Changes in version 2.0.0 (2026-03-19)                  

Breaking Changes

  - All solvers now return solver_result objects (class vector:
    c("solver_result", "mle_fit_numerical", "mle_fit")). Per-solver
    subclasses (mle_gradient_ascent, mle_bfgs, etc.) are removed.
  - Solver metadata moved to $solver_info nested list. Access solver
    name via result$solver_info$name, iterations via
    result$solver_info$iterations, trace data via
    result$solver_info$trace_data.
  - Composition metadata moved to $solver_info$composition. Access
    strategy via result$solver_info$composition$strategy, chain results
    via result$solver_info$composition$chain, etc.
  - Solver-specific diagnostics moved to $solver_info$diagnostics (e.g.,
    $diagnostics$final_temp for sim_anneal, $diagnostics$cycles for
    coordinate_ascent).
  - is_mle_numerical() replaced by is_solver_result().
  - compose() removed from exports — use chain() instead (superset with
    early_stop).
  - mle_problem() is now an S3 generic dispatching on its first
    argument.
  - Requires algebraic.mle >= 2.0.0.

New Features

  - mle_problem.likelihood_model(): Bridge to the likelihood.model
    package. Pass a likelihood_model object and data to auto-extract
    loglik/score/fisher.
  - fisher_scoring() now correctly reports "fisher_scoring" as solver
    name (previously reported "newton_raphson").
  - coordinate_ascent() gains learning_rate parameter for
    non-line-search mode.
  - grid_search() now errors if the grid would exceed 1 million points.
  - race() warns when parallel = TRUE but future is not installed.

Bug Fixes

  - sim_anneal() now correctly reports convergence based on temperature
    cooling (previously always reported converged = TRUE).
  - coordinate_ascent() non-line-search mode uses scaled gradient step
    instead of hardcoded 0.1 step.
  - print.mle_trace_data() no longer accesses vector elements before
    null-check.
  - print.mle_problem() correctly detects default (unconstrained)
    problems.

                        Changes in version 1.0.0                        

CRAN Release

  - First CRAN release
  - Moved algebraic.mle from Imports to Depends so that generics
    (params(), se(), confint(), loglik_val(), aic(), nparams(), vcov())
    are available immediately when the package is loaded
  - Converted \dontrun{} to \donttest{} in examples (CRAN policy)
  - Added hypothesize to Suggests
  - Added CI via GitHub Actions

                        Changes in version 0.2.0                        

New Features

  - Simulated annealing solver (sim_anneal())
  - Coordinate ascent solver (coordinate_ascent())
  - race() function for explicit parallel solver racing with future
    support
  - chain() function for sequential composition with early stopping
  - Convergence diagnostics: plot.mle_numerical() and
    optimization_path()
  - Derivative caching in mle_problem() to avoid redundant computation
  - Verbose/progress output via cli for long-running solvers
  - Trace aggregation with merge_traces() across composed solvers

Documentation

  - Four vignettes: getting-started, theory-and-intuition, case-studies,
    strategy-design
  - pkgdown site at https://queelius.github.io/compositional.mle/

Bug Fixes

  - Fixed R CMD check issues (missing imports, example errors)
  - Fixed LICENSE copyright holder

                        Changes in version 0.1.0                        

  - Initial release with core solver factories and composition operators
  - Solvers: gradient_ascent, newton_raphson, bfgs, lbfgsb, nelder_mead,
    grid_search, random_search
  - Composition: %>>% (sequential), %|% (racing), with_restarts(),
    unless_converged()
  - Problem specification with mle_problem() and constraint support
  - Tracing system for optimization iteration recording
  - Penalty/regularization transformers (L1, L2, elastic net)