Replace silent stub fallback in InferenceEngine with a tracing-aware
loader:
- emit `tracing::info!` when pose_v1.safetensors loads into the candle
backend, including the resolved on-disk path
- emit `tracing::warn!` (instead of silently falling through) when the
configured weights file is missing or no path is set at all, so the
appliance log surfaces "no model — running stub" instead of just
emitting confidence=0 frames forever
- inline unit tests covering both branches: stub fallback shape +
confidence, and a real-weights forward pass that asserts the output
decomposes into 17 (x, y) pairs all finite and in [0, 1]
The forward pass itself already matched the training script
(Conv1d 56->64->128->128 with dilations [1, 2, 4], GlobalMeanPool,
Linear 128->256->34, sigmoid) and the tensor names in the
safetensors file (`enc.c1/c2/c3.weight|bias`, `head.fc1/fc2.weight|bias`)
line up with the VarBuilder prefixes — no architecture changes
required, only loader hygiene.
Confidence is the published validation PCK@50 (0.185); v0.0.1 doesn't
emit per-frame confidence and we're not fudging that here.
339d9d70dc