fix: truncate prompt to 500 chars to prevent Gemma token overflow (1087>1024)

The Gemma prompt enhancer was crashing on long prompts, producing
garbage like '/imagine synchronized lipsync.' which caused the model
to generate nonsensical text images instead of video content.

app.py

@@ -999,6 +999,10 @@ def generate_video(
         tiling_config = TilingConfig.default()
         video_chunks_number = get_video_chunks_number(num_frames, tiling_config)
 
+        # Truncate prompt to prevent Gemma token overflow (max 1024 tokens ≈ 500 chars)
+        if len(prompt) > 500:
+            prompt = prompt[:500]
+
         video, audio = pipeline(
             prompt=prompt,
             seed=current_seed,