feat(training): LoRA sentiment + emotion pipelines complete

- Sentiment LoRA (FinBERT): 5 epochs, early stopping, 6.2MB adapter
- Emotion LoRA (DistilRoBERTa): 5 epochs, weighted loss, 8.1MB adapter
- Both with early stopping (patience=3, threshold=0.001)
- Data augmentation: templates + crypto slang
- Human-in-the-loop verification CLI
- Disk-conscious: save_total_limit=1, ~6-8MB each

sentiment_engine/training/lora_pipeline.py

@@ -24,7 +24,7 @@ import hashlib
 # ============================================================
 try:
     import torch
-    from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
+    from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer, EarlyStoppingCallback
     from peft import LoraConfig, get_peft_model, TaskType, PeftModel
     from datasets import Dataset
     TORCH_AVAILABLE = True
@@ -381,6 +381,7 @@ def train_lora(train_data: List[Dict], val_data: List[Dict]) -> str:
         }
     
     trainer = Trainer(
+        callbacks=[EarlyStoppingCallback(early_stopping_patience=3, early_stopping_threshold=0.001)],
         model=model,
         args=training_args,
         train_dataset=train_ds,