Topic 4: Filtering

The problem: ANN indexes are built to find nearest vectors, not "nearest vectors that also match a condition". How the filter and the index work together decides whether you get the right results, too few results, or silently worse results.

A shared setup for this topic: every vector in the Topic 1 dataset gets a tenant label, with tenants of very different sizes, so we can test filters from "half the data" down to "a few dozen vectors".