Added unit tests for arg checks, prep, Fisher, core ML, and pipeline

[eboyer221]

Summary

Adds a testthat test suite covering helpers and the full runMLPipeline() flow.

Coverage

• tests/testthat/test-arg-checks.R : every internal .checkArg* validator in R/arg_check_ml.R
• tests/testthat/test-prep-ml.R : .getTargetVarName, getNumFeat, shuffleLabels, calculateMinSamples, loadMLInputTibble.
• tests/testthat/test-fisher.R : encodePhenotype, runFisherTests, applyBenjaminiHochberg, computeFeatureFreq.
• tests/testthat/test-core-ml.R : splitMLInputTibble, buildRecipe, buildLRModel, buildWflow, buildTuningGrid, predictML / getConfusionMatrix / calculateEvalMets/ extractTopFeats.
• tests/testthat/test-pipeline.R : end-to-end runMLPipeline() with cross-validation instead of a train/validation/test split: output structure, signal recovery, return_fit / return_pred / return_tune_res, shuffle_labels baseline, multi-class rejection for non-LR models, removeTopFeats.

Helpers in helper-fixtures.R and helper-ml-fixture.R build small tibbles for fast pipeline tests.

Known issue -

This note is just a heads-up that the tests uncovered a pre-existing bug in runMLPipeline() that's worth fixing in a follow-up PR:
When you call runMLPipeline() with a train/validation/test split (e.g. split = c(0.6, 0.2)), it crashes. The pipeline turns off cross-validation in this mode, but the tuning step still tries to run cross-validation and errors out. Until that's fixed, the pipeline tests only cover the cross-validation mode (split = c(1, 0)). The tests are sufficient as-is so no action should be needed from the reviewer to merge them - this is just a note for a future fix.

How to run the tests

Locally, from the package root (amRml/):

# Option A — devtools (recommended during development)
devtools::test()

# Option B — R CMD check (full package check, including tests)
devtools::check()

Or from the shell:
cd amRml
Rscript -e 'devtools::test()'
Expected: [ FAIL 0 | WARN 0 | SKIP 0 | PASS 200 ].

[amcim]

@eboyer221 I looked into the preexisting train/validation/test split bug and added a test showing that runMLPipeline() now returns the expected structure with split = c(0.6, 0.2).

The change is line 264 of core_ml.R where initial_validation_split now uses:

rsample::validation_set(data_split)

instead of calling vfold_cv().

In run_ml_pipeline.R, there is this conditional for when 0 is passed is for the split where the nfold is set to NA.

  # Set `n_fold` to `NA` if not using cross-validation.
  if (split[2] != 0) {
    n_fold <- NA
  }

The crash was because the original vfold_cv should not be taking an NA for the number of folds, in this mode the existing validation partition should be used rather than creating new folds. I ran with the validation split and a cv split and I see similar, overlapping top importance score features.

If this solution makes sense to you, ill approve for merging