OK, Ollama and Magistral and I are getting along better. I've made a model file with
PARAMETER num_ctx 16384
PARAMETER num_predict 16384
to give it more memory for input and output, and I've put a lot of background knowledge for the task in the system prompt. And I've remembered to invoke the new version and not the default one. The development is still experimental, and Magistral's explanation of its reasoning is very helpful so far. We'll see how long the honeymoon lasts.