Coding assistant + evals

Pick a synthetic encounter note, or paste your own, and a Sidekiq job hands it to a model adapter. The model can only answer through one tool, suggest_codes, whose schema admits codes from a small made-up catalog, and every suggestion has to quote the sentence that justifies it; a quote that isn't in the note gets flagged. Suggestions come back over Turbo Streams, one card per code, as the job appends them. The Evals tab is the other half of the demo: a prompt change is only safe when a golden set tells you what it broke, so you can edit the prompt there, re-run thirty graded cases, and see the score move before anything ships. Your note is never stored; it lives in the job's arguments and is gone when the job finishes.