TRL 的 SFTTrainer 一行就能训，为什么还有人手写 SFT 循环？我把 forge 仓库 Stage 1 的每个决策（chat template、prompt masking、collate 与 loss、bf16、LR schedule、数据、验证）逐一摆到台面上，对照两套教材（rlhf-book、hands-on-modern-rl）、四个生产框架（alignment-handbook、trl、open-instruct、OpenRLHF）和六路信息源：哪些决策有共识，哪些有真实的配方之争，哪些两本教材各自糊弄过去了。
A note on RAG and context engineering: retrieval quality, evidence shape, memory boundaries, and why context is a product surface.
A note on evaluation as an instrument: failure cases, metrics, benchmark design, product loops, and the discipline of measuring agents.
A field note on designing agents as observable loops, with tools, memory, failure recovery, and product boundaries.