Phase D: Multi-model support for AMICAMLXNG

[neuromechanist]

Summary

Epic #74 Phase D (final phase): extend the MLX Apple-GPU backend (AMICAMLXNG) to n_models > 1. The Phase B benchmark showed multi-model had no GPU path (MLX was single-model, MPS loses); this closes that gap so the ~7x single-model MLX win extends to multi-model AMICA.

Ports the multi-model machinery from AMICATorchNG:

• comp_list indirection (params go (n_mix, n_channels) -> (n_mix, n_comps)), per-model W = inv(A[:, comp_list[:,h]]) + slogdet on the CPU stream (hoisted per iteration).
• E-step per-model loop -> logV (batch, n_models), cross-model responsibilities v = softmax(logV), u = v_h * z.
• M-step: gm, the per-model exact-EM bias c update, and the gm-weighted A-update scattered through comp_list.

Single-model (#76) stays byte-for-byte unchanged (the loop runs once, gm=1, identity comp_list -> the A-update collapses to the old A - lrate*(dA.T@A)). Component sharing remains a fast-follow.

Validation (real sample EEG; Apple GPU)

• Multi-model (n_models=2) matches PyTorch float32: one-iteration sufficient stats agree to float32 precision (dWtmp 2.7e-3, dmu_d 6.2e-4, rest <=1e-4); converged LL gap ~1e-5 (MLX -3.38967 vs torch -3.38966). New test_multimodel_matches_torch_float32.
• Single-model byte-identity + torch-float32 tests still pass (5/5 mlx tests green).

Benchmark

MLX now runs in the multi-model configs (dropped the single-model gate). MLX wins multi-model too: ~38 ms/it (32ch) / 45 ms/it (70ch) -- ~5x over torch-CPU, MPS still loses. Findings updated (.context/issue-77/benchmark_findings.md).

Scope note

The comprehensive native-Fortran + CPU-core-scaling (4/6/8/12) + CUDA cross-platform benchmark is intentionally NOT in this PR -- it is being split into its own epic (native x86 Linux + CUDA host, since Apple's Fortran binary is x86-under-Rosetta and Mac CPU timing is unreliable).

Test plan

uv run pytest pyAMICA/tests/mlx_tests/ (Apple Silicon; self-skips in CI). Full torch suite unaffected (mlx_impl is isolated + optional). ruff clean; mlx_impl omitted from the coverage gate (Apple-only).

Closes #81
Part of epic #74

[neuromechanist]

Review summary (4 Sonnet reviewers, worktree-isolated)

The code reviewer independently verified byte-identity (single-model ll_history exactly identical pre/post-PR) and multi-model correctness -- no blocking bugs; the port is faithful. All findings were quality/coverage/doc improvements, addressed in e8e5d53:

Silent-failure

• Ported the dead-model warning to the multi-model c update (a collapsed zero-responsibility model kept its prior c but was surfaced-free; torch logs it).
• Added gm/c to the per-iteration mx.eval + params_finite guard -- they are new multi-model state feeding the next E-step, so a last-iteration blow-up must be caught (nan_params), and c's cross-iteration dependency must materialize each iteration rather than growing the lazy graph unbounded (a real, if narrow, bug).

Tests

• Expanded the multi-model one-iteration stat comparison to every accumulator, including dWtmp (the gm-weighted A-update input, transposed to align (n_models,n,n) vs (n,n,n_models)) and the scattered mixture stats.
• Added a direct c-update value check (responsibility-weighted data mean; the two models' c must differ) and a single-model c==0 regression guard.

Comments/docs

• Fixed two wrong AMICATorchNG citations (c-update 1083-1092, A-update 1231-1247), the stale _lgamma_table shape comment, and dropped the leftover "MVP/v1" framing that contradicted the new multi-model scope.

Not changed (with rationale)

• Dead-model 0/0 in alpha/A-direction -> NaN (silent-failure Finding 3): pre-existing, bit-for-bit matches AMICATorchNG, and correctly caught by params_finite (nan_params stop + warning), not silent. A "freeze the dead model gracefully like c" refactor is a cross-backend follow-up, not this PR.
• 3-model / dead-model-branch unit tests (test-analyzer, low priority): the loops are generic over range(n_models); n_models=2 exercises the comp_list arithmetic. Noted as nice-to-haves.

mlx suite green (5 passed).