test_that("cross_design creates correct factorial grid", {
  chats <- list(gpt = "mock_chat", claude = "mock_chat")
  design <- cross_design(
    prompts = c("a", "b"),
    chats = chats,
    temperatures = c(0.0, 1.0)
  )

  expect_s3_class(design, "tbl_df")
  # 2 prompts x 2 chats x 2 temps = 8

  expect_equal(nrow(design), 8)
  expect_true(all(c("condition_id", "prompt", "chat_name", "temperature") %in% names(design)))
  expect_setequal(unique(design$prompt), c("a", "b"))
  expect_setequal(unique(design$chat_name), c("gpt", "claude"))
  expect_setequal(unique(design$temperature), c(0.0, 1.0))
})

test_that("cross_design validates inputs", {
  expect_error(cross_design(prompts = character(0), chats = list(a = 1)))
  expect_error(cross_design(prompts = "x", chats = list(1)))
})

test_that("replicate_design expands correctly", {
  chats <- list(gpt = "mock")
  design <- cross_design(prompts = c("a", "b"), chats = chats)
  rep_design <- replicate_design(design, n = 3)

  expect_equal(nrow(rep_design), 6)  # 2 conditions x 3 reps
  expect_true("replicate" %in% names(rep_design))
  expect_setequal(unique(rep_design$replicate), 1:3)
})

test_that("randomize_design is reproducible with seed", {
  chats <- list(a = "m")
  design <- cross_design(prompts = letters[1:5], chats = chats)
  r1 <- randomize_design(design, seed = 123)
  r2 <- randomize_design(design, seed = 123)

  expect_equal(r1$condition_id, r2$condition_id)
})

test_that("randomize_design changes order", {
  chats <- list(a = "m")
  design <- cross_design(prompts = letters[1:20], chats = chats)
  randomized <- randomize_design(design, seed = 99)

  # Very unlikely to be in original order

  expect_false(all(randomized$condition_id == design$condition_id))
})