[Bug] Resume training results in very high rollout logprobs diff and very low pass@1/raw_reward for a few steps

Bug Description

Steps to Reproduce

1. train qwen-30b-a3b-thinking-2507
2. load checkpoint:

CKPT_ARGS=(
   --hf-checkpoint /tmp/instance_storage/${MODEL_CONFIG}/
   --ref-load /tmp/instance_storage/${MODEL_CONFIG}_torch_dist/
   --load /tmp/instance_storage/checkpoints/${MODEL_CONFIG}_slime/
   --save /tmp/instance_storage/checkpoints/${MODEL_CONFIG}_slime/
   --save-interval 20
)

Expected Behavior

no numerical error

Actual Behavior

very high log probs diff / very low pass@1

Environment

• slime version: c25982336374bcdcb44635410771bd78166684c1
• Python version: 3.12
• PyTorch version: 2.9.1
• CUDA/ROCm version:
• GPU type and count: 24xH200
• OS:
• SGLang version (if relevant):
• Megatron-LM version (if relevant):

Logs

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and my issue is not addressed there.
• I have searched for existing issues and this is not a duplicate.
• I have provided a minimal, reproducible example.

Start by reproducing the resume flow with the provided checkpoint arguments and compare rollout logprobs, pass@1, and raw_reward before and after loading the checkpoint. The issue is resolved when resumed training no longer produces the reported numerical discrepancy or temporary metric collapse.