integration task: bisect

once the failures is reproducible we should use `bisect` to try to detect when it starts to give more context to the LLM

see https://github.com/huggingface/transformers/pull/47937

Start by reproducing the failure described in the issue, then read the linked Transformers pull request for relevant integration context. Identify the repository entry point that runs the failing task and determine how bisect results should provide additional context to the LLM; done means the reproducible failure can be bisected and that context is emitted for the LLM.

Written by the indexing model from the issue text.