> found that just building a BM25 index, and giving the LLM a BM25 "search tool" and casting a "wide net" wins every single time compared to learned methods

I've been saying this for a while now. Lexical is approximately the same thing as semantic when a frontier LLM is driving the search terms.