Make PassthroughRNG dispatch survive overlay method-table shadowing

[ChrisRackauckas-Claude]

Closes the root cause behind SciML/JumpProcesses.jl#588.

Summary

PassthroughRNG previously defined only the three no-second-arg methods (rand, randexp, randn). On Julia 1.12+, GPU back ends like CUDA.jl install device-side overlay tables via Base.Experimental.@consistent_overlay. Julia's OverlayMethodTable.findall (Compiler/src/methodtable.jl:73–92) returns overlay matches without consulting the base method table whenever the overlay fully covers the signature:

if nr ≥ 1 && result[nr].fully_covers
    # no need to fall back to the internal method table
    return result
end

CUDA.jl's @device_override Random.randexp(rng::AbstractRNG) therefore shadows our specific Random.randexp(::PassthroughRNG) on the device — the more specific base-table method is never seen by inference. The override's body then runs with rng::PassthroughRNG and calls Random.rand(rng, UInt52Raw()). The stdlib Sampler chain for that bottoms out at _rand52(r, rng_native_52(r)) → rand(r, UInt64); PassthroughRNG had no rng_native_52 and no typed-arg rand, so the chain statically reached throw(MethodError, ...) and GPUCompiler refused to lower the kernel:

InvalidIRError: ... resulted in invalid LLVM IR
Reason: unsupported call to an unknown function (call to jl_f_throw_methoderror)
Stacktrace:
 [1] rand        @ Random/src/generation.jl:114
 [2] rand        @ Random/src/Random.jl:255
 [3] randexp     @ CUDACore/src/device/random.jl:339   (device_override)

Confirmed on Julia 1.12.6 by directly calling Random.rand(PassthroughRNG(), Random.UInt52Raw()) on the CPU — the top two frames match the GPU InvalidIRError frames exactly.

Fix

Two forwarding methods so the Sampler chain bottoms out at bare rand(T):

Random.rng_native_52(::PassthroughRNG) = UInt64
Random.rand(rng::PassthroughRNG, ::Type{T}) where {T} = rand(T)

These preserve PassthroughRNG's ''use whatever default_rng() returns here'' semantics — bare rand(T) goes through default_rng(), which GPU back ends device-override to their device RNG (Philox2x32 in CUDA.jl). Verified locally on Julia 1.12.6:

rng_native_52(prng):  UInt64
rand(prng, UInt52Raw()):  2033734290634309481
rand(prng, UInt64):  11777094854860645014
rand(prng, Float32):  0.98086137
rand(prng, Float64):  0.9338644122927825
pois_rand(prng, 3.0):  3
pois_rand(prng, 50.0):  52

Added a regression test (@testset "PassthroughRNG dispatch") covering each of these calls. Bumped version 0.4.7 → 0.4.8.

Test plan

• Full Pkg.test() green on Julia 1.12.6 — Aqua (10/10), JET static analysis (6/6), ExplicitImports (2/2), BigFloat (401/401), PassthroughRNG dispatch (5/5), Allocation Tests (10/10), and the count/ad/mixed statistical samplers.
• CI green.

Notes

• Please ignore until reviewed by @ChrisRackauckas.
• This doesn't pull in a dependency on RandomNumbers.jl (alternative would have been to make PassthroughRNG <: RandomNumbers.AbstractRNG{UInt64} and inherit the rng_native_52 fallback — heavier and changes the package's dep graph).
• There is a separate, downstream CUDA.jl issue exposed once this PR unblocks the dispatch chain: @noinline randexp_unlikely in the device randexp override (line 345) triggers julia.get_pgcstack on Julia 1.12 — that needs an upstream fix in CUDA.jl and is not in scope here.

🤖 Generated with Claude Code