Fix Dockerfile: replace HHEM pre-warm with nli-deberta-v3-small

HHEM checkpoint is broken (zero embeddings) — replaced in grader.py last
session. Dockerfile still tried to download it, failing the build.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Dockerfile

@@ -12,12 +12,9 @@ RUN pip install --no-cache-dir -r requirements.txt
 
 # Pre-download models so first request isn't slow on HF Spaces
 RUN python -c "\
-from sentence_transformers import SentenceTransformer; \
-from transformers import T5Tokenizer, pipeline; \
+from sentence_transformers import SentenceTransformer, CrossEncoder; \
 SentenceTransformer('all-MiniLM-L6-v2'); \
-tok = T5Tokenizer.from_pretrained('t5-small'); \
-pipe = pipeline('text-classification', model='vectara/hallucination_evaluation_model', tokenizer=tok, trust_remote_code=True); \
-pipe(['test document test response'])"
+CrossEncoder('cross-encoder/nli-deberta-v3-small')"
 
 COPY knowledge/ ./knowledge/
 COPY backend/   ./backend/