chore(build): reconcile auto-detected CUDA arch with the shipped shape

[inureyes]

Summary

With MLX_CUDA_ARCHITECTURES unset, sm_arch_with_suffix appended CUDA's architecture-specific a suffix to every SM from 90 up, so a default cargo build --features cuda on this project's GB10 resolved the whole MLX target to 121a. That is the shape PR #1938 decided against, and it is self-defeating: the per-source injection in src/lib/mlx-cpp/CMakeLists.txt matches plain entries only and skips an a-carrying one as already architecture-specific, so on an auto-detected build it never fired at all. The defect is visible in this checkout's own target/, where two debug build directories from 2026-09-20 configured 121a and their libmlx.a carries every image architecture-specific, 98 of 98. A fixed auto-detect release build carries 1 of 99.

Auto-detection now stops suffixing at a shared FIRST_PLAIN_SM boundary, which is what makes the shipped mechanism work: Hopper keeps 90a, Blackwell is plain. Nothing here rests on throughput; that A/B was run and refuted, sitting inside the measurement method's own floor (docs/benchmark_results/data/blackwell-arch-specific-gb10-2026-09-21/).

The Hopper decision

Hopper keeps 90a because release.yml ships 90a and the job of auto-detection is to agree with it. Dropping it would create on Hopper exactly the local-versus-shipped divergence this closes on Blackwell.

The stated rationale for it is false at the current pin, so it is replaced rather than repeated. MLX_CUDA_SM90A_ENABLED has no definition in the pinned tree (81ba1c6a): upstream 44540d12, "[CUDA] JIT-compile qmm_sm80/qmm_sm90/gather_gemm", deleted the line defining it and moved those kernels to runtime NVRTC compilation, and jit_module.cpp derives the NVRTC --gpu-architecture from the running device, appending a itself from compute capability 9 up. Cross-compiled the pinned tree at 90 and at 90a here, with the real build's flags from its own compile_commands.json, and compared with cuobjdump. No Hopper device needed.

translation unit	kernels emitted	SASS lines changed	of which .headerflags	PTX changed
qmm_sm90.cu	0	0	0	0
qmm.cu	0	0	0	0
qmm_sm80.cu	0	0	0	0
qmv.cu	2898	5796	5796	0
fp_qmv.cu	54	108	108	0

So the suffix changes no instruction at this pin: the three JIT-compiled units emit no device function either way, and the two that emit code differ only in EF_CUDA_ACCELERATORS on the .headerflags line, the ELF flag marking a cubin architecture-specific. That is a reason to stop claiming 90a gates a kernel, not a reason to drop it: CUTLASS_ARCH_MMA_SM90A_ENABLED still keys on __CUDA_ARCH_FEAT_SM90_ALL, the way CUDA_ARCH_CONDITIONAL(1210) gates the Blackwell equivalent, so a later pin can make it matter. The no-detection fallback literal stays 90a on the same shipped-list argument: a fallback spelled 90 would hand a host without nvidia-smi a shape no published archive uses.

cuda_arch.rs is a deliberate extension beyond the issue's scope list: suggested_architecture applies the same rule, and the startup mismatch message is the only place the project hands an operator an MLX_CUDA_ARCHITECTURES value to paste. On a Blackwell card it named 121a, the one spelling this repository rejects everywhere else. The guard reads both functions, since rule 3 cannot see a list no workflow writes down. Two doc comments in that file were showing the shape the guard now rejects: compiled_cuda_architectures illustrated the string with "80;86;89;90a;100a-real;100;120a-real;120", replaced by the real x86_64 release list, and the test module's release-list header still justified Hopper's 90a by the deleted gate.

A sweep for other prose carrying the same dead rationale found eight live sites in total, all corrected here and none of them behavior: both release.yml jobs cited the deleted macro, three CUDA jobs in ci.yml justified their pin by auto-detection "yielding 121a" (it now yields 121, so each keeps the reason that survives rather than losing its rationale), and scripts/bench_block_width.sh carried the one claim that inverts rather than merely going stale, telling readers a row from an auto-detected build is not comparable with records pinned at plain 121 when after this change it is. No architecture list is touched. The draft-block-width-gb10-2026-09-20/harness/ directory counts as live rather than archival, and its four files are corrected too: the post-#1939 sweep and the #1943 A/B both re-used that harness, and draft-block-width-post-1939-gb10-2026-09-21/ ships no harness/ of its own because of it. The record prose under docs/benchmark_results/ and the TECHNICAL_REPORTS entry are left as written, since they document what was true when they were measured. The surface kept being wider than assumed at every pass, which is the pattern worth noting more than any single site.

Test plan

• check_cuda_arch_lists.py passes here and fails on unmodified main, three findings per site naming this issue; check_cuda_arch_lists_test.sh 13 cases pass.
• cargo fmt --check and cargo clippy -p mlxcel --lib --tests -- -D warnings at default features both exit 0. The warnings in that log are C++ build-script output from files this branch does not touch.
• cargo test -p mlxcel-core --release --features cuda --lib -- --test-threads=1 cuda_arch in the auto-detect arm at this head: 27 passed.
• Two arms, each in its own CARGO_TARGET_DIR because build.rs declares cargo:rerun-if-env-changed=MLX_CUDA_ARCHITECTURES and reuses one OUT_DIR. Both caches read 121; qmv.cu.o and fp_qmv.cu.o list sm_121 only; their --dump-sass hashes are equal across arms (84f0e788…, cc2c4d46…); each archive holds 1 architecture-specific image of 99, the auto arm's being fp_quantize.cu.o with 216 F2FP.SATFINITE.E2M1 instructions, so the injection fired.
• Raw .o bytes do differ between arms, which the change does not explain: compiling one unmodified unit twice with identical flags in one directory also yields two different objects with the same SASS hash. nvcc is not byte-reproducible here, which is why the criterion is stated on SASS.
• cargo check --features cuda,xla-iree --all-targets left to CI: IREE_DIST is unset on this host.

Closes #1943