Implement weighted samples and robust TRF (Issue #17)

[Hugo-W]

Summary

Implements weighted samples and robust TRF estimation as described in Issue #17.

Changes Made

utils.py

• Added robust_loss() function: d(e) = log(1 + (e/σ)²) for robust error computation
• Added robust_weights() function: computes weights for IRLS based on residuals
• Added apply_sample_weights() function: applies W to X and y for weighted least squares (maps X to WX)

models.py (TRFEstimator)

• Added weights parameter to fit() method for weighted samples
  • When weights are provided, applies them to X and y (WX transformation)
  • Uses same solver as usual after weighting
• Added robust parameter for robust TRF with IRLS
  • Uses log-based error function for outlier resistance
  • Implements Iteratively Reweighted Least Squares (IRLS)
• Added solver parameter to __init__(): 'auto', 'svd', or 'cg'
• Added _solve_trf() helper method to choose between SVD and Conjugate Gradient
• Updated _fitlists() to support weights and robust mode
• Conjugate Gradient solver used for robust TRF (iterative reweighting)

Usage Example

# Weighted samples
weights = np.exp(-0.1 * np.arange(n_samples))  # temporal decay
trf = TRFEstimator(tmin=-0.5, tmax=1.2, srate=125)
trf.fit(X, y, weights=weights)

# Robust TRF with IRLS
trf = TRFEstimator(tmin=-0.5, tmax=1.2, srate=125)
trf.fit(X, y, robust=True, max_irls_iter=10, robust_sigma=1.0)

# Access standardized coefficients
print(trf.standardized_coef_)

Mathematical Background

• Weighted samples: min ||W(y - Xβ)||² → transform to ||W^(1/2)y - W^(1/2)Xβ||²
• Robust TRF: Uses d(e) = log(1 + (e/σ)²) to reduce outlier influence
• IRLS: Iteratively reweight based on residuals until convergence

Testing

• Basic weighted least squares with synthetic data
• Robust TRF converges with IRLS
• Conjugate Gradient solver works correctly
• Backward compatibility (weights=None, robust=False)

Related Issues

Closes #17

[Hugo-W]

Closing this PR as superseded.

The branch was built on an old snapshot of ai-revamp and conflicts with work that has since landed on the base branch — in particular the quadratic-regularization refactor (_build_quadratic_regularizer, with alpha as the M-strength knob, commit 5d8f1eb) and the standardized_coef_ block. Reusing this PR would regress those fixes.

We're reopening the work as a smaller, focused fix for issue #17 on a fresh branch off current ai-revamp. Scope is reduced to the sample-weight mechanism only (apply_sample_weights + a weights parameter on fit/_fitlists); the robust-loss / IRLS / CG-solver additions are dropped for now as they add complexity without clear near-term payoff.

Thanks for the work — the weighted-least-squares formulation and apply_sample_weights helper will be reused.

[Hugo-W]

Closing as superseded — see comment above.