Day 1 expanded schema: 5 tables, 14B fine-tuned model

perf: load seed data from Parquet instead of generating on every query