v0.2.0

Model-quality release: better generalization and weak classes, calibrated confidence, and a retrain on all 15 subjects. No API contract change (512-sample window, 8 classes).

Added
Train-only data augmentation (jitter / scaling / time-shift / magnitude-warp), pure numpy in preprocessing.py, applied to the train split only (val/test untouched).
Config-driven regularization: encoder Dropout1d + head dropout via ModelConfig.
Cosine LR schedule + early-stopping (patience) in training.
Temperature-scaling confidence calibration: fit on validation (fit_temperature, torch-free), stored in the checkpoint, applied at inference in predict.py (softmax(logits / T); backward-compatible, defaults to T=1.0).

Changed
Retrained on all 15 subjects (S6 included); regenerated model/metrics/*.json.
Multimodal test accuracy 0.650 → 0.776; walking F1 0.36 → 0.81. The old overfitting gap closed — best validation now lands at epoch ~10 and test ≈ validation.
ECG-only baseline test 0.371 → 0.609 on the same pipeline.

Honest note
working is now the weakest class (test F1 0.47, confused with lunch_break); with only 2 validation + 2 test subjects the metrics are noisy (val/test ordering can flip).