NEWS

tsissm 1.0.4

The status slot of a tsissm.estimate object now holds a unified solver convergence code: 0 = converged, 1 = the solver stopped at an iteration/evaluation limit and returned a solution which is not certified as converged, -1 = failure. It previously held the raw nloptr status code. The raw solver output is now available in solver_diagnostics, and the solver used is recorded in the new solver field.
tsbacktest now uses future.seed = TRUE, so parallel simulation uses proper L'Ecuyer RNG streams and results are reproducible across future plans. Results for a given seed may differ from earlier versions.
Switched Eigen::all to Eigen::indexing::all for compatibility with Eigen 5.0 / the upcoming RcppEigen (contributed by Dirk Eddelbuettel PR # 2).

Work around RTMB AD eigen failures on M1 macOS. CRAN M1MAC/R-devel checks fail while taping RTMB's AD overload for non-symmetric eigen decompositions, with an invalid advector error inside the complex eigenvector normalization path. This appears to be platform and toolchain sensitive: Apple silicon, especially M1 (but not M3), does not provide the same extended precision behavior as some x86 paths, and the reduced floating-point headroom can expose fragile complex eigen/AD calculations. Added a lazy runtime capability check for RTMB AD eigen support. Normal platforms continue to use RTMB's AD eigen constraints. If the capability check fails, the package falls back only for eigenvalue constraints: constraint values are computed with numeric base::eigen() and constraint Jacobians are computed by central finite differences. The likelihood, scores, and other TMB/RTMB AD paths remain unchanged.
Exposed tsissm_capabilities() to report the detected RTMB AD eigen capability, source of the decision, RTMB/TMB versions, and platform details.
Documented the tsissm.rtmb_ad_eigen option under estimate() so users can force RTMB AD eigen constraints or force the finite-difference fallback for diagnostics.
Also avoided a fragile vignette conversion through as_flextable() by printing the model summary directly.

The SLSQP solver (from nloptr) sometimes will return NaN for the parameters during the optimization phase. Added a stop clause else it may crash the process (due to RTMB eigen crashing).
Added a solver control function (issm_control)

Initial CRAN submission based on complete re-write of model on github which was originally created on 2020-09-16.