EMEP Unit Test Plan Per Component

Unit test coverage targets and test categories for every EMEP component from ModelRegistry to DeploymentManager.

Unit tests cover every EMEP component in isolation. Each component has a coverage target, a list of critical test categories, and a defined mock strategy for external dependencies.

Coverage Table

Component	Coverage Target	Critical Test Categories	Mock Strategy
ModelRegistry	90%	CRUD, lifecycle transitions, search, concurrency	In-memory DB
ModelLoader	85%	Format parsing, checkpoint loading, error handling	Mock filesystem
ModelCompatibilityAnalyzer	90%	Tensor shape validation, tokenizer compatibility, architecture compatibility, state transitions	Mock ModelRegistry
TensorEngine	85%	Element-wise ops, broadcasting, dtype coercion, shape mismatch	CPU-only backend
MergeEngine	85%	Strategy dispatch, parameter validation, artifact storage	Mock TensorEngine, Mock ArtifactStore
MergeStrategy	90%	SLERP, TIES, DARE, Task Arithmetic, Franken-Merge correctness	Fixed random seeds
CandidateGenerator	80%	Genome encoding, decoding, validation	Mock MergeEngine
EvolutionEngine	80%	NSGA-II, NSGA-III, CMA-ES loop correctness	Mock CandidateGenerator, Mock FitnessEngine
EvaluationEngine	85%	Benchmark dispatch, result aggregation, timeout handling	Mock BenchmarkEngine
BenchmarkEngine	85%	Harness execution, seed control, backend switching	Mock InferenceBackend
FitnessEngine	80%	Objective computation, normalization, multi-objective aggregation	Mock EvaluationEngine
ExperimentTracker	85%	State transitions, schema validation, query	In-memory store
ArtifactStore	80%	Upload, download, signing, encryption, retention	Mock S3 / local filesystem
DatasetRegistry	80%	Split management, format validation, caching	Mock filesystem
QuantizationEngine	85%	NF4, INT8, FP8 quantization and dequantization	CPU-only
InferenceBackend	80%	Load, forward, batching, unload	Mock GPU layer
DeploymentManager	75%	Promotion, rollback, stage validation	Mock ArtifactStore

Test Patterns

Lifecycle Transition Tests

For every component with state machines (ModelRegistry, ExperimentTracker), unit tests verify all valid transitions and reject invalid ones. Example: a model cannot transition from DISCOVERED directly to PROMOTED.

Error Path Tests

Every public method must have tests for:

• Null or missing required arguments
• Invalid enum values
• Resource not found
• Timeout or resource exhaustion
• Concurrent modification

Determinism Tests

MergeStrategy and TensorEngine tests use fixed random seeds and compare outputs against golden values. Any nondeterminism is treated as a bug.

Mock Strategy

External dependencies are mocked at the interface boundary. Mocks are validated against the real implementation in integration tests. Mock contracts are documented in the component’s test directory.

Integration Points

• CI: unit tests run on every commit.
• Coverage: coverage reports are uploaded to ArtifactStore and tracked in ExperimentTracker.
• Numerical Testing: unit tests for TensorEngine and MergeStrategy reference tolerance thresholds defined in Numerical Testing.