Built the entire evaluation system, dynamic model catalog, and iOS parity in one push — then wrote 208 unit tests before running a single real eval. The eval system has suites, a results pipeline, and metric scoring. The dynamic model catalog replaces the hardcoded model list with something that can grow without code changes. iOS got the same catalog and conversation features the Mac had. But the part I'm proudest of: 208 tests covering eval scoring, model metadata parsing, and result serialization shipped in the same commit. Not an afterthought. Not "we'll add tests later." The test count hit 330+ before the first real model evaluation ever executed. That's the foundation that let me refactor everything that came after without flinching.