# Parity tests for the TabICL feature preprocessing (R/tabicl-preprocess.R)
# against scikit-learn. Fixtures are generated by dev/tabicl/dump_preprocess.py:
# fit on the training rows, transform a held-out set that includes an outlier.
# The Yeo-Johnson lambdas come from an optimizer but match scipy's Brent to ~1e-7,
# so all comparisons use the float32 storage tolerance.

# Load a preprocessing fixture as plain R matrices / vectors.
tabicl_load_prep <- function(name) {
  f <- tabicl_load_fixture(name)
  lapply(f, function(t) {
    arr <- as.array(t)
    if (length(dim(t)) == 1L) as.numeric(arr) else as.matrix(arr)
  })
}

test_that("tabicl standard scaler matches sklearn", {
  skip_if_no_tabicl_fixtures("prep_standard_scaler")
  f <- tabicl_load_prep("prep_standard_scaler")

  p <- brulee:::tabicl_standard_scaler_fit(f$x_fit)
  expect_lt(max(abs(p$mean - f$mean)), 1e-5)
  expect_lt(max(abs(p$scale - f$scale)), 1e-5)

  out <- brulee:::tabicl_standard_scaler_transform(p, f$x_apply)
  expect_lt(max(abs(out - f$out)), 1e-5)
})

test_that("tabicl outlier remover matches sklearn", {
  skip_if_no_tabicl_fixtures("prep_outlier_remover")
  f <- tabicl_load_prep("prep_outlier_remover")

  p <- brulee:::tabicl_outlier_remover_fit(f$x_fit, threshold = 4.0)
  expect_lt(max(abs(p$lower - f$lower)), 1e-5)
  expect_lt(max(abs(p$upper - f$upper)), 1e-5)

  out <- brulee:::tabicl_outlier_remover_transform(p, f$x_apply)
  expect_lt(max(abs(out - f$out)), 1e-5)
})

test_that("tabicl preprocessing pipeline (none) matches sklearn", {
  skip_if_no_tabicl_fixtures("prep_pipeline_none")
  f <- tabicl_load_prep("prep_pipeline_none")

  pp <- brulee:::tabicl_preprocess_fit(f$x_fit, normalization_method = "none")
  out <- brulee:::tabicl_preprocess_transform(pp, f$x_apply)
  expect_lt(max(abs(out - f$out)), 1e-5)
})

test_that("tabicl preprocessing pipeline (Yeo-Johnson) matches sklearn", {
  skip_if_no_tabicl_fixtures("prep_pipeline_power")
  f <- tabicl_load_prep("prep_pipeline_power")

  pp <- brulee:::tabicl_preprocess_fit(
    f$x_fit,
    normalization_method = "YeoJohnson"
  )
  expect_lt(max(abs(pp$normalizer$lambdas - f$lambdas)), 1e-5)

  out <- brulee:::tabicl_preprocess_transform(pp, f$x_apply)
  expect_lt(max(abs(out - f$out)), 1e-5)
})

test_that("Yeo-Johnson transform keeps a matrix for a single row", {
  skip_on_cran()
  set.seed(396187)
  x_fit <- matrix(rnorm(40), ncol = 4)
  pp <- brulee:::tabicl_preprocess_fit(
    x_fit,
    normalization_method = "YeoJohnson"
  )

  one_row <- brulee:::tabicl_preprocess_transform(pp, x_fit[1, , drop = FALSE])
  expect_true(is.matrix(one_row))
  expect_equal(dim(one_row), c(1L, ncol(x_fit)))

  multi <- brulee:::tabicl_preprocess_transform(pp, x_fit[1:2, , drop = FALSE])
  expect_equal(one_row[1, ], multi[1, ])
})

test_that("unsupported normalization methods error", {
  skip_on_cran()
  x <- matrix(rnorm(40), ncol = 4)
  expect_snapshot(
    error = TRUE,
    brulee:::tabicl_preprocess_fit(x, normalization_method = "quantile")
  )
})