testthat::test_that("max_split_depth retains deepest-splittable-node semantics", {
  set.seed(10)
  invisible(capture.output(
    fit <- boostPM::fit_boostpm(
      matrix(c(0.1, 0.3, 0.7, 0.9), ncol = 1L),
      add_noise = FALSE,
      Omega = matrix(c(0, 1), nrow = 1L),
      max_marginal_trees = 1,
      max_dependence_trees = 0,
      c0 = 0.1,
      gamma = 0,
      max_split_depth = 0,
      min_node_observations = 1,
      prior_split_prob = 1,
      n_bins = 2
    )
  ))

  testthat::expect_identical(fit$tree_diagnostics$node_count, 3L)
  testthat::expect_identical(fit$tree_diagnostics$max_depth, 1L)
})

testthat::test_that("split equality is assigned left during fit and evaluation", {
  set.seed(11)
  invisible(capture.output(
    fit <- boostPM::fit_boostpm(
      matrix(c(0.25, 0.5, 0.75), ncol = 1L),
      add_noise = FALSE,
      Omega = matrix(c(0, 1), nrow = 1L),
      max_marginal_trees = 1,
      max_dependence_trees = 0,
      c0 = 0.9,
      gamma = 0,
      max_split_depth = 0,
      min_node_observations = 1,
      prior_split_prob = 1,
      n_bins = 2
    )
  ))

  testthat::expect_equal(fit$trees[[1]]$theta[[1]], 0.65, tolerance = 1e-15)
  expected_importance <-
    (2 / 3) * log(0.65 / 0.5) +
    (1 / 3) * log(0.35 / 0.5)
  testthat::expect_lte(
    abs(as.numeric(fit$variable_importance) - expected_importance),
    1e-15
  )

  at_boundary <- stats::predict(
    fit,
    matrix(0.5, ncol = 1L),
    type = "details"
  )
  testthat::expect_equal(
    as.numeric(at_boundary$log_density),
    log(0.65 / 0.5),
    tolerance = 1e-15
  )
})

testthat::test_that("density outside Omega has log density negative infinity", {
  fit <- make_one_split_fit(theta = 0.25, location = 0.5)
  outside <- matrix(c(-0.1, 0.25, 1.1), ncol = 1L)

  result <- stats::predict(fit, outside, type = "details")

  testthat::expect_identical(
    as.numeric(result$log_density),
    c(-Inf, log(0.5), -Inf)
  )
  testthat::expect_identical(result$mean_log_density_path, -Inf)
})

testthat::test_that("a rejected early-stopping tree is not stored or applied", {
  data <- small_two_dimensional_data()
  set.seed(12)
  invisible(capture.output(
    fit <- boostPM::fit_boostpm(
      data,
      add_noise = FALSE,
      Omega = cbind(c(0, 0), c(1, 1)),
      max_marginal_trees = 5,
      max_dependence_trees = 5,
      c0 = 0.1,
      gamma = 0,
      max_split_depth = 1,
      min_node_observations = 1,
      early_stop = c(200, 2),
      prior_split_prob = 0.9,
      n_bins = 4
    )
  ))

  testthat::expect_length(fit$trees, 0L)
  testthat::expect_equal(nrow(fit$tree_diagnostics), 0L)
  testthat::expect_length(fit$heldout_diagnostics$mean_log_density_improvement, 3L)
  testthat::expect_false(any(fit$heldout_diagnostics$accepted))
  testthat::expect_equal(fit$residual_coordinates, t(data), tolerance = 1e-15)
})