Unify node categories to AutoPromptEngineer, add Qwen3‑VL nodes, and fix JS registration

[ifilipis]

Motivation

• Consolidate node registration so all nodes appear under a single ComfyUI category for discoverability and consistency.
• Correct a miscategorization where the Director front-end registration did not match the unified category and caused UI/JS mismatch.
• Add first-class Qwen3‑VL support so users can load and run multimodal Qwen3‑VL checkpoints from the UI.
• Record model configuration and declare transformer-related dependencies required by the new nodes.

Description

• Changed node CATEGORY values (e.g. DirectorGemini, ImageRouterSink, LatestImageSource, and Qwen nodes) to AutoPromptEngineer and aligned front-end registration by updating the LiteGraph.registerNodeType call in director_actor_executor.js to AutoPromptEngineer/DirectorActorExecutor.
• Added qwen3_vl_nodes.py implementing Qwen3VL and Qwen3VLAdvanced nodes and qwen3_vl_models.json containing model entries and VRAM hints.
• Updated __init__.py to merge node mappings from the director and Qwen modules into NODE_CLASS_MAPPINGS and NODE_DISPLAY_NAME_MAPPINGS, and updated AGENTS.md to reflect the shared category.
• Added transformer stack requirements to requirements.txt by declaring transformers and huggingface_hub.

Testing

• No automated tests were executed for these changes.
• Basic repository-wide edits were validated by searching and updating matching CATEGORY occurrences and the JS registration string to ensure UI mapping consistency.
• The new model config file is loaded by the Qwen node loader fallback, and the node code includes runtime checks to require transformers/huggingface_hub when used.
• Manual inspection of modified files confirmed the files added or updated (qwen3_vl_nodes.py, qwen3_vl_models.json, nodes.py, director_actor_executor.js, __init__.py, AGENTS.md, requirements.txt).

---

Codex Task