Flaky: reductions of host-accessible arrays return 0 on Julia 1.13 (ALCF CI) #661

The reductions of host-accessible arrays testset added in #649 fails intermittently on the ALCF (Aurora, LTS stack) CI with Julia 1.13:

Error in testset reductions of host-accessible arrays:
Test Failed at test/array.jl:116
  Expression: sum(a) == 1024
   Evaluated: 0.0f0 == 1024
Error in testset reductions of host-accessible arrays:
Test Failed at test/array.jl:117
  Expression: maximum(a) == 1
   Evaluated: 0.0f0 == 1

So sum/maximum sometimes still return 0 on a SharedBuffer/HostBuffer-backed array: the symptom #649 was meant to fix. Two of the four tests in the loop fail, i.e. presumably both tests for one of the two buffer types (the output doesn't say which).

Occurrences

All on test:lts:1.13; the corresponding test:lts:1.10 jobs passed every time:

#655 only touches accumulate/cumsum/cumprod, which sum/maximum don't use, and it resolves the same GPUArrays (11.5.15) and KernelAbstractions (0.9.43) as the passing main run. So this looks like a race, not a regression from any one change. It first appears on the branch that introduced cooperative synchronization (#654).

Environment: Julia 1.13.1, NEO 26.18.38308, IGC 2.34.4, Intel Data Center GPU Max 1550, test workers spread over 12 GPUs (ONEAPI_TEST_SPREAD_GPUS=1), run with --quickfail.

Notes

• This is likely another instance of the host-visibility problem described in Track outstanding GPU work per allocation for host-accessible arrays #652: the CPU reads the buffer (or the reduction's host-accessible result) before the GPU work has completed or become visible. Possibly the cooperative wait from Make synchronization cooperative, using GPUToolbox #654 returns before the work it's waiting on is actually done.
• Not investigated further yet. A first step would be to loop this testset on 1.13 on Aurora to find out which buffer type fails, and whether the input array or the reduction output is the stale one.