Neural Train

This implements the DISTILL and CONSOLIDATE phases of a 4-step neural pattern pipeline, letting you capture what worked after successful task completions and fold them into long-term storage. The flow is trajectory tracking (start, step by step, end with pass/fail) followed by neural_train to crystallize patterns. SONA handles single-domain micro-adaptation under 0.05ms, MicroLoRA kicks in when you have three or more distinct domains to avoid overloading one adapter. The consolidate phase uses EWC++ to prevent catastrophic forgetting when you fold in new patterns. It's very specific about namespaces and phases, which means you need to understand the broader pipeline to use this effectively. Good if you're building agents that actually learn from their own execution history.