Build a per-matter cost model from usage logs

The project has three parts: a generator that writes a synthetic JSONL usage log, a pure call_cost that turns one usage object into dollars, and a report that groups costs by matter type, token class and agent. The decision that matters most is pricing each call from the row of a dated price table in force on the call's date, with the token classes kept apart, so the report can say which term to attack and a price change never rewrites old months. No model is called: everything runs offline.