fix(compute): nine GPU tests panicked without an adapter — the nightly coverage run was RED six days on an environment fact (PMAT-1106)

[noahgift]

What

backends::gpu::device::backward::tests called GpuDevice::new().expect("GPU device") nine times. Under cargo llvm-cov on yoga-build2 (docker runner, no GPU passthrough) two of them panicked and the whole nightly coverage run exited 2 — run 34575134766, and the run is RED six days running (the other half of that redness is the same class in aprender-cgp, #2848, pulled into 0.67.0).

A missing adapter is an environment fact, not a kernel defect. device_or_skip() prints SKIP: no GPU adapter on this host (<err>) and the test returns; the box that has an adapter stays the gate.

Proven both ways (lambda-vector, cargo test -p aprender-compute --lib --features gpu -- backends::gpu::device::backward --nocapture)

polarity	result
VK_ICD_FILENAMES=/nonexistent (no adapter)	9 SKIP: lines, 9 passed in 0.00 s
real adapter	0 SKIP: lines, 9 passed in 1.42 s
cargo fmt --check; cargo clippy -p aprender-compute --all-targets -- -D warnings -A unused-variables (CI's chain, default features)	clean, 0 errors
cargo clippy -p aprender-compute --lib --tests --features gpu -- -D warnings	1 error, pre-existing and not in this diff: len() == 0 at backends/gpu/pool.rs:169, a gpu-feature-only test target no gate lints today (same class as the aprender-gpu dark targets noted in memory)

Falsifier of record: the next coverage-nightly.yml run (03:00 UTC) — with this and #2848 on main it must reach the report step and print a line-coverage figure instead of exiting 2 in cargo test.

Refs PMAT-1106, #2848, operator rule 2026-09-11 ("nightly we should have a low priority code coverage run with goal of eventually getting to 95% coverage").

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3116 head=df67d3e00432f047731cbba655cfd900778c32b2 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

Falsifier ran (2026-09-11 10:19Z, coverage-nightly.yml dispatched on this branch, run 34588621366, yoga-eph): the run reached the report step — TOTAL: 815132/927205 lines covered (87%) — where every run of the last six days died in cargo test on the GPU-conditional panics. So the compute half is fixed. The job still exits 2 because the measured 87.91 % is 0.09 pt under the enforced COV_FLOOR := 88 (the denominator grew from 885,829 to 927,205 lines since the July measurement); that is the next Kaizen row, not this PR's. The cgp half (#2848) is still needed for a green run on intel, where a nvidia-smi binary exists without a device.

[noahgift]

Superseded: this branch's commits (device_or_skip for the wgpu device tests) are already merged into #3112 (PMAT-1102), which also carries the aarch64 lint fixes and two dark-test fixes surfaced by the quick tier. Cancelled this PR's rerun to free X64 runner slots for #3112 (the queue was 17 deep at 15:04Z); will close this PR once #3112 lands. — Claude, autonomous fleet loop