Avoid overflow when summing probability-selection fitness

[ktz03]

Roulette-wheel (rws) and stochastic-universal (sus) selection reduce fitness in its original NumPy dtype. Four positive numpy.int64(2**62) fitness values therefore wrap their sum to zero, and GA.run() raises ZeroDivisionError. Large positive uint64, float16 and float32 values also overflow these reductions; multi-objective fitness can overflow before its row sums are normalized.

Use an accumulator dtype of numpy.result_type(fitness.dtype, float) for both reductions in each selector. This promotes native integer and narrow floating arrays while preserving object arrays containing arbitrary-size Python integers. The existing exception for a genuine zero fitness sum remains unchanged.

Validation on Windows, Python 3.12.14 and NumPy 2.5.4:

• The parent-selection regression file has 117 tests: baseline 16 failures/101 passes, fixed 117 passes. New coverage includes both selectors, single/multiple objectives, four numeric dtypes, genuine zero sums and arbitrary-size Python integers.
• All eight files used by the Windows CI test step pass locally: 582 tests, with one existing warning that crossover and mutation are disabled. These were run from the source checkout; a wheel install and the complete CI matrix were not run.
• A separate public GA.run() script changes from 16 failures/20 normal controls to 36 completed calls; every previously normal row, selected-parent index and caller input remains identical.
• git diff --check passes.

No full repository, other Python/NumPy versions, other platforms, model training or performance validation is claimed.

[github-actions]

📊 Change summary

2 files changed · +66 / −4 lines

pie showData
    title Lines changed per file (70 total)
    "tests/test_parent_selection_regressions.py" : 62
    "pygad/utils/parent_selection.py" : 8

Loading

Per-file breakdown

File	+	−	total
tests/test_parent_selection_regressions.py	62	0	62
pygad/utils/parent_selection.py	4	4	8
Total (2 files)	66	4	70