v0.1.0

Model

• Replaced learnable channel-wise scaling with a fixed constant.
• Added an exponential moving average (EMA) of model weights during training.
• Changed the final profile convolution to kernel_width=1.
• Set the default model size to 96 filters.
• Tuned Muon and AdamW learning rates and weight decay values.
• Best-model selection now monitors validation profile loss rather than total validation loss.

Training

• Default max_jitter for fitting lowered from 500 to 50 (CLI fitting defaults).
• Library PeakGenerator default max_jitter lowered from 128 to 50.

Packaging and tooling

• Migrated from setup.py to pyproject.toml with uv support.
• Refactored the CLI from a monolithic script into the cherimoya_cli modular package.
• Raised the minimum Python version to 3.10 and minimum PyTorch to 2.9.
• Added macs3, bam2bw, bpnet-lite, triton, and joblib as dependencies.
• Added a Sphinx documentation site hosted on Read the Docs.