Phase 3: preserve pamica-specific metadata through the MNE wrapper

[neuromechanist]

Summary

An mne.preprocessing.ICA carries no source-density family, GG shape, or component-sharing state, so those pamica-specific fitted parameters would be silently dropped by the to_mne_ica export. This phase exposes them as read accessors on AMICAICA (and, for parity, on AMICA/AMICATorchNG).

Closes #142. Part of epic #139. Read-only metadata accessors — no algorithm or fit-path change.

What's new

• get_pdftype(model_idx=...) -> (n_components,) source-density family code (0-4) per component. PDFTYPE_NAMES (exported from pamica.mne_compat and pamica.torch_impl) maps codes to names: generalized_gaussian, super_gaussian_cosh, gaussian, logistic, sub_gaussian_cosh. Codes differ per component only under the adaptive switcher (pdftype=1, Adaptive-PDF selection for AMICATorchNG #26).
• get_rho(model_idx=...) -> (n_mix, n_components) generalized-Gaussian shape parameters (meaningful for pdftype=0).
• shared_components() -> one group of (model_idx, component_idx) pairs per column merged across models by share_comps (Implement component sharing (share_comps) in AMICATorchNG #60); empty when nothing is shared (a single model, or a default multi-model fit with sharing off).

All three exist on AMICATorchNG (guarded on unfitted), delegate through AMICA (with the degenerate-fit _check_usable guard), and are surfaced on AMICAICA with keyword-only model_idx validated against the fitted backend's model count.

Test plan

• Backend pamica/tests/torch_tests/test_ng_metadata.py (base CI, real EEG): default family is generalized Gaussian; a fixed pdftype=2 fit reports Gaussian; rho shape/bounds; shared_components empty for 1-model and default 2-model, and non-empty with a valid (>=2 models) structure when share_comps merges (real fit, comp_thresh=0.9); unfit guards; PDFTYPE_NAMES covers 0-4.
• MNE wrapper (test-mne job): per-model get_pdftype/get_rho, PDFTYPE_NAMES mapping, empty shared_components by default, model_idx bounds, and the unfitted guard.
• Local gates green: ruff, ty check ., mkdocs build --strict, typos, and the backend (7) + mne (39) suites on real sample EEG.

[neuromechanist]

Review (4 Sonnet reviewers: code, tests, silent-failure, comments)

All actionable findings addressed in commit 4a... The code-reviewer found nothing above its confidence bar (verified the comp_list indexing, shared_components grouping against the merge algorithm, and PDFTYPE_NAMES values). No false positives.

Correctness / robustness

• CRITICAL (silent-failure): negative/out-of-range model_idx gave a silent wrong-model result at the AMICATorchNG/AMICA layer (get_pdftype(model_idx=-1) returned the last model's data via torch negative indexing; out-of-range gave an opaque IndexError). Added a backend AMICATorchNG._check_model_idx (rejects negatives and non-ints with a clear ValueError) and applied it to the new accessors AND, closing the pre-existing gap once, to transform/get_mixing_matrix/get_unmixing_matrix/mir/variance_order (pmi inherits it via transform). The AMICA wrapper inherits the guard by delegation.
• HIGH (silent-failure): get_rho could return a silent NaN on a degenerate multi-model fit whose aggregate LL didn't trip nan_ll (the wrapper's _check_usable only inspects stop_reason). Added an isfinite guard matching state_dict()'s defense-in-depth.

Docs / comments

• get_rho docstring said rho is inert for "families 2-4" but it's inert for all non-zero codes 1-4 (self.dorho = pdftype == 0 gates the only rho update; the comment reviewer confirmed empirically). Corrected.
• Added a caveat to shared_components that a merge synchronizes the comp_list-routed mixture params (mu/alpha/beta/rho) but not the per-source pdtype family code, so under pdftype=1 a shared pair can report different get_pdftype codes.

Tests (the reviewers' gaps)

• The AMICA-delegation layer had no base-CI coverage: added get_pdftype/get_rho/shared_components to the existing test_amica_ng_wrapper.py unfitted + degenerate-fit guard enumerations, plus a direct AMICA happy-path delegation test.
• The shared_components test hit total collapse (all 32 merged), so its exclude-filter branch was never exercised. Switched to partial-merge params (share_start=8, share_iter=10, do_newton=True, deterministic 10 groups / comp_used=54/64) and assert 0 < len(groups) < n_sources so non-shared columns are provably excluded.
• Added: model_idx bounds rejection (positive and negative) at the backend and AMICA layers; per-model get_rho(model_idx=1) differs from model 0; and the adaptive pdftype=1 path surfacing mixed {1,4} codes through get_pdftype.

Local gates green: ruff, ty check ., mkdocs build --strict, typos, and the backend metadata (11) + mne (39) + sharing/posterior (26) suites on real sample EEG.