aprender-gpu test_gpu_allocation_under_pressure SIGKILLs on GB10 unified memory (host OOM) — size pressure from MemAvailable or refuse on unified-memory devices #3713

Seen as first responder on batch-1 #3707 at eb454d4e1: gpu-quick on gx10-eph, step "aprender-gpu cuda unit tests (--features cuda --lib --release)" (job 106397875044):

test memory::resident::tests::test_gpu_allocation_under_pressure has been running for over 60 seconds
error: test failed, to rerun pass `-p aprender-gpu --lib`
  process didn't exit successfully: `…/trueno_gpu-92a9e66850846a53` (signal: 9, SIGKILL: kill)

The same job passed on the same runner one push earlier (32dd7b666), and the push changed no GPU code. This is the unified-memory host-OOM shape on GB10: an allocation sized for "pressure" on a device whose memory IS host RAM makes the kernel OOM-kill the test process, and under concurrent load it can take the runner's other work with it. gpu-quick is not in gate's needs, so it blocked nothing. A test that can SIGKILL a runner is still a defect.

done_when (cop ruling 2026-09-21)

• test_gpu_allocation_under_pressure sizes its pressure from the host's MemAvailable (read at test time, with headroom) on a unified-memory device, OR it is refused there by name with an explicit reason. It never allocates past what the host can give.
• It can never SIGKILL the runner. Proven on a GB10 (gx10) under concurrent load: the test ends in PASS or in a named refusal, never signal 9.
• A case row or a mutant shows that the sizing reads MemAvailable, not a constant.

Refs #3707