Prevent hardcoded 'cpu' device - use config file

[rogermt]

Problem

The YOLO plugin has 70+ occurrences of hardcoded "cpu" as the default device parameter. This causes the plugin to run on CPU even when the config file (models.yaml) specifies device: "cuda".

Locations affected:

• plugin.py - Tool function defaults, manifest schema defaults
• inference/*.py - All detection/tracking functions
• video/*.py - All video processing functions
• utils/team.py - Team classification

Current workaround:

The run_tool() method now reads from config, but if someone calls the underlying functions directly, they'll still get CPU.

Proposed Solution

1. Add a CI check to prevent new hardcoded "cpu" defaults:

   • Create a script that scans for device: str = "cpu" patterns
   • Fail CI if found outside of config fallback

2. Refactor to use config throughout:

   • Create a get_default_device() helper in configs/__init__.py
   • Replace all device: str = "cpu" with device: str = get_default_device()

3. Document the pattern:

   • Add to plugin development guide
   • Ensure all plugins follow the same pattern

Impact

• GPU acceleration not working by default on GPU-enabled systems
• Performance degradation on Kaggle/Colab where GPU is expected
• Confusing UX when config file is ignored

Priority

High - Affects core functionality (GPU vs CPU inference)