vectors: the per-bucket PRODUCT guard refuses small buckets the engine searches exactly #173

VectorIndex._ensure_product_quantization_ready() (src/arcadedb/vector.py, called by find_nearest, find_nearest_approximate and build_graph_now) raises "PRODUCT quantization in this ArcadeDB version requires at least pq_clusters indexed vectors per bucket" when a bucket index holds fewer vectors than pq_clusters. It was added in March (fc44025, ArcadeData/arcadedb#3723). Since ArcadeData/arcadedb#5417 (b37d610cd0, July), which every engine these wheels ship includes, the engine handles that case itself: a bucket with fewer vectors than pq_clusters is built and searched exactly, without PQ (LSMVectorIndex.isPQTrainable, and findNeighborsFromVectorApproximate falls back to the exact path).

Measured on both 26.11.1.dev0 dev wheels, 4 buckets of 5, 15, 16, 17 and 40 vectors, 8-d and 32-d, pq_clusters 16: with the guard bypassed, the JVM build behaves exactly like the native build. Below 16 per bucket the build succeeds, PQ is unavailable per bucket, and find_nearest_approximate returns the exact neighbours with exact scores (recall@1 and @5 of 1.0, score difference 0); from 16 per bucket up PQ is trained. With the guard, the JVM build raises at 5 and 15 per bucket. The native build applies the same check to the whole type (one shim), so it can refuse a small type that the engine would search exactly.

Fix: remove the guard and its calls, and replace the tests that pin it with a shared test of the engine behaviour (a small PRODUCT index builds and its approximate search returns the exact nearest neighbour).