WasmAgent turns agent runs into training data without human labeling

WasmAgent's compliance engine captures every agent run, evaluates it, and exports a typed ComplianceEvalRecord ready for SFT or DPO training. In tests with Qwen2.5-1.5B, full_pcl mode achieved 54.7% pass rate, 8.7 points higher than prompt_retry. The system records repair traces, turning failures into learning data automatically.