Add Ministral3 DeepSpec reproduction experiments

[andylolu2]

Why

This experiment checks whether the DeepSpec Eagle3 and DSpark recipes reproduce paper-level accepted-length results when adapted to mistralai/Ministral-3-3B-Instruct-2512. The reports are intentionally included even though the result is negative: the implementation and eval pipeline work, but the Ministral3 runs do not match the Qwen3-4B reference behavior.

Summary

• Add Ministral3 Eagle3 model/trainer/evaluator support and 10k Open-PerfectBlend reproduction scripts.
• Add save-only checkpoints, incremental eval JSON writing, split eval task launching, and plotting utilities.
• Add DSpark block-7 Ministral3 follow-up launchers, arena-only rescue eval launchers, and a validated eval JSON merge helper.
• Add final markdown reports with training curves and accepted-length comparison plots.

Reviewer guide

• Real logic changes: deepspec/modeling/eagle3/ministral3, deepspec/trainer/eagle3_trainer.py, deepspec/eval/base_evaluator.py, eval.py, and checkpoint handling in deepspec/trainer/base_trainer.py.
• Experiment/config changes: config/eagle3/eagle3_ministral3_3b.py, experiments/ministral3_3b_eagle3/*.sbatch, and experiments/ministral3_3b_followup/*.sbatch.
• New follow-up utility: experiments/ministral3_3b_followup/merge_eval_json.py validates and merges a single-task rescue eval into a complete aggregate JSON.
• Report artifacts: experiments/ministral3_3b_eagle3/report.md covers the Eagle3 baseline reproduction; experiments/ministral3_3b_followup/README.md covers the completed DSpark follow-up. These reports are the best entry points for reviewing the outcome.

Test plan

• Ran Python syntax checks for eval.py, the follow-up plotting/merge helpers, and shell syntax checks for split eval sbatch scripts.
• Ran smoke data/cache/train jobs for Ministral3 support before full runs.
• Completed the 3000-step DeepSpec Eagle3 training run and final 9-task eval. Final eval JSON: /mnt/vast/runs/andy/deepspec_ministral3_3b_eagle3/eval_10k_step3000_metrics.json.
• Completed the 3000-step DSpark block-7 training run and final 9-task eval. Final eval JSON: /mnt/vast/runs/andy/deepspec_ministral3_3b_followup/eval_dspark_8gpu_step3000_metrics.json.
• Re-ran arena-hard-v2 as a one-task rescue eval after the full DSpark eval timed out at 8/9 rows, then validated and merged it with merge_eval_json.py.
• Regenerated report plots from TensorBoard and the completed eval JSONs.