Adding a domain without a fork

Pluggability is easy to assert and the assertion is usually worth nothing, so here is the version git can check. Weather and esports each implement one protocol and register. Covenant, being first, is served by the service directly and is not itself a plug-in.

So the training and explanation code never learned that weather or esports exist. The protocol did: the second domain is where decode-and-extract became one call the domain owns, and the third is where construction was deferred so a deployment running only esports never loads weather's fitted state. That is what the second and third implementations are for.

class DomainProtocol(Protocol):
    """Protocol that all domain implementations must satisfy.

    A domain provides event codecs, feature extraction, alert context
    generation, and Kafka topic configuration. The GenericStreamingWorker
    delegates all domain-specific logic through this protocol.
    """

services/covenant-radar-api/src/covenant_radar_api/domains/protocols.py, the protocol both plug-ins implement.

Covenant compliance is exact rule evaluation: formulas, thresholds, frequencies. Breach risk is the learned path. They are separate code paths on purpose.

Conflating them would make a compliance answer probabilistic, which is worse than useless. A lender asking whether a borrower is in breach needs an exact answer from the contract terms, not a confidence score. The temptation is real because one model could nominally produce both.

The same idea runs through the model layer and the explanation layer. Twelve backends sit behind one interface, including gradient boosting I wrote from scratch competing against the production libraries in the same harness, and they are selectable per request.

Explainability follows the model rather than the endpoint: permutation importance for anything, SHAP tree explanations for tree models, plain and integrated gradients for the neural ones. One contract, three strategies, chosen by what is actually loaded.

Training on 40 real public credit, loan, default and bankruptcy datasets rather than generated data is the unglamorous half. Synthetic data is the usual shortcut, and it flatters everything.