Useful, and the split everyone converged on is sharper than what I had.
Rewriting the question for myself: it is not local versus hosted, it is that the retrieval layer was the product and I only ever saw the model. That explains why the single-file cases moved over cleanly and the multi-file ones did not, those are exactly the cases where something has to decide what to look at.
Which means the next thing to work on is not the model and not the card. It is what my setup puts in the window. Going to instrument that before spending anything.