Shared memory, so workers read stores in place.
Runs clustering off-thread when worthwhile. Resolve to null to have the service run it in-process instead.
Runs a search for FindNearestAsync, possibly off-thread. The result may
be approximate or a superset — the service re-ranks it exactly. With
TopK: null it must include every row that passes the threshold.
Native code for FindNearest: approximate when ANN is on, otherwise a verified exact scan.
In-process accelerator. Subclass it and register with
@RegisterClass(BaseVectorAccelerator, VECTOR_ACCELERATOR_KEY)from a package that loads after this one to take over vector work process-wide.