Use the existing `*.OpenAI_*.log` file as the source artifact, replay every captured raw response row in order, and inspect the exact completion output the current backend parser produces now.