Cost-engineer an agent pipeline

The project is one pipeline with switches. matters.py generates the month once as agent runs (how many calls or loop steps each agent took on each matter). cost_model.py turns runs into usage objects and prices them, and a Config says which levers are on. Each milestone adds one lever to the previous configuration, so the cost-per-matter table in RESULTS.md shows exactly what each lever bought. The decision that matters most is that simulated usage has the same shape as real usage. The cache simulator emits Claude-style usage dicts (cache reads, 5-minute and 1-hour writes, iterations for compaction), so one pricing function prices the baseline, every optimisation and the ledger, and nothing can disagree by construction.