Add Beam serverless-GPU captioner (Qwen2.5-VL) with batching

Surface Generate errors in the guide panel; raise LLM max_tokens to 4096

Auto-extract: generate the LLM step outline first, anchor frames to step times

Make auto-frame extraction narration-accurate

DocuMaker: video to step-by-step DOCX guide (Whisper + HF LLM + BLIP)