Token-budget windows: pack by Qwen3.5 tokenizer counts (6400/window), 24202 windows, max templated len ~7.1K

Plan B windowed line-anchored dataset (4407 docs -> 20443 windows, same splits as legacy)

initial commit