Replaces the v0.10 ML model with a properly-trained version.

- balanced_acc: 0.55 -> 0.81
- specificity: 4.5% -> 80% (20x fewer false positives)
- threshold: 0.85 (hack) -> 0.5 (natural)
- architecture: custom CNN -> ResNet-18 pretrained
- feature sample rate: 22050 -> 44100 (CRITICAL: was erasing the MP3 signal)