docs: trim cookbook content and widen evaluator conciseness range for v2

- Trim composite and dataset-aware recipes (~4K chars removed)
- Tighten score-range and task-model sections
- Widen cookbook evaluator sweet spot: 10K→18K, hard limit 20K→30K
- All score baselines passing (cookbook: 0.9088 vs 0.9002 baseline)

Author: cellarius