Improve deduplication for function-calling training data

Goal

Reduce overexposure from function-calling data generated from the same seeds while preserving useful alternative trajectories.

• Quantify repeated training examples produced from the same seed task or conversation, including variants that survive ordinary duplicate checks.
• Reduce overrepresentation from high-yield seeds while retaining valid alternative executions, recoveries, and outcomes.
• Compare seed-level capping, selection, or weighting policies against the current mixture at matched training exposure.