Riemannian tangent-space EEG path

[aaygan29]

Builds the Riemannian EEG arm as the proper follow-up flagged in #3, in three small commits. This is the one I said needed a dedicated path rather than a base-learner swap — here it is, done correctly.

Why it needed its own path

EEG carries much of its discriminative signal in inter-channel covariance, which lives on the manifold of SPD matrices, not a flat vector space. The framework's StandardScaler z-scores each covariance entry independently (destroying the positive-definite structure) and SMOTE interpolates in straight lines (off-manifold). So you can't just slot a "riemann" base learner into the flat pipeline — the covariance has to be projected into a flat tangent space first.

What's shipped

BaggingClassifier
  └── Pipeline(RiemannianTangentSpace → StandardScaler → SMOTE → logistic)

• models/riemann.py — RiemannianTangentSpace transformer (scipy-only, log-Euclidean metric). Isometric vectorisation (√2 off-diagonal so Euclidean norm = Frobenius), identity-shrinkage regularisation so rank-deficient short-epoch covariances still admit a matrix log, and a training-set log-Euclidean mean reference computed in fit only (leakage-safe).
• make_balanced_pipeline gains pre_steps, and build_modality_model's riemann path prepends the tangent transform, classifies with logistic in tangent space, and forces max_features=1.0 (a column subset of a flattened covariance isn't a covariance).
• EEGLoader covariance mode — emits per-trial inter-channel covariance, sharing flatten_spd with the transformer as one source of truth (a round-trip test guards against drift). Mutually exclusive with band power / ERP.

block = EEGLoader(include_bandpower=False, include_erp=False,
                  include_covariance=True).from_arrays(epochs, sfreq, times, subj, stim)
# config: model.base_learner.eeg = riemann

Two honest caveats (in the code and docs)

• It does not beat plain logistic on clean synthetic data, because raw covariance entries are already linearly separable there. Its advantage is on ill-conditioned real EEG, the regime the method exists for. The tests assert the path recovers covariance structure above chance — not a synthetic "beats logistic" win, which would be cherry-picking.
• This is the log-Euclidean metric (closed-form, no dependency). pyriemann's affine-invariant metric is a further upgrade, noted as the next step if a real EEG dataset warrants it.

Tests

18 new (test_riemann.py 12, test_eeg_covariance.py 6): isometric/invertible vectorisation, rank-deficient survival, leakage-safe reference mean, non-covariance rejection, pipeline wiring, loader↔transformer round-trip, mutual-exclusivity, and end-to-end above-chance positive controls. Full suite 147 passing, no new dependency.