val <- prepare_largeraw_sample()
small_val <- val[seq_len(600), ]

model_seen <- list()
feature_input_model_adapter <- LKTCustomModel(
  name = "feature-input-smoke",
  fit = function(request, verbose = FALSE) {
    model_seen$request_names <<- names(request)
    model_seen$feature_input_names <<- names(request$feature_input)
    model_seen$feature_input_class <<- class(request$feature_input)
    model_seen$column_names <<- request$feature_input$column_names
    model_seen$response_length <<- length(request$feature_input$response)
    model_seen$feature_count <<- length(request$feature_input$feature_spec)

    LibLinearModel()$fit(request, verbose = verbose)
  }
)

feature_input_model <- LKT(
  data = small_val,
  interc = FALSE,
  components = c("Anon.Student.Id", "KC..Default.", "KC..Default."),
  features = c("intercept", "intercept", "lineafm"),
  model = feature_input_model_adapter,
  verbose = FALSE
)

check_true("model request includes feature_input", "feature_input" %in% model_seen$request_names)
check_true("feature_input has LKTFeatureInput class", "LKTFeatureInput" %in% model_seen$feature_input_class)
check_true("feature_input exposes formula", "formula" %in% model_seen$feature_input_names)
check_true("feature_input exposes design matrix", "design_matrix" %in% model_seen$feature_input_names)
check_true("feature_input exposes column names", length(model_seen$column_names) > 0)
check_true("feature_input response length matches data", model_seen$response_length == nrow(small_val))
check_true("feature_input feature spec length", model_seen$feature_count == 3)
check_true("custom model preserves common outputs", all(c(
  "model",
  "coefs",
  "model_specification",
  "r2",
  "prediction",
  "loglike",
  "model_name"
) %in% names(feature_input_model)))
check_true("custom model reports its name", feature_input_model$model_name == "feature-input-smoke")
check_true("custom model predictions align", length(feature_input_model$prediction) == nrow(small_val))

prediction_only_model_adapter <- LKTCustomModel(
  name = "prediction-only",
  fit = function(request, verbose = FALSE) {
    p <- mean(request$response, na.rm = TRUE)
    p <- pmin(pmax(p, .00001), .99999)
    LKTModelFit(
      model = list(kind = "constant-probability", probability = p),
      pred = rep(p, length(request$response))
    )
  }
)

prediction_only_model <- LKT(
  data = small_val,
  interc = FALSE,
  components = c("KC..Default."),
  features = c("intercept"),
  model = prediction_only_model_adapter,
  verbose = FALSE
)

check_true("prediction-only model returns model", !is.null(prediction_only_model$model))
check_true("prediction-only model has finite loglike", is.finite(prediction_only_model$loglike))
check_true("prediction-only model has finite r2", is.finite(prediction_only_model$r2))
check_true("prediction-only model reports its name", prediction_only_model$model_name == "prediction-only")
check_true("prediction-only model can omit coefficients", nrow(prediction_only_model$model_specification[[1]]) == 0)
check_true("prediction-only model uses helper constructor", inherits(prediction_only_model_adapter, "LKTModel"))

wrong_length_model <- LKTCustomModel(
  name = "wrong-length",
  fit = function(request, verbose = FALSE) {
    list(
      model = list(kind = "bad-model"),
      pred = rep(.5, max(0, length(request$response) - 1))
    )
  }
)

wrong_length_error <- tryCatch(
  {
    LKT(
      data = small_val,
      interc = FALSE,
      components = c("KC..Default."),
      features = c("intercept"),
      model = wrong_length_model,
      verbose = FALSE
    )
    FALSE
  },
  error = function(err) grepl("returned predictions with length", conditionMessage(err), fixed = TRUE)
)
check_true("wrong-length model fails clearly", wrong_length_error)

bad_probability_model <- LKTCustomModel(
  name = "bad-probability",
  fit = function(request, verbose = FALSE) {
    list(
      model = list(kind = "bad-model"),
      pred = rep(1, length(request$response))
    )
  }
)

bad_probability_error <- tryCatch(
  {
    LKT(
      data = small_val,
      interc = FALSE,
      components = c("KC..Default."),
      features = c("intercept"),
      model = bad_probability_model,
      verbose = FALSE
    )
    FALSE
  },
  error = function(err) grepl("finite probabilities", conditionMessage(err), fixed = TRUE)
)
check_true("non-probability model predictions fail clearly", bad_probability_error)