* fix(genetic): sum the self-consumption dot products in a fixed order
GENETIC gave different plans for the same input and seed on different
CPUs. SelfConsumptionProbabilityInterpolator computed its two dot
products with np.dot, which calls BLAS; OpenBLAS picks a ddot kernel per
CPU at run time (and macOS uses Accelerate), and the kernels sum in
different orders, so the result differs in the last bit. The optimizer
compares fitness values exactly, so one bit is enough for a different
plan.
fixed_order_dot() adds the rounded products in index order with Python
floats, without BLAS: the same result on every CPU. The vectors have one
entry per 50 W load bin and the results are cached per (load, PV) pair,
so the run time does not change noticeably. Plans change once in the
last bits compared with the previous np.dot result of a given CPU.
* fix(genetic): sort equal prices of the terminal value window in slot order
np.argsort with the default kind (quicksort) leaves the order of equal
values to the platform: NumPy sorts with SIMD kernels on x86-64
(x86-simd-sort) and with introsort elsewhere. Slots with the same price
then fill the terminal value curve in another order, the battery
capacity cuts it elsewhere, and the same input and seed gave another
plan on x86-64 than on aarch64 (seen as a native-vs-Python compare
difference on x86-64 in
test_plan_keeps_reserve_for_a_peak_beyond_the_control_horizon).
The window is now sorted with kind="stable": slots with the same price
keep their slot order on every platform. Same goal as the dot product
change in this PR: the same input gives the same plan on every CPU.
Assisted-by: Claude Code (claude-opus-5-5)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
b4d2d7b2aa