Add a macOS Apple Silicon CI job

[neuromechanist]

Closes #246 (first half: the MLX/Accelerate gap).

Why

Every job in ci.yml was ubuntu-latest. Two things were therefore untested:

• pamica/tests/mlx_tests/ had never run in CI. MLX is the recommended
  Apple GPU backend per AGENTS.md; pyproject.toml even excludes
  pamica/mlx_impl/* from coverage with the comment that CI "cannot exercise
  it and always measures 0%".
• The Accelerate BLAS numerical path. Not theoretical: the
  test_mir_history_survives_keep_best_restore assertion was green on Weight nd by pre-update model weights #241's
  PR CI and failed immediately after merge on macOS. The cause was not a stale
  base — the ~1e-6 trajectory shift moved the stopping iteration from 55 to 54,
  and the assertion only held when the stop landed on a multiple of mir_step.
  Linux never saw it. Adding CI on dev pushes would not have caught this
  either; only a macOS runner would.

What

One test-macos job on macos-26 (current arm64 image; macos-14 is arm64 too
but deprecated). Standard runners are free and unmetered for public repos, so
this costs nothing — deliberately not a -large/-xlarge variant, which are
billed even here.

Verified locally on macOS 26.6.1 arm64 before pushing:

• mlx.core.__version__ -> 0.32.0, so the import assertion is real
• LAPACK_LIBS="-framework Accelerate" bash native/build.sh builds cleanly
• pamica/tests/mlx_tests/ collects 11 tests and runs

The MLX import is asserted explicitly rather than left to mlx_tests'
self-skip: a silent skip would leave this job green while testing nothing,
which is the exact failure mode it exists to prevent.

--no-cov because the Linux job owns the coverage gate and the Codecov upload;
a second partial run would report a different total for the same commit.

Deliberately not in scope

• Enabling the slow Fortran-parity tests on macOS. They drive the bundled
  x86_64 amica15mac, which needs Rosetta on an arm64 runner. Pointing them at
  the arm64 binary this job already builds is the natural follow-up; left in
  CI runs Linux only: MLX backend and the Apple Silicon numerical path are never tested #246.
• push: branches: [dev]. Worth doing, but a separate concern, and note it
  would not have caught the failure described above.
• A self-hosted runner. sccn/pAMICA is public and GitHub advises
  self-hosted runners only for private repos (a fork PR can run arbitrary code
  on an unsandboxed runner). Hosted macos-26 covers the same ground with no
  exposure. Self-hosted remains the only option for CUDA.

[neuromechanist]

Findings addressed in da41ebd. One correction on the Critical, with evidence.

Critical 1 — "the MLX tests will very likely still self-skip": empirically not the case. I checked the completed run's log rather than reasoning about it. On the hosted macos-26 runner, all 11 MLX tests PASSED:

PASSED pamica/tests/mlx_tests/test_mlx_backend.py::test_sufficient_stats_match_numpy_reference
PASSED pamica/tests/mlx_tests/test_mlx_backend.py::test_converged_ll_matches_torch_float32
PASSED pamica/tests/mlx_tests/test_mlx_backend.py::test_multimodel_matches_torch_float32
... (11 total)
============ 402 passed, 10 skipped, 1 xfailed in 582.91s (0:09:42) ============

The 10 skips are test_fortran_adapter.py (3) and test_ng_pdf_families.py (5) plus two others — no MLX test among them. So this runner does expose Metal to MLX and mx.default_device().type == gpu holds. The cited upstream reports are real but do not describe this image.

That said, the finding is right as insurance and I have implemented it. You correctly spotted the second gate at test_mlx_backend.py:28-33 that my import assertion did not cover — if a future runner image loses GPU passthrough, the job would go green with 11 silent skips. The step now asserts mx.default_device().type == mx.DeviceType.gpu.

Important 2 — the -m "not slow" rationale was wrong. Confirmed both halves. test_sample_data.py's slow tests never invoke a binary, and the two in test_ng_backend.py (:1096, :1179) pass binary_path=None, which resolves through PAMICA_NATIVE_BINARY first (native/resolver.py:112) — already set by this job to the arm64 binary it builds. Rosetta was never involved. Comment corrected to say they are excluded for wall-clock, and that enabling them would give CI its first Fortran parity coverage.

Important 3 — PYTORCH_ENABLE_MPS_FALLBACK. Added. This is indeed the only job whose hardware can auto-select MPS.

Important 4 — EXTRA_LDFLAGS. Added -static-libquadmath so the recipe genuinely matches release-binaries.yml. Verified the build locally with both flags: arm64 Mach-O, clean.

Suggestion — stale coverage comment. Updated to state why the omission is still required (this job runs --no-cov, so the total still comes only from the ubuntu job) rather than leaving a claim a future reader would find false.

Suggestion — release-binaries.yml still on macos-14. Good catch on the November 2 2026 EOL; filing separately rather than widening this PR.