feat: Switch V4 to Qwen2.5-1.5B for HF memory compatibility

- Replace Phi-3-mini (3.8B) with Qwen2.5-1.5B-Instruct
- Phi-3-mini still exceeded 16GB limit even with V2 disabled
- Qwen2.5-1.5B: ~3GB RAM (comfortable within 16GB)
- 3x larger than 0.5B version (much better quality)
- Good structured output and NDJSON generation capability
- Memory: V4: 3GB + system: 2GB = ~5GB (plenty of headroom)

app/core/config.py

@@ -107,9 +107,9 @@ class Settings(BaseSettings):
         description="Enable V4 model warmup on startup (uses 1-2GB RAM with quantization)",
     )
     v4_model_id: str = Field(
-        default="microsoft/Phi-3-mini-4k-instruct",
+        default="Qwen/Qwen2.5-1.5B-Instruct",
         env="V4_MODEL_ID",
-        description="Model ID for V4 structured output (3.8B params, optimized for instruction following)",
+        description="Model ID for V4 structured output (1.5B params, fits HF 16GB limit)",
     )
     v4_max_tokens: int = Field(
         default=1024, env="V4_MAX_TOKENS", ge=128, le=2048, description="Max tokens for V4 generation"