DEVLOG #19 - New lessons can overwrite old ones

A continual-learning system like SL-LLM-R can forget older knowledge if it trains only on the newest mistake. I use replay, which means mixing some older verified examples back in while a new lesson is learned. The goal is to make new learning fit alongside old knowledge instead of pushing it out.