Distill training pairs for summarizer and explainer

[William-Hill]

Summary

Generate training pairs for the summarizer and explainer tasks via Claude API distillation. These are the two existing task types in the pipeline.

Depends On

• Merge distillation-pipeline and shap-explainability to main #96

Tasks

• Run python -m training.distill --school bishop-state --task summarizer (~1,500 pairs)
• Run python -m training.distill --school bishop-state --task explainer (~1,500 pairs)
• Run python -m training.prepare --school bishop-state (validate, dedup, 80/10/10 split)
• Verify pair quality: spot-check 10 pairs per task for correctness
• Commit training data to training_data/bishop-state/ (JSONL files)
• Track distillation cost (expected: $3-5 total)

Acceptance Criteria

• = 1,200 validated pairs per task after dedup

• Train/val/test splits at training_data/bishop-state/final/{task}/{split}.jsonl
• All pairs are valid JSON with correct schema