v2.2: filing prompt uplift + XBRL FIN_MAP expansion + diagnose-loop narrative

v2.1: SEC 10-K live eval (0.56 F1 at 0.06/doc) + CI ruff fixes

eval: multi-model benchmark - nano is Pareto-optimal (0.896 micro F1 at 0.0116/doc)

feat(eval): first live run + real README metrics (micro F1 ~0.92, $

Initial commit: schemas, extraction engine, dataset prep, 54 passing tests