GPU device loss fails loudly in one test file and silently in the other

[ocots]

Summary

test/helpers/capabilities.jl implements the Handbook's capability-gated test policy
(philosophy/testing.md, §"Capability-gated tests"), including the loud-failure guard
on_gpu_runner(). But the guard is wired into one of the two GPU skip sites. On the kkt
runner, a lost device fails loudly in one file and skips silently in the other.

The policy

Handbook/philosophy/testing.md:122-133:

On the runner that is supposed to have the capability, make its absence fail loudly
instead of silently.

on_gpu_runner() = get(ENV, "RUNNER_NAME", "") in ("kkt", "occidata")

if on_gpu_runner()
    Test.@test is_cuda_on()   # fails loudly if the GPU runner lost its device
end

Reference: CTSolvers.jl#189 / #190, CTFlows.jl#375.

What we have

Both device-tier tests correctly use Test.@test_skip rather than a silent early return, so
they show as Broken in the summary rather than vanishing. That half is right, and both carry
a comment explaining the choice.

The loud guard is not symmetric:

site	@test_skip	on_gpu_runner() guard
test/suite/flows/test_gpu_routing.jl:197-204	yes	yes
test/suite/builders/test_options_forwarding.jl:66-70	yes	no

test_options_forwarding.jl does not import on_gpu_runner at all — its using .TestCapabilities line takes only is_cuda_on.

So on a kkt whose driver broke, test_gpu_routing.jl fails as designed, but
backend (CUDA) in test_options_forwarding.jl reports Broken — indistinguishable from a
laptop run. That is a smaller version of the failure class the pattern exists to prevent.

Also worth a look

• on_gpu_runner() matches the single literal "kkt", while the Handbook snippet now reads
  in ("kkt", "occidata"). For this repository "kkt" is still correct — .github/workflows/CI.yml:39
  is runs_on: '[["kkt"]]' and there is no occidata job. But CTParser 0.9.0 has already
  retired kkt for occidata on its side. If OptimalControl follows, this literal must move
  with it — and its own docstring notes it "stops firing silently rather than failing loudly"
  if the runner is renamed.
• gpu_extension_armed() is defined and documented but never asserted on the GPU runner. On
  kkt the extension being unarmed (the MadNLPGPU + CUDA + CUDSS trigger, CTSolvers#189)
  is as much a broken-runner condition as a missing device.

Proposed fix

1. Import on_gpu_runner in test_options_forwarding.jl and add the same
   if on_gpu_runner(); Test.@test is_cuda_on(); end guard.
2. Consider asserting gpu_extension_armed() under the same guard.
3. Keep the runner literal in capabilities.jl and CI.yml in sync; a comment on each
   pointing at the other would make the coupling visible.

Found while raising the compat bounds (CTBase 0.29 / CTModels 0.18 / CTFlows 0.17 /
CTParser 0.9 / CTDirect 1.1 / ExaModels 0.12); the full suite is green, 2242 pass / 0 fail,
with these two as the only Broken.