fix(R-3): the cuBLAS-backward training banner is an event (launch counter), and --gpu-backend cuda on a cpu-only build is FeatureDisabled exit 9 — never a silent CPU run (PMAT-991, #2906)

[noahgift]

PP-066 DAG row R-3 (spec §5 T-6, D-7; epic #2873; ticket PMAT-991; receipt docs/audits/impl-PMAT-991-receipt.md, status: partial until merged). feedback_apr_finetune_lora_cpu_only: the "cuBLAS backward" training banner was printed from the request, before any step, on any build; --gpu-backend cuda on a cpu-only build fell through to the CPU silently. Now the banner is an event and the impossible request is a refusal.

What lands

• entrenar (crates/aprender-train): backward_kernel_launches() / reset_backward_kernel_launches() — an AtomicU64 bumped at every cuBLAS backward launch (cuda_forward/matmul.rs a, a_accumulate, b; matmul_f16.rs a, b); training_backend_banner(requested) -> Option<String> is Some iff requested == "cuda" and the counter > 0. RED-first test tests/banner_truth.rs (both feature arms; the cuda arm drives one cuBLAS backward on this box's RTX 4090 and watches the counter move).
• apr finetune: gpu_backend_decision(requested, build_has_cuda, build_has_wgpu) — cuda/wgpu on a build without the feature is CliError::FeatureDisabled (exit code 9, read from error.rs:106); auto/cpu never refuse; pre_training_notice never says "cuBLAS backward"; post_training_banner prints entrenar's line after training. Case table commands/finetune_gpu_backend_truth_tests.rs (5 tests, both polarities).
• contracts/apr-train-banner-truth-v1.yaml (kind: pattern, TBT-OB-001..003; pv validate valid, pv lint 0/0). README contract count 1811 → 1812.
• test_run_training_creates_adapter hard-coded "cuda" and asserted the old silent fallback; it now names the CPU path it always ran.

Acceptance, re-run by the orchestrator on eaea476

A_i	rc
cargo test -p aprender-train --test banner_truth	0 (2 passed)
cargo test -p aprender-train --test banner_truth --features cuda (lambda, sm_89)	0 (2 passed)
cargo test -p apr-cli --lib gpu_backend_truth	0 (5 passed)
cargo test -p apr-cli --lib	0 (7213 passed, 12 ignored)
cargo test -p aprender-train --lib	101: 7621 passed, 3 failed in prune::snapshot_tests — identical on a pristine origin/main worktree at 65680cd while CI's workspace-test there is green: local insta drift, src/prune untouched here
cargo fmt --all -- --check · cargo clippy -p {aprender-train,apr-cli} --lib -- -D warnings	0 · 0 · 0
check_contract_test_binding.sh · check_contract_enforcement.sh · check_readme_claims.sh · check_no_claim_literals.sh · check_roadmap_diff_additive.sh	0 ×5

Mutations — RED, then restored GREEN

mutation	result
training_backend_banner returns Some for "cuda" regardless of the counter	banner_is_none_for_cuda_with_zero_launches FAILED, cpu_only::cpu_backward_never_increments_the_device_counter FAILED (0 passed, 2 failed) → restored 2 passed
gpu_backend_decision returns Ok(Cpu) for "cuda" on a build without cuda	cuda_request_on_cpu_only_build_is_a_refusal_not_a_fallback FAILED, gpu_backend_decision_case_table FAILED (3 passed, 2 failed) → restored 5 passed

Findings recorded in the receipt

• The crate dir is aprender-train, the lib is entrenar (S0-10 class).
• crates/apr-cli/src/lib.rs cannot be committed on this box: the pre-commit complexity hook follows its include!() graph and refuses on pre-existing debt (dispatch.rs:103 cognitive 41, dispatch.rs:508 cognitive 30, help_producer_truth.rs:51 cognitive 73). The integration-test seam was dropped; the case table lives in commands::finetune's own test module. Not discharged with #[allow] or --no-verify.
• Not instrumented: custom-PTX backward kernels under autograd/cuda_backward/ — a PTX-only backward would leave the banner None (open question for the 3-lane review; the spec names the cuBLAS sites).
• gx10 leg not run here (lambda only); C4 dogfood covers it.

Review round (3-lane quorum, 2026-09-06 — 3/3 mergeable with changes; applied at a137ba6, re-verified)

• Unanimous: the process-wide counter with a test-only reset leaked across fine-tunes → the banner is now scoped to the run (training_backend_banner(requested, launches_at_start), snapshot before trainer.train(), no reset API); the tests snapshot and compare (race-free under parallel tests).
• Lane 1 alone said gemm_f16_to_f32_backward_a launched uncounted; lanes 2/3 said all sites were covered. Lane 1 was right, and three more uncounted device backward launches were found in matmul.rs (gemm_nf4_backward_a_cublas, the NF4 PTX and tensor-core backward_a — the QLoRA path). Nine launch sites are counted now.
• -m lora --gpu-backend cuda now REFUSES (ValidationFailed, exit 5) instead of running the CPU path under a GPU flag; auto + plain LoRA is the CPU path and the notice says so (gpu_backend_decision gained method_has_cuda_path).
• Dead gpu_backend_notice/GpuBackendPlan (still holding the request-derived cuBLAS string) and their five tests removed.
• Re-run on a137ba6: banner_truth 2/2 (default and --features cuda), gpu_backend_truth 8/8, cargo test -p apr-cli --lib 7211 passed, cargo test -p aprender-train --lib 7621 passed + the same three pre-existing prune snapshot failures, clippy/fmt clean, pv validate valid. Mutations: 1 (banner unconditional) 0/2 passed; 2 (cuda on a cpu-only build → Cpu) 6/8; 3 (plain LoRA + cuda → Cuda, no refusal) 7/8; all restored GREEN.

Quorum verdicts and the adjudication are in the receipt. Receipt for this PR itself: advisory, not produced (driver A1).

no-close: #2873 is the epic this row implements one part of; an epic is not closed by one of its rows.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3001 head=3eaee3b0a4a6487fe88364aa3ebe57ba5854f27e 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]

Rebased onto origin/main @ 3792afa3d (after G-10/G-11 #3011/#3020 landed). Dropped commit d5fde6b5b (chore(roadmap): PMAT-991 kind:code label) — it wrote docs/roadmaps/roadmap.yaml, which G-11a now forbids in a row PR. The README.md count-line bump in ccda35736 needed no action: the rebased base already carried 1812 (bumped since by the G-11 ratchet process), so the contract commit now adds only contracts/apr-train-banner-truth-v1.yaml with no README diff — row PRs do not write README.md per driver rule. The other 8 commits cherry-picked cleanly onto the new base with no conflicts.

Re-verified at HEAD (b5b2b044f):

• cargo check -p apr-cli -p aprender-train --lib → clean
• cargo test -p aprender-train --test banner_truth → 2/2 pass
• cargo test -p apr-cli --lib finetune → 79/79 pass
• cargo test -p apr-cli --lib gpu_backend → 10/10 pass (gpu_backend_truth_tests::*, test_finetune_rejects_unknown_gpu_backend_2583, test_finetune_accepts_its_own_gpu_backends_2583)
• pv validate contracts/apr-train-banner-truth-v1.yaml → 0 errors, 0 warnings
• bash scripts/check_row_pr_write_set.sh --base origin/main --head HEAD --branch agent/R-3 → PASS, 11 changed paths, no shared file touched

[noahgift]

Auto-merge armed. Currently BLOCKED in the queue on two unrelated fleet issues: #3028/PR #3030 (roadmap.yaml duplicate ids redding ci / security on every PR) and a fleet-wide gap in scripts/pmat_bin.sh's pinned analyser on some runner hosts (paiml/infra#468). Not this PR's diff — will merge once both clear.