Revert to Whisper lazy loading - build OOM persists even with /data. First request downloads to persistent storage.

Dockerfile

@@ -37,15 +37,11 @@ RUN python -c "from transformers import AutoTokenizer, AutoModelForSequenceClass
     AutoModelForSequenceClassification.from_pretrained('Cyberlace/swara-structure-model', cache_dir='/.cache'); \
     print('✅ Structure Model cached!')" && chmod -R 777 /.cache
 
-# 2. Download Whisper medium model (~1.5GB) 
-# Using /data for HF Pro Persistent Storage (survives restarts)
-RUN mkdir -p /data/.cache && \
-    python -c "import whisper, os; \
-    os.environ['TORCH_HOME'] = '/data/.cache'; \
-    print('📥 Downloading Whisper medium to persistent storage...'); \
-    whisper.load_model('medium', download_root='/data/.cache'); \
-    print('✅ Whisper medium cached!')" && \
-    chmod -R 777 /data/.cache
+# 2. Whisper medium: LAZY LOADING on first request
+# Build OOM - HF Space build container has RAM limit
+# Will download to /data/.cache on FIRST REQUEST (~2-3 min)
+# With HF Pro persistent storage, download persists across restarts
+# Subsequent requests will be fast using cached model
 
 # 3. Download Sentence Transformer for Keywords (~420MB)
 RUN python -c "from sentence_transformers import SentenceTransformer; \