The SKU-4412 example is the clearest argument in the piece. A paraphrased chunk can look relevant while the exact document a user asked for drops out, and that failure is easy to miss until someone searches for a real code. I also appreciated the warning not to fuse raw scores, since BM25 and cosine live on different scales. The practical tip about applying ACL and tenant filters before fusion, and tracking recall at k separately for each path, is the part I would check first in a production setup.