[ml31415]

simple.py is gone

It printed three usage examples, checked nine functions against a group-and-loop reference and timed those nine over 100,000 values. All three are covered elsewhere: the examples are in the README, the comparison is test_compare.py (over every implementation pair and far more functions), and the timing is this module's table (40 functions, --size 1e5 gives the same data set). It also still went through the aggregate_np/aggregate_py aliases.

The one thing worth keeping was its np-grouploop column: the same numpy implementation with its fast path disabled, i.e. what a func the implementation does not recognise costs. generic.py still carried the helper for it, unused since the rework - it is now _grouploop.py and an opt-in column, in line with _pandas.py and _ufunc.py:

function              ufunc         numpy     grouploop         numba
---------------------------------------------------------------------
sum                   0.130         0.094         4.259         0.060
arbitrary              ----         7.092         7.158         1.705
sort                   ----         4.216          ----         0.593

---- where no per-group callable can express the function (the cumulatives, the argreductions, sort); arbitrary is a custom callable and lands on the group-loop path by itself, which is why numpy and grouploop agree there.

The entry point

generic.py is the package __init__.py and a five-line __main__.py runs it, because python -m on a package executes __main__.py and not the __init__.py:

python -m numpy_groupies.benchmarks --pandas --numbagg
uv run -m numpy_groupies.benchmarks --sort --runs=3

The --help header, the README and the two commands in docs/benchmarks.md follow the new name; the --grouploop column is documented next to the other opt-in ones.