Summary

• Adds configurable checkpoint retention to PipelineTrainer
• Introduces keep_recent_and_top(...), where strategies return eligible checkpoint steps to keep
• Protects current, leased, and scheduled-eval checkpoints from deletion
• Records checkpoint retention metadata in history.jsonl without changing normal metric routing
• Aligns Unsloth loaded-LoRA adapter pruning with Megatron so both unload non-retained vLLM adapters

Validation

• ruff check
• ty check
• git diff --check
• pytest tests/unit/test_metric_routing.py tests/unit/test_checkpoint_retention.py tests/unit/test_pipeline_trainer_local_backend.py tests/unit/test_multi_checkpoint_inference.py
• scratch smoke test for actual local checkpoint dir deletion and lease protection