test_that("ai_generalizability + ai_dstudy run end to end on a crossed design", {
  skip_if_not_installed("lme4")
  set.seed(42)
  d <- expand.grid(case = 1:10, prompt = 1:2, model = 1:2, run = 1:2)
  d$score <- 5 + rnorm(10, sd = 1)[d$case] + rnorm(nrow(d), sd = 0.4)

  g <- suppressWarnings(ai_generalizability(d))
  expect_s3_class(g, "aiEvalR_gtheory")
  expect_true("case" %in% g$variance_components$component)

  ds <- ai_dstudy(g, n_prompt = 2, n_model = 2, n_run = 2)
  expect_s3_class(ds, "aiEvalR_dstudy")
  expect_true(ds$phi >= 0 && ds$phi <= 1)
  expect_true(ds$g_coef >= 0 && ds$g_coef <= 1)
})

test_that("ai_dstudy with fewer conditions yields lower or equal Phi", {
  skip_if_not_installed("lme4")
  set.seed(7)
  d <- expand.grid(case = 1:10, prompt = 1:3, model = 1:2, run = 1:2)
  d$score <- 5 + rnorm(10, sd = 1)[d$case] + rnorm(nrow(d), sd = 0.5)
  g <- suppressWarnings(ai_generalizability(d))

  ds_full <- ai_dstudy(g, n_prompt = 3, n_model = 2, n_run = 2)
  ds_reduced <- ai_dstudy(g, n_prompt = 1, n_model = 1, n_run = 1)
  expect_true(ds_reduced$phi <= ds_full$phi)
})

test_that("ai_generalizability errors on missing required columns", {
  skip_if_not_installed("lme4")
  d <- data.frame(x = 1:5)
  expect_error(ai_generalizability(d), "Missing column")
})

## ---- print methods, conditional SEM, decision consistency ----------

test_that("print.aiEvalR_gtheory and print.aiEvalR_dstudy run without error", {
  skip_if_not_installed("lme4")
  set.seed(30)
  d <- expand.grid(case = 1:10, prompt = 1:2, model = 1:2, run = 1:2)
  d$score <- 5 + rnorm(10, sd = 1)[d$case] + rnorm(nrow(d), sd = 0.4)
  g <- suppressWarnings(ai_generalizability(d))
  expect_output(print(g), "Variance Components")
  ds <- ai_dstudy(g, n_prompt = 2, n_model = 2, n_run = 2)
  expect_output(print(ds), "D-Study")
})

test_that("ai_conditional_sem returns per-bin CSEM as a data frame", {
  set.seed(31)
  # rows = prompts, cols = occasions; larger spread at higher score levels
  responses <- t(sapply(1:40, function(i) {
    mu <- i / 10
    rnorm(4, mean = mu, sd = 0.1 + mu * 0.05)
  }))
  out <- ai_conditional_sem(responses, n_bins = 4)
  expect_s3_class(out, "data.frame")
  expect_true(all(c("score_bin_mid", "csem", "n") %in% names(out)))
  expect_true(all(out$csem >= 0))
})

test_that("ai_decision_consistency is perfect when all occasions agree", {
  # every prompt is clearly on one side of the cutpoint on every occasion
  responses <- rbind(
    c(10, 10, 10),
    c(0, 0, 0),
    c(9, 9, 9),
    c(1, 1, 1)
  )
  out <- ai_decision_consistency(responses, cutpoint = 5)
  expect_equal(out$consistency, 1)
  expect_equal(out$kappa, 1)
})

test_that("ai_decision_consistency detects inconsistent decisions", {
  # prompt 1 flips across occasions relative to cutpoint 5
  responses <- rbind(
    c(4, 6, 4),
    c(0, 0, 0),
    c(10, 10, 10)
  )
  out <- ai_decision_consistency(responses, cutpoint = 5)
  expect_lt(out$consistency, 1)
})

test_that("ai_decision_consistency returns NA kappa with a single occasion", {
  responses <- matrix(c(10, 0, 9), ncol = 1)
  out <- ai_decision_consistency(responses, cutpoint = 5)
  expect_true(is.na(out$kappa))
})

test_that("ai_generalizability errors without lme4-style data columns", {
  skip_if_not_installed("lme4")
  expect_error(ai_generalizability(data.frame(score = 1:5, case = 1:5),
                                    facets = c("prompt")),
               "Missing column")
})