What does this PR do?

Expands test_retrieve_online_milvus_documents (in sdk/python/tests/integration/online_store/test_universal_online.py) with three additional cases that exercise code paths adjacent to the bug fixes in #6275.

The existing test covered a single L2 vector query on a 3-row dataset with top_k=2. This PR adds:

1. Empty-store query

Queries retrieve_online_documents_v2 before write_to_online_store, so the Milvus collection exists but is empty. Asserts 0 rows returned. Exercises the "no hits" path of the v2 retrieval code.

2. Oversized top_k

Queries with top_k=5 on a 3-row dataset. Asserts all 3 rows are returned. Exercises the hit-parsing loop when the result set is smaller than the requested top_k — adjacent to the territory of bug #4 from #6275 (the bytes.fromhex(None) crash on a hit with a missing composite key).

3. Cosine-metric variant

Creates a second FeatureView (item_embeddings_cosine) with vector_search_metric="COSINE", writes the same dataset, and queries with distance_metric="COSINE". Verifies the cosine index is created and queried end to end. The existing test only covered the L2 path.

Motivation

Bug #4 in #6275 (TypeError: fromhex argument must be str, not None) was caught by a unit test (test_milvus_retrieve_online_documents_v2_missing_entity_key), but the universal integration test for Milvus retrieval did not exercise any of the adjacent edges (empty results, trailing hits, non-L2 metric). This PR closes that gap.

Tests

Verified locally against Milvus Lite:

No new imports are added; all types (FeatureView, Field, Array, Float32, String, timedelta, item) were already imported in the module.

Out of scope

• Adding Milvus to the test_retrieve_online_documents_v2 universal parametrization: that test asserts a BM25-style text_rank contract that Milvus's current LIKE-based text search does not satisfy. A capability-matrix refactor of the universal v2 test is a larger design question and will be raised separately.

🤖 Generated with Claude Code