Generate a task-specific `prepare.py` (read-only after this point) that parses raw data, builds vocab/labels/splits/cache, and emits an EDA report. Reads `brief/parsed/data_spec.json` and `task_type.json` to pick the right template (CV / NLP / audio / RL / tabular / multimodal). Use after workspace-init when raw data is available (locally or on remote machines). Triggers when the user has data ready and asks to prepare / preprocess / set up data loading.
