Neural Training

This handles neural pattern training with a proper architecture stack: SONA for self-optimization, MoE routing across 8 experts, HNSW for pattern retrieval that's 150x to 12,500x faster than baseline, and EWC++ to prevent catastrophic forgetting. The pipeline is retrieve, judge, distill, consolidate. You'd use this when you need pattern learning, model optimization, or knowledge transfer, not for simple one-off tasks. Commands let you train patterns, check status, predict outcomes, and optimize for targets like latency. The performance numbers are aggressive (sub-0.05ms for SONA, 2.49x to 7.47x speedup with Flash Attention) and the approach assumes you're doing ongoing learning where consolidation matters. Worth it if you're building adaptive agents that need to learn and remember.