fix: f16 subnormal overflow + OpenChat 3.5 Q8_0 integration test Fix signed arithmetic overflow in f16_to_f32 for subnormal exponents. Add integration test that streams OpenChat 3.5 Q8_0 (7.7 GB) through the bgz17 indexer → 42.6 MB output (679× overall compression). Results: Attention 328×, FeedForward 920×, Embedding 3765×. Peak RAM: 524 MB. Time: 185s. 226 tensors indexed, 65 skipped. https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7#47