Fine-tuned intent classifier v1

DistilBERT fine-tuned on the bundled 14-intent synthetic ticket corpus, exported to ONNX and dynamically quantised to int8.

• base model: distilbert-base-uncased
• trained on 1,916 records, validated on 336, 6 epochs with early stopping on validation macro-F1, seed 20260831
• validation macro-F1 0.7353, accuracy 0.7440
• 2,887.8 s of CPU training (Linux x86_64, 8 threads)
• 67 MB int8, max sequence length 160

The finetuned-encoder arm downloads these three files on a cache miss. Reproduce the whole thing with just finetune.