Fix AIRadar packaging: add missing README and declare real dependencies

[egeboy35]

Hi! While setting the project up from the root README's Quick Start, step 3 (flit install --symlink in AIRadar/) failed on a fresh clone:

Config error: Description file README.md does not exist

AIRadar/pyproject.toml references readme = "README.md", but the file was not present. The package also declared only numpy as a dependency, while the AIRadarLib modules import scipy, matplotlib, torch, einops, and sklearn — so even a successful install would produce import errors at use time.

Changes

• Add AIRadar/README.md: a short overview of the AIRadarLib modules plus install instructions.
• pyproject.toml: declare the dependencies AIRadarLib actually imports, and expose the TensorFlow/Sionna-based helpers as an optional [sionna] extra (they are only used behind lazy imports).
• Give AIRadarLib/__init__.py a package docstring.

Verification

In a clean venv (Python 3.14):

• Before: python -m flit build → Config error: Description file README.md does not exist
• After: python -m flit build → Built sdist: dist/airadarlib-0.1.0.tar.gz + Built wheel: dist/airadarlib-0.1.0-py3-none-any.whl
• pip install -e . (PEP 660) succeeds and import AIRadarLib works.

Happy to adjust the README wording or the dependency split if you prefer a different structure.

🤖 Generated with Claude Code

[egeboy35]

One consequence of this PR I should have spelled out, since it changes what your own documented setup step does.

Adding torch to [project] dependencies makes it a hard install requirement, so the README's flit install --symlink now pulls a multi-gigabyte wheel — including for someone who only wants the classical DSP helpers. Looking at what actually imports what, the split is clean:

• numpy / scipy / matplotlib only: signal_processing, waveform_utils, channel_simulation, datautil, target_utils, AIradar_autopara, visualization
• needs torch (plus einops / scikit-learn): modeling_RadarNet, modeling_TimeNet, modeling_transformer, ofdm_decoder, pretrain_dataset, radar_det

So an alternative shape would be core dependencies numpy, scipy, matplotlib with the model side behind an extra, e.g. pip install -e ".[train]" for torch, einops, scikit-learn, keeping the existing [sionna] extra as is. That makes a bare install light and fast, at the cost of one more thing to remember when training.

I deliberately did not do that in this PR, because it changes the install contract rather than just fixing the broken one, and I would rather you decide which you prefer. Happy to push either version — or to leave it exactly as it is if a single install command matters more to you than install size.

Separately, requires-python is still >=3.8 from before this PR; recent torch releases no longer publish 3.8 wheels, so that bound is optimistic now regardless of how the dependencies are grouped.

🤖 Generated with Claude Code