[tae898]

Closes #116.

A VectorIndex wrapper held across a SQL REBUILD INDEX now keeps serving. Before, the old Java index object (replaced by the rebuild, isValid() false) answered searches with no rows and no error: on arcadedb-embedded 26.10.2.post1 with 2,000 vectors of dimension 16, find_nearest went from 5 rows to 0 and totalVectors from 2000 to 0, while a wrapper re-fetched with db.schema.get_vector_index returned 5 and 2000.

The fix is in src/arcadedb/vector.py: a new _check_live() runs at the start of every public method that touches the Java index (find_nearest, find_nearest_by_key, find_nearest_approximate, get_size, get_metadata, get_stats, get_quantization, warm_up, build_graph_now). When the cached index is no longer valid it looks the index up by name in the schema, replaces _java_index, and resets _lsm_indexes and _pq_ready. If the name no longer resolves (the index was dropped), it raises ArcadeDBError naming the index, never an empty result.

Hot path cost: one added JPype crossing per public call, isValid() on the cached object. The lookup by name runs only after an invalidation. find_nearest_by_key and find_nearest_approximate check once and then call internal helpers (_find_nearest_live, _get_quantization_live), so neither pays the crossing twice. find_nearest therefore adds exactly 1 crossing per call.

Native build: NativeVectorIndex names the index in every statement, so it has no cached engine object to go stale. _check_live is a no-op there. One related change: get_stats now reads the bucket index name afresh, because a rebuild builds new bucket indexes and the memoized name can be gone.

Tests, in tests/test_vector_index_introspection.py (shared by both backends, no Java object used): test_old_wrapper_serves_after_rebuild_index and test_old_wrapper_follows_a_second_rebuild hold the wrapper across REBUILD INDEX and assert 5 rows from find_nearest and the right totalVectors, size, and metadata; test_dropped_index_raises_and_never_answers_empty drops the index and asserts ArcadeDBError, with the index name in the message on the JVM build. On the old code the rebuild tests fail by the measurement above (0 rows instead of 5, totalVectors 0 instead of the row count), and the drop test fails because the old wrapper returned an empty list instead of raising. I did not run the suite locally; CI does.

Engine side, filed upstream as ArcadeData/arcadedb#10897: the old bucket LSMVectorIndex answers findNeighborsFromVector, get, countEntries and getStats with 0 instead of throwing (JDK 21 and 25, main c3af6d9028), while the old TypeIndex throws. This PR does not depend on that fix: the wrapper checks isValid() itself.

🤖 Generated with Claude Code