Your hybrid retriever is fusing scores that were never on the same scale

Most hybrid search tutorials end at the same place: run BM25, run a vector search, add the two scores with a weight, sort. It works on the demo corpus. Then the corpus grows and the ranking quietly ch