test_that("fast p-values agree with diptest across all calibrations", {
  skip_if_not_installed("diptest")
  set.seed(21985)
  sizes <- unique(c(
    BIDistances:::.fastdip_sample_sizes,
    17, 63, 999, 72001
  ))

  for (n in sizes) {
    x <- switch(
      as.character(n %% 3),
      "0" = rnorm(n),
      "1" = round(rnorm(n), 1),
      "2" = c(rnorm(n %/% 2, -1), rnorm(n - n %/% 2, 1))
    )
    got <- fast_dip_test(x)
    ref <- suppressMessages(diptest::dip.test(x))

    got_statistic <- as.numeric(got$statistic)
    ref_statistic <- as.numeric(ref$statistic)
    statistic_error <- abs(got_statistic - ref_statistic)

    # Compare independent C++ and C implementations numerically, not bitwise.
    expect_true(
      statistic_error <= 1e-12,
      info = paste(
        "n =", n,
        "absolute statistic error =",
        format(statistic_error, digits = 17)
      )
    )
    expect_equal(
      got$p.value,
      ref$p.value,
      tolerance = 2e-14,
      info = paste("n =", n)
    )
    expect_equal(got$nobs, ref$nobs, tolerance = 0)
  }
})

test_that("calibration choices are explicit at the 72000 boundary", {
  set.seed(1)
  x <- rnorm(72000)
  at_limit <- fast_dip_test(x)
  expect_identical(at_limit$p.value.method, "table (last row)")

  x_large <- c(x, rnorm(1))
  expect_warning(
    above <- fast_dip_test(x_large, warn.asymptotic = TRUE),
    "limiting calibration"
  )
  expect_identical(above$p.value.method, "asymptotic")
  expect_error(
    fast_dip_test(x_large, p.value.method = "table"),
    "ends at n = 72000"
  )
})

test_that("bootstrap calibration is reproducible", {
  set.seed(7)
  x <- c(rnorm(50, -1), rnorm(50, 1))
  set.seed(2026)
  a <- fast_dip_test(x, p.value.method = "bootstrap", B = 50)
  set.seed(2026)
  b <- fast_dip_test(x, simulate.p.value = TRUE, B = 50)

  expect_equal(a$p.value, b$p.value, tolerance = 0)
  expect_identical(a$p.value.method, "bootstrap")
})