Fix: wrap as Qwen3_5ForConditionalGeneration + model.language_model.* prefix (drop-in for vLLM --language-model-only)

Replace MLX 4-bit with BF16 merged model (SFT best@6k, val loss 0.015). Compatible with vLLM/transformers.

Intent Classifier 4B — SFT best@6k (val loss 0.015, ~99% accuracy). Qwen3.5-4B base, 13 intents, 5 languages.