Summary

Adds an MNE-Python compatibility layer for pamica, additive throughout: the scikit-learn-style AMICA API and the byte-identical EEGLAB I/O are unchanged. MNE-Python users get a native entry point via from pamica.mne_compat import AMICAICA, shipped as the optional pamica[mne] extra so import pamica never requires MNE. Works with current MNE, no fork or extension.

Closes #139.

Phases

• Phase 1 (Phase 1: core MNE wrapper (single-model) #140): core single-model wrapper. AMICAICA.fit(raw/epochs, picks=...) and to_mne_ica() mapping pamica's mean/symmetric-ZCA sphere/unmixing into a real mne.preprocessing.ICA (round-trips AMICA.transform to float64 precision); get_sources/apply/get_components/plot_components/plot_sources delegate to it.
• Phase 2 (Phase 2: multi-model exposure for MNE users #141): multi-model exposure. Each model exports as its own mne.preprocessing.ICA (to_mne_ica(model_idx=...)); the per-sample model dominance MNE cannot represent is exposed via get_model_probability/plot_model_probability, built on a new public AMICA.model_loglik/model_probability accessor (validated bit-for-bit against the E-step's Lht).
• Phase 3 (Phase 3: preserve pyAMICA-specific metadata through the MNE wrapper #142): pamica-specific metadata (get_pdftype/get_rho/shared_components + PDFTYPE_NAMES) surfaced rather than dropped by the ICA export.
• Phase 4 (Phase 4: expose PMI/MIR through the MNE wrapper #143): the Epic: port MIR and PMI separation-quality metrics from postAmicaUtility/pre_ICA_cleaning #133 separation-quality metrics (mir/pmi) callable directly on an MNE object.

Design notes

• MNE has no multi-model concept, so multi-model AMICA is exposed per-model plus a dominance accessor, not forced into MNE's single-unmixing ICA.
• The per-model export folds each model's data-space center c into pca_mean_; single-model (c=0) stays byte-identical.
• pca_components_ is kept orthonormal (sphere written as V diag(1/sqrt(e)) V^T) so MNE's scalp maps come out in channel space.
• PCA reduction (pcakeep/pcadb) is rejected (the export assumes full-rank whitening); non-finite input and degenerate fits are refused.

Test plan

• New pamica/tests/mne_tests/ (46 tests) run on the real bundled eeglab_data.set via a dedicated test-mne CI job (uv sync --extra mne); the base job skips them via importorskip. New backend suites (test_ng_model_posterior.py, test_ng_metadata.py) run in base CI. No mocks.
• Every phase went through a 4-5 reviewer Sonnet pass; all findings addressed (see the per-phase PRs Phase 1: core MNE wrapper (single-model) #187-Phase 4: expose PMI/MIR through the MNE wrapper #190). CI green on all four.

Merge

Per the repo policy, an epic PR uses a regular merge (not squash) to preserve the per-phase history.