GeneralisedFilters: Stratified resampler is exported but has no CPU method

Stratified is exported from GeneralisedFilters, but sample_ancestors has no method for a plain Vector, so any CPU use of it fails with a MethodError.

Reproducer

using Random, GeneralisedFilters
using GeneralisedFilters: Stratified, Systematic, Multinomial, Metropolis, Rejection,
    sample_ancestors

w = [0.25, 0.25, 0.25, 0.25]
sample_ancestors(Random.default_rng(), Stratified(), w)

ERROR: MethodError: no method matching sample_ancestors(::TaskLocalRNG, ::Stratified, ::Vector{Float64})

The other four exported resamplers all work on the same input:

It surfaces the same way through a filter, so BF(64; resampler = Stratified()) fails while the other four return a log-likelihood.

Cause

Stratified is defined as a bare struct, and its only sample_ancestors and sample_offspring methods are typed on CuVector. On 0.4.2 those sit at src/resamplers.jl:281-292; on 0.5.0 the struct is at src/resamplers.jl:240 and the CuVector methods have moved to ext/CUDAExt.jl:61-69. Either way there is no CPU implementation, so the exported name only works on GPU arrays.

Version

GeneralisedFilters 0.4.2, Julia 1.12, macOS. Read from the source, 0.5.0 looks the same in this respect, though I have not run it.

This was opened by a bot. Please ping @seabbs for any questions.

Start by running the direct sample_ancestors reproducer and inspect the Stratified definitions in src/resamplers.jl and ext/CUDAExt.jl. Compare its CPU behavior with the other resamplers and verify the fix through BF(64; resampler = Stratified()) as well as the direct Vector call. Done means both return successfully on CPU while existing CuVector behavior remains intact.