graphiti#6 Bolt dropped native datetime params → inbound structs are now decoded + the engine normalizes native temporal params (arcadedb#4905 / #4906); outbound temporals return as native structs (arcadedb#4907 / #4921); and native offset-datetime writes now persist instead of being dropped (arcadedb#4922 / #4923).
• Cleanup maintenance utilities + add podcast runner getzep/graphiti#5 WHERE valid_at <= $reference_time never matched → fixed by Update Maintenance LLM Queries and Partial Schema Retrieval getzep/graphiti#6; plain comparison works, no datetime(...) wrapping.
• fix(arcadedb): wire up fulltext/similarity search (BM25 currently crashes) #2 No MVCC retry → conflicts now surface as a Neo4j transient error, so session.execute_write auto-retries (arcadedb#4908 / #4910).
• Refactor maintenance structure, add prompt library getzep/graphiti#4 Single edge save no-op → $param.field inside a MATCH pattern now resolves, so no UNWIND workaround needed (arcadedb#4909 / #4910).

So on a build with these, four workarounds can go:

• ✅ the dehydrate_datetime monkeypatch (Update Maintenance LLM Queries and Partial Schema Retrieval getzep/graphiti#6)
• ✅ the datetime(...) wrapping in graph_data_operations.py (Cleanup maintenance utilities + add podcast runner getzep/graphiti#5) - and revert that shared-file change so it no longer touches the other providers
• ✅ the MVCC retry loop (fix(arcadedb): wire up fulltext/similarity search (BM25 currently crashes) #2)
• ✅ the UNWIND [$edge_data] edge-save workaround (Refactor maintenance structure, add prompt library getzep/graphiti#4)

Keep the two genuine driver-side fixes:

• 👍 the cross-tenant database_ routing fix (fix: 6 bugs found running the ArcadeDB driver end-to-end (cross-tenant leak, MVCC retry, DDL syntax, edge save, temporal filter) #1)
• 👍 the Cypher CREATE INDEX DDL rewrite (fix(arcadedb): work around index-scan datetime comparison bug (3 code paths) #3)

One heads-up: ArcadeDB stores a datetime as an instant (zone normalized to UTC) and reads sub-millisecond precision back as a zone-less value - the instant is preserved and comparisons stay correct, just noting the offset isn't round-tripped. Non-issue for graphiti's semantics.

Could you trim the PR to the two keepers? Happy to do that pass and push to your branch if you'd prefer. Really appreciate you driving this.

[agc-63]

Thanks for the fast, thorough review — and for tracking down and fixing those 4 bugs upstream so quickly. Really appreciate it.

Trimmed the PR down to the two driver-side keepers, per your review.

Removed (verified end-to-end against a live arcadedata/arcadedb:26.8.1-SNAPSHOT container):

• MVCC retry-with-backoff loop in execute_query() — 20 concurrent writes to the same node, 0 losses, with zero retry logic on our side (the neo4j driver's own managed-transaction retry now handles it, since ArcadeDB maps the conflict to a proper transient error code).
• UNWIND [$edge_data] wrap in the single edge-save query — reverted to the plain MATCH (...) {uuid: $edge_data.source_uuid}, relationship still created correctly.
• datetime(...) wrap in retrieve_episodes() (graph_data_operations.py) — fully reverted, file now has zero diff against feat/arcadedb-backend, so it no longer touches the other providers. Temporal filter still matches correctly.
• (The dehydrate_datetime monkeypatch was never part of this PR — it lived at the application layer, as noted in the original PR description.)

Kept: database_ routing fix and the Cypher CREATE INDEX DDL rewrite. Re-ran the full unit suite (21/21 driver tests + 339 non-integration tests) plus ruff check/format — all clean.

[lvca]

@agc-63 perfect thanks!