Sprint 10C: lora_plan.py + distill.py — adapter/distillation planning (dry-run, no GPU required)

Sprint 10B: optimizer.py — agentic decision policy (improving/plateau/degrading/rollback)

Sprint 10A: shadow_eval.py — compare candidate vs baseline safely

Sprint 9C: prompt_pack.py — epigenetic optimization (prompts before weights)

Sprint 9B: dataset.py — trace → filtered training dataset builder

Sprint 9A: fingerprint.py — capability fingerprint from execution traces

Sprint 9-10: optimization package init