Phase C: MLX backend AMICAMLXNG (v1 MVP)

[neuromechanist]

Summary

Adds AMICAMLXNG (pyAMICA/mlx_impl/core.py), an optional Apple-Silicon GPU backend that runs the natural-gradient EM E/M-step on the Apple GPU via MLX. Epic #74 Phase C. Now that Phase A made full-data float32 converge (the precision Apple GPUs are limited to), an MLX backend becomes viable.

v1 MVP scope: single-model, generalized Gaussian (pdftype=0), natural gradient. Newton, the other PDF families, component sharing, multi-model, outlier rejection and save/load are fast-follows (rejected with NotImplementedError, not silent).

Hybrid design (forced by MLX 0.32, verified by probing)

• GPU is float32-only (Apple has no FP64) and all mlx.core.linalg is CPU-only. So the elementwise/matmul hot path runs on the GPU float32 stream (carrying the Phase A ufp/y divide-by-zero guard), while inv(A)/slogdet(W) run on the CPU stream, hoisted to once per iteration (the torch _forward recomputes slogdet 60x/iter). Measured: CPU linalg ~42 us/iter vs a ~13 ms GPU E-pass (0.3%) -- not the bottleneck; mx.eval placement is (one per iteration bounds the lazy graph).
• lgamma/digamma are absent in MLX -- the GG normalizer and the rho update are computed host-side (SciPy) on the small rho array, so the GPU E-step has no special functions.

Optional dependency (CI-safe)

mlx is under [project.optional-dependencies], not the default deps; mlx_impl is imported lazily so import pyAMICA never requires MLX. CI (ubuntu) doesn't install mlx, so the MLX tests self-skip (pytest.importorskip + skip-if-no-Apple-GPU).

Test plan (real sample EEG, Apple GPU; NO MOCKS)

pyAMICA/tests/mlx_tests/test_mlx_backend.py (4 tests, all pass locally in ~20s):

• Sufficient stats vs the NumPy float64 reference (one block, identical shared-seed params): rtol 1e-3 (measured max relerr ~1.6e-4). Tight algorithmic faithfulness against an independent oracle.
• Converged LL vs the PyTorch float32 backend (the epic's acceptance): matched settings/seed, |gap| < 1e-2 -- measured gap ~2e-6 (both reach the same NG-GG fixed point).
• Full-data stability: finite ll_history, non-degenerate stop, ascent (the Phase A guard holds in MLX).
• MVP boundaries fail loudly (NotImplementedError for n_models>1, do_newton, pdftype!=0).
• Full non-slow torch suite unaffected (124 passed).

Honest note

This is an enablement + validation step, not a proven speed win at 32 channels -- whether MLX actually beats CPU/PyTorch-MPS is what Phase B (dimension-sweep benchmark) measures.

Closes #76
Part of epic #74

[neuromechanist]

Review summary (Sonnet reviewers, worktree-isolated)

Two reviewers completed with substantive findings (silent-failure, comments/docs); the comment reviewer ran the actual MLX suite on Apple Silicon and verified the numeric claims (converged-LL gap 2.15e-6, sufficient-stat max relerr 1.59e-4, import isolation, MLX-linalg-CPU-only). The code and test reviewers were stopped after they ballooned to ~40 min running full suites inside their isolated worktrees; the code reviewer had already reached a clean verdict (no blocking findings), and the test reviewer's angle overlaps the verified suite run. All findings below addressed in cfd22b2.

Silent-failure (parity gaps vs the torch backend's guards)

• CRITICAL: fit() had no parameter finiteness check, so a final-iteration M-step blow-up could complete as stop_reason="max_iter" with silently NaN params (torch backstops this in state_dict; the MLX MVP had none). Now checks A/mu/alpha/beta/rho each iteration and stops with a degenerate "nan_params" reason + nan final_ll_.
• HIGH: missing NaN-in-rho reset-to-rho0 + warning (present in torch) -- ported, so a rho blow-up doesn't poison the lgamma table unattributably.
• MEDIUM (real bug): doscaling multiplied mu by raw scale, silently zeroing mu on a zero-norm column while A/beta were left intact -- fixed to safe_scale (matches torch's masked handling).
• Comment on the lazy-graph exception site; transform() now raises a clear NotImplementedError instead of AttributeError.

Comment/doc accuracy

• _preprocess docstring wrongly claimed MLX eigh is float32 -- the reviewer tested and MLX's CPU eigh is full float64 (only the GPU stream is unsupported), so using numpy is a code-reuse choice, not a precision workaround. Corrected.
• Fixed the A-update citation (core.py:1176-1184, was 1156-1164) and tightened the _score_gg/_log_pdf_gg/_get_block_updates ranges.
• Module docstring now states accurately which deferrals raise NotImplementedError vs are simply absent.
• Reworded the leftover mps_pathways.md "Cost: ... v2 option" paragraph that contradicted the "MVP LANDED" status above it.

MLX suite still green (4 passed); the finiteness checks did not materially slow the fit.