Module replacement 

Survivor-selection / replacement operators.

Each function takes the current generation’s population and fitness plus the offspring population and fitness, and returns the (population, fitness) pair that becomes the next generation’s starting state.

All selection logic operates on host-side &[f32] fitness slices (lower is better), and winners are lifted back to the device via a single Tensor::select gather. None of these functions draw random numbers; they are deterministic given their inputs.

Fitness convention

Fitness is treated as a cost: smaller values are better. This matches the convention used throughout crate::ops::selection.

Choosing a replacement strategy

Strategy	Parent survival	Use when
generational	none	offspring quality is trusted; GA / CMA-ES
elitist	top-k	preserving known-good solutions matters
mu_plus_lambda	best of μ+λ pool	ES / DE with strong elitism
mu_comma_lambda	none (offspring only)	ES with deliberate age-based forgetting

Functions

elitist

Elitist replacement: keeps the k best parents and the best remaining offspring.

generational

Replaces the entire current generation with the offspring (no elitism).

mu_comma_lambda

(μ, λ) replacement: discards parents and keeps the μ best offspring.

mu_plus_lambda

(μ + λ) replacement: keeps the μ best individuals from the merged parent and offspring pool.