[tae898]

Brings the vector guide, the rerank_oversample check and their tests in line with three ArcadeDB 26.11.1 engine changes.

Engine change	Upstream	What changes here
A BINARY index scores nodes on its 1-bit codes and reranks k * rerankOversample candidates (default 10), or the whole efSearch beam when larger; the rerank option works on it	ArcadeData/arcadedb#10975 (fixes #10968)	create_vector_index(rerank_oversample=...) accepts PRODUCT and BINARY on both backends; NONE, INT8 and no quantization keep the ValueError, as do pq_* and fused_pq without PRODUCT. Guide, docstrings updated
db.index.vector.queryNodes(index, k, vector, {where, filter}) options map	ArcadeData/arcadedb#10808 (feat of #10685), first in 26.11.1; 26.10.2 answers expects 3 arguments but got 4	"has no filter" sentence corrected in the guide, example 29 and the examples README; the map is shown once
A selective where on an indexed property takes the property index; broad and unindexed conditions unchanged	ArcadeData/arcadedb#10957 (fixes #10921)	One paragraph in "Filtered search" (index the property you filter on, no numbers, link to #10921); example 29 now creates NOTUNIQUE on Fact.tenant

Why rerank_oversample is not opened for every quantization: LSMVectorIndex.findNeighbors documents the setting as "Ignored when the index is neither PRODUCT nor BINARY", and LSMVectorIndexMetadata marks it "PRODUCT and BINARY". The metadata parser itself accepts the key on any index, so the stricter Python rule stays for NONE and INT8.

Verification:

• Tests (python -m pytest, pinned taskset -c 4,5, Python 3.13) ran against a scratch layout of the JVM wheel with the engine jars of upstream 313f23e234 (contains all three changes) and this branch's Python sources overlaid; no wheel was built. 139 passed across test_vector_product_quantization.py, test_vector_sql.py, test_vector.py, test_vector_index_introspection.py and test_docs_vectors_page.py. The same files pass on e353330c5c (has #10808, not #10957 or #10975) and the new queryNodes test and the BINARY tests pass on 26.10.2 jars through their old-engine branches.
• The BINARY tests fail on the old check (5 failed on origin/main sources over the 313f23e234 jars).
• Verified from engine source only, not by a test: that a BINARY search reranks k * rerankOversample candidates (a recall difference of a few percent with no deterministic assertion) and that the indexed pre-filter plan is taken (the new test pins the returned rows, which are the same whichever plan runs). Both are covered by upstream's own tests.
• Version gating follows the repository's patterns (has_pq_rerank probe and _engine_version()), no skips.
• Example 29 runs against 313f23e234 with its checks passing.
• black, isort, pre-commit, both bandit commands clean; the changed .py files compile.

🤖 Generated with Claude Code