ci: un-dark aprender-gpu and aprender-cuda-edge in workspace-test (651 tests, 0.193s)

[noahgift]

T0 of docs/specifications/nvidia-cuda-rust-library-integration.md, pulled forward into 0.66 because nothing else in that spec is measurable without it. Tracked by #3062 (0.67).

What this is

Two crates were excluded from the only required test check for a CUDA/cuBLAS link their default features do not have. Both are default = [] with no build.rs; the cuda feature is optional and off.

Crate	Tests	Run time
aprender-gpu	444	0.10s
aprender-cuda-edge	207	0.03s
	651	0.193s

Nothing else covered them. .github/workflows/cuda-nightly.yml:242 is the only other CI reference to either crate and it runs a single perf053 filter.

Correction (2026-09-09). An earlier version of this description said this PR gives the 109k-LOC PTX generator required-check coverage. It does not. In aprender-gpu, driver, kernels, memory and ptx are all #[cfg(feature = "cuda")] (lib.rs:138–151). The 444 default-feature tests this PR un-darks are monitor (308), testing (72), backend (39), error (20), graph (4), plus 13 new launch_budget tests from #3064. The PTX generator's ~2,160 tests still run only under --features cuda, i.e. the gx10 nightly's single perf053 filter. This PR is a first step, not the fix for that; the gap is tracked separately.

Measured with the exact CI harness

Not cargo test — the command CI actually runs:

cargo nextest run --profile ci -p aprender-gpu -p aprender-cuda-edge --lib
  Starting 651 tests across 2 binaries
   Summary [0.193s] 651 tests run: 651 passed, 0 skipped

Re-run in a CI-like environment (CUDA_VISIBLE_DEVICES="", empty HOME, scrubbed XDG_CACHE_HOME) — still green, so the pass does not depend on this box having a GPU or a warm ~/.cache/trueno.

Cost: ~31s extra cold build (263 dep crates, measured in an empty target dir) against the job's timeout-minutes: 150.

aprender-compute stays excluded — it SIGSEGVs at exit and already runs as its own step below.

Coupled state checked

• scripts/ci_test_tier.sh does not reference these crates
• no other workflow excludes them
• no guard script asserts the exclusion list
• YAML re-parsed after the edit (10 jobs)

Risk, stated plainly

workspace-test is a required check with fail-fast = true. If either crate has a failure that reproduces only in the clean-room container, this blocks every open PR. The scrubbed-environment run above is evidence, not proof.

That is why this lands alone rather than batched — the revert is this single commit. Please watch the first run rather than assuming it.

Touches .github/workflows/*, so per repo policy this needs a web-UI merge click.

🤖 Generated with Claude Code

https://claude.ai/code/session_01J9cSQynVPYeUkQ2i7ccvrs

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3063 head=a829836612f18c54e71daaa25b2990384256223a 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]

Pushed c6ce084ec (PMAT-1096, the 0.66.0 cut): the merge-queue run 34411725611 failed workspace-test with nextest exit 100 — driver::cublas_tests::cta64_vs_cta32_vs_cublas_fp16 and cta_wmma_dbuf_bench_fp16 panicked CudaNotAvailable("CUDA driver not found") on the clean-room runner.

Why the PR-level measurement was green and the queue was red. -p aprender-gpu -p aprender-cuda-edge --lib is a per-package resolve: cuda stays off (cargo nextest list … -p … on main lists 0 driver:: tests, 13 launch_budget::). The line this PR changed is --workspace, where cargo unifies features across every member, and aprender-explain depends on aprender-gpu with features = ["cuda"] non-optionally — so cargo nextest list --workspace --lib -E 'package(aprender-gpu)' lists 217 driver:: tests. The PR's own body already warned about exactly this class.

The fix that keeps the claim. The workspace line keeps its three excludes; the two GPU crates run as their own full-tier step (cargo nextest run --profile ci -p aprender-gpu -p aprender-cuda-edge --lib, same container and mounts as the compute step). Auto-merge is re-armed. The aprender-explain → aprender-gpu/cuda non-optional edge is a finding on its own (every --workspace build compiles the CUDA driver FFI); noted on #3067.