trust-region-least-squares

A dense trust-region-reflective nonlinear least-squares solver that reproduces SciPy's least_squares (method='trf') bit-for-bit on its unbounded path for arbitrary dimension n, every loss (linear/soft_l1/huber/cauchy/arctan), and 2-point Jacobians, with an injectable SVD/BLAS seam so a pinned host LAPACK's numerical trajectory can be reproduced exactly. For anyone who needs scipy-identical least-squares results in Rust.