Add ONNX fp32 export for onnxruntime consumers (1.73x CPU speedup vs PyTorch, zero accuracy loss)

Phase 4: LDAM head fine-tune (test F1 0.962, has_next_action 0.907, has_quality_finding 0.917)

Phase 3 v3: autonomous traces + tightened labels (test F1 0.962, decided recall 0.917 vs P2 0.833)

Initial Rust integration bundle (test F1 0.968, bf16 safetensors, candle-transformers compatible)

initial commit