Enum FitResult 

pub enum FitResult {
    Standard(StandardFitResult),
    GaussianLocationScale(GaussianLocationScaleFitResult),
    BinomialLocationScale(BinomialLocationScaleFitResult),
    DispersionLocationScale(DispersionLocationScaleFitResult),
    SurvivalLocationScale(SurvivalLocationScaleFitResult),
    SurvivalTransformation(SurvivalTransformationFitResult),
    BernoulliMarginalSlope(BernoulliMarginalSlopeFitResult),
    SurvivalMarginalSlope(SurvivalMarginalSlopeFitResult),
    LatentSurvival(LatentSurvivalTermFitResult),
    LatentBinary(LatentBinaryTermFitResult),
    TransformationNormal(TransformationNormalFitResult),
    SplineScan(SplineScanFit),
    ResidualCascade(ResidualCascadeFit),
}

Variants

Standard(StandardFitResult)

GaussianLocationScale(GaussianLocationScaleFitResult)

BinomialLocationScale(BinomialLocationScaleFitResult)

DispersionLocationScale(DispersionLocationScaleFitResult)

SurvivalLocationScale(SurvivalLocationScaleFitResult)

SurvivalTransformation(SurvivalTransformationFitResult)

BernoulliMarginalSlope(BernoulliMarginalSlopeFitResult)

SurvivalMarginalSlope(SurvivalMarginalSlopeFitResult)

LatentSurvival(LatentSurvivalTermFitResult)

LatentBinary(LatentBinaryTermFitResult)

TransformationNormal(TransformationNormalFitResult)

SplineScan(SplineScanFit)

Exact O(n) state-space cubic/linear/quintic smoothing-spline scan (#1030/#1034). A scan-bearing model IS a Gaussian-identity model with a different (exact) representation: rather than a dense design + coefficient vector it carries the Durbin–Koopman smoother posterior directly (knots, smoothed states, pointwise variances, σ², log λ, exact diffuse-REML EDF, and an exact per-row predict). Library callers that want the fitted posterior get it here without paying the dense O(n·k²)+O(k³) route; the CLI/FFI save paths build the persistence payload from the same SplineScanFit via assemble_spline_scan_payload.

ResidualCascade(ResidualCascadeFit)

O(n log n) multiresolution residual-cascade smooth (#1032). UNLIKE the 1-D scan, the cascade is NOT the same posterior as the Duchon/Matérn term it stands in for (a different finite basis — the multilevel Wendland frame), so it is never a silent swap: this variant is produced only when the structural detector residual_cascade_fast_path fires on an eligible scattered-low-d Gaussian fit past the dense-kernel cliff AND the in-cascade quasi-uniformity guard certifies the metric; every other shape (and a rejected metric) falls through to the dense fit_model path. The cascade-bearing model carries the ResidualCascadeFit directly — knots-free nested geometry, coefficients, the factored precision, and an exact per-row predict; the CLI/FFI save paths build the persistence payload from its to_state snapshot.