ci(gpu-tests): review fixes — rc capture, ssh keepalive, log artifact, sed guard, doc nits

[MauroToscano]

Review follow-ups for #747, targeting its branch. All small, behavior-preserving fixes; shellcheck and actionlint pass.

Fixes

.github/workflows/gpu-tests.yml

• rc=$? after if wait_ready ...; fi was always 0 — bash gives a false if with no else exit status 0, so the rc=1 timeout / rc=2 image failure diagnostic in the provisioning warning never worked. Now captured in an else branch (verified empirically).
• ssh keepalive on the test session — ConnectTimeout only covers connection setup; a box that wedged or dropped off the network mid-suite would hang the step until the 240-minute job timeout, stalling the merge queue. ServerAliveInterval=60 + ServerAliveCountMax=10 fails ~10 minutes after the box goes dark.
• upload the full test log as an artifact (gpu-test-log, 14-day retention) — the workflow's own comment notes the UI truncates/rotates huge step logs; the artifact keeps the complete multi-hour output retrievable. Summary footer now points to it.

scripts/gpu_test.sh

• Guard the cudarc-pin sed anchors — if crypto/math-cuda/Cargo.toml's cudarc features are ever renamed or reformatted, the sed would silently no-op, fallback-latest would stay active, and the driver-symbol panic the pin exists to prevent (cuDevSmResourceSplit etc.) would come back with a confusing signature. Now fails loudly.
• Restore the mutated Cargo.toml on exit — CI re-checks-out before every run so it never noticed, but a manual run on a dev GPU box was left with a silently dirty tracked file that could be committed by accident.

Makefile

• Add test-prover-debug to .PHONY (pre-existing omission; every other test target is listed).
• test-cuda-integration comment said the test asserts R1/R2/R3 counters — it also asserts the R4 DEEP-composition and FRI-commit counters.
• Scope the test-prover-comprehensive-cuda parity comment: CPU CI's merge-queue comprehensive job also runs test_recursion_execute, which has no GPU leg (see 'Not fixed here' below).

Docs

• docs/roadmap.md: GPU FFT, GPU Merkle tree, and GPU FRI were still listed as Planned — all three are implemented (crypto/math-cuda/kernels/), dispatch-counted (crypto/stark/src/gpu_lde.rs), and CI-tested by this very PR. Now Done.
• prove_elfs_tests.rs: cargo test --ignored → cargo test -- --ignored in the ignore comment.

Deliberately NOT fixed here (bigger than one-liners)

1. Vast box leak window: teardown is if: always() in the same job and the cleanup label embeds run_attempt — if the GitHub runner itself dies (infra failure, 6h hard kill), the box leaks and nothing ever sweeps it. Needs a small scheduled sweeper workflow (also covers benchmark-gpu.yml, which has the same gap).
2. Merge-queue status-check timeout: repo setting, not code — if it's still the 60-min default when gpu-tests becomes a required check, worst-case provisioning (~75 min) fails every queue entry. Check when flipping the required check on.
3. Recursion on the GPU path: test_recursion_execute is never run with cuda enabled anywhere; adding it needs recursion ELF builds on the box and a test-time budget decision.