[medgan-pr-prep] T8: Integration tests (8 MECE in-memory + 4 MIMIC-III skip-gracefully)

[jalengg]

Goal

Create tests/integration/test_medgan_end_to_end.py covering all 8 integration seams, following the same bootstrap pattern used for CorGAN.

Bootstrap pattern

Same as CorGAN: stub pyhealth.datasets and pyhealth.models via sys.modules, load model via importlib.util.spec_from_file_location. This avoids litdata/pyarrow/dask import failures in CI.

Small test config

_SMALL_MODEL_KWARGS = {
    "latent_dim": 4,
    "hidden_dim": 4,
    "autoencoder_hidden_dim": 4,
    "discriminator_hidden_dim": 8,
    "ae_epochs": 1,
    "gan_epochs": 1,
    "batch_size": 4,
    "save_dir": None,
}

Sample data: flat list ["A", "B", "C"] (5–8 codes per patient, ≥4 patients).

Category A tests (8, always run)

Test	Seam
TestMedGANIsBaseModelInstance	Inheritance contract
TestMedGANFeatureKeys	model.feature_keys == ["visits"]
TestMedGANInputDim	model.input_dim == processor.size()
TestMedGANForwardRaisesNotImplementedError	forward() contract
TestMedGANTrainModelRuns	Training loop (compute)
TestMedGANSynthesizeCount	synthesize_dataset(5) returns 5 dicts
TestMedGANSynthesizeOutputStructure	Output is flat list of str (not nested)
TestMedGANSaveLoadRoundtrip	Checkpoint I/O

Category B tests (4, skip when MIMIC-III absent)

_MIMIC3_PATH = os.environ.get(
    "PYHEALTH_MIMIC3_PATH",
    "/srv/local/data/physionet.org/files/mimiciii/1.4",
)

Test	What it verifies
TestMedGANMIMIC3TaskFunction	MedGANGenerationMIMIC3 produces valid samples
TestMedGANMIMIC3FullPipeline	Dataset → task → model → train (1 epoch)
TestMedGANMIMIC3Synthesize100	synthesize_dataset(100) succeeds
TestMedGANMIMIC3SynthesizeCodeQuality	All generated codes exist in training vocab

Resource cleanup

Every test using tempfile.mkdtemp() must have tearDown: shutil.rmtree(self.tmpdir, ignore_errors=True).

Acceptance criteria

• All 8 Category A tests pass on a machine without MIMIC-III
• Category B tests are skipped gracefully (not errored) when MIMIC-III is absent
• No hardcoded absolute paths in bootstrap
• tests/integration/__init__.py exists (already created for CorGAN)