docs(train): L1–L3 learning model design (#94)

[loootte]

Summary

#94 design doc: L1–L3 supervised OMR layout model plan (parent #92 task 2).

Decisions

• Path A (selected): lightweight CNN + L1/L2 detection heads + L3 1D heatmap on L2 crops
• MVP-1 (feat(train): 训练 Framework 骨架与 toy 训练 #95): L2 systems + L3 interior splits first; L1 optional later
• Metrics: line-level split count + mean abs x error (aligned with refactor(core/desktop): L3 改为「L2 行内纵向分割线」模型（识别切分 + 拖线编辑 + 线级评估） #85/Feature: 为流水线各节点搭建分层准确度量化评测工具，配套参数调优、误差定位与可视化分析 #86); measure IoU secondary only
• Integration: map to existing structure.items / barlines / IR; rule pipeline remains fallback

Files

• docs/train/l1-l3-model-design.md
• links from data-spec / architecture / core README

Closes #94