Struct LatentInnerSolver 

pub struct LatentInnerSolver<'a, A: ArrowSystemAssembler> {
    pub beta: Array1<f64>,
    pub latent: &'a mut LatentCoordValues,
    pub assembler: A,
    pub options: LatentInnerOptions,
}

Joint (t, β) inner Newton solver exploiting arrow structure.

Usage

1. Call LatentInnerSolver::new with the initial β, a mutable LatentCoordValues, an ArrowSystemAssembler, and options.
2. Call LatentInnerSolver::solve to run the inner Newton loop. Both β and the latent field are updated in place.
3. The returned LatentInnerOutcome::factor_cache is the artifact Piece 3’s IFT warm-start consumes when the outer loop next perturbs (β, ρ).

Fields

beta: Array1<f64>

latent: &'a mut LatentCoordValues

assembler: A

options: LatentInnerOptions

Implementations

impl<'a, A: ArrowSystemAssembler> LatentInnerSolver<'a, A>

pub fn new( beta: Array1<f64>, latent: &'a mut LatentCoordValues, assembler: A, options: LatentInnerOptions, ) -> Self

pub fn solve(&mut self) -> Result<LatentInnerOutcome, String>

Run the joint Newton loop.

Numerical-stability invariants:

• initial_ridge_t, initial_ridge_beta are clamped to ≥ 0.
• On a per-row or Schur PD failure, both ridges are escalated by lm_grow (or seeded at 1e-6 when starting from 0).
• max_ridge is the cold-restart trigger: if we exhaust the ramp and the Hessian is still non-PSD, the loop bails with a clear diagnostic citing the iteration index and both ridge levels reached — the outer driver should treat this as an identifiability failure (missing gauge-fixing penalty, collinear basis, etc.).