Struct FamilyLinearizationState 

pub struct FamilyLinearizationState<'a> {
    pub beta: &'a [f64],
    pub family_scalars: Option<Arc<dyn Any + Send + Sync>>,
    pub channel_hessian: Option<Arc<dyn FamilyChannelHessian>>,
    pub probit_frailty_scale: f64,
}

β-linearization state passed to BlockEffectiveJacobian::effective_jacobian_at.

At pre-fit initialization, pass beta = &[] / zeros and family_scalars = None. Families that need β-dependent scalars (e.g. survival marginal-slope’s q0, q1, g, c, z) store them in family_scalars as a concrete type behind Arc<dyn Any + Send + Sync> and downcast inside their impl.

Fields

beta: &'a [f64]

family_scalars: Option<Arc<dyn Any + Send + Sync>>

Optional family-shared scalars at this β linearization. Downcast via state.family_scalars.as_ref().and_then(|a| a.downcast_ref::<T>()).

channel_hessian: Option<Arc<dyn FamilyChannelHessian>>

Optional per-subject channel Hessian for multi-output families. When Some, the identifiability canonicalisation step and the Gram builder use the channel-stacked Fisher information instead of the scalar-weight approximation. Single-output families leave this None.

probit_frailty_scale: f64

Probit frailty scale factor s_f = 1/√(1+σ²).

For survival marginal-slope families the logslope η contribution is s_f · g · z, so any Jacobian callback that depends on g or z must read s_f from here rather than from a captured-at-construction value. When σ = 0 (no frailty) or for non-frailty families, set this to 1.0.

Since σ is always fixed (not jointly optimised with β) in the survival family, s_f is a static scalar for the entire inner fit; ∂s_f/∂σ never appears in the β-Jacobian. The field is nonetheless carried through state so that Jacobian callbacks are not required to capture s_f at spec-construction time — they can read it at evaluation time and thus stay correct across outer-loop σ updates.