[MamadouSDiallo]

calibrate(bounds=(lo, hi)) and calibrate_matrix(bounds=) no longer raise NOT_SUPPORTED: the factor g = new/old weight is held within the bounds during the solve, as R survey's calibrate(bounds=) with the linear distance.

• svy-rs: calibrate_bounded and calibrate_bounded_by_domain, a port of grake with cal.linear (Newton steps on the MASS::ginv pseudo-inverse, R's stopping rule |misfit| / (1 + |t|) < epsilon, and its pre-scaling of weights far below the controls).
• svy: new tol= (1e-7, R epsilon) and max_iter= (50, R maxit), used with bounds. A solve that stops short raises WEIGHTING_NOT_CONVERGED naming the domains or replicate weights; on_nonconvergence="warn" keeps the last iterate (R force=TRUE). Failed replicates follow on_nonconvergence, as the closed-form replicate check does; R keeps them with a warning. bounds= with trimming= raises NOT_SUPPORTED. No side set keeps the closed form, unchanged.
• Variance: no change; the existing GREG record already matches R.

Tests: R 4.5 fixtures on apiclus1, apistrat and apisrs (two-sided, one-sided, tol=0.01, scale=, stratum interaction, by=, unit weights, non-convergence, JK1 replicates). Weights match at 1e-12, Taylor SEs at 1e-8, and the failing replicates (9, 11, 12) and their miss match R. Rust unit tests pin hand-computed cases. The five tests that asserted the refusal are replaced.

Needs an svy-rs release before svy.