coconots 2.0.4

Bug fixes

Fixed a heap buffer over-read reported by CRAN’s valgrind check. The post-hoc log-likelihood for Poisson models with covariates reused the Generalized-Poisson parameter slicing, which drops eta. Poisson models carry no eta, so the slicing also discarded the first covariate coefficient and passed a coefficient vector one element shorter than the number of covariate columns to likelihoodGP1cov() / likelihoodGP2cov(), which then read one double past the end of that vector on every observation. The stored $likelihood (and hence AIC/BIC) was wrong for these models; parameter estimates and standard errors were unaffected.
likelihoodGP2cov() clamped an underflowing per-observation likelihood to 1/10^-12 = 1e+12, rewarding the optimizer for entering the degenerate region. It now clamps to 1e-12.
The order-2 starting-value shrink loops assigned to a dead variable, so alpha3 never shrank and the loop could fail to terminate.
cocoReg() with a data.frame xreg overwrote data instead of coercing xreg.
cocoSoc() now forwards its julia argument to the internal cocoReg() fits and to cocoScore(), instead of hardcoding julia = TRUE when scoring.
Julia result dictionaries are now read by key name rather than by integer position, which depended on Julia’s hash iteration order and could shift silently across Julia versions.

cocoReg() gained a live cores argument. The order-2 likelihood kernels of the RCPP backend are multithreaded via std::thread; accumulation stays serial in the original order, so results are bit-identical for any thread count. cores defaults to the number of physical cores minus one, capped at 2 while _R_CHECK_LIMIT_CORES_ is set.
Estimation kernels and the assessment tools (cocoScore(), cocoPit(), cocoResid(), cocoBoot(), rgenpois()) were rewritten around batched density evaluations, lookup tables and memoization. Numerical results are unchanged (order-1 paths exactly, order-2 paths to ~1 ulp).
JULIA_NUM_THREADS is now set to the number of physical cores minus one before the first JuliaConnectoR call, unless the user already set it. A Julia process otherwise starts single-threaded, leaving Coconots.jl’s threaded likelihood idle.

Added a softplus link function for the conditional mean of the innovation rate in cocoReg() (covariate models). Uses a numerically stable formulation, is smooth everywhere, and guarantees a positive rate.
Added S3 summary methods for forecast objects (cocoForecast, cocoForecastCollection) returning a data frame of point forecasts (mean, median, mode) and prediction intervals per forecast horizon.