CE Fallback Test

You are writing tests that validate fallback chain behavior. This is a deliberate opt-in — normal tests must NOT trigger fallbacks (the disable_fallbacks autouse fixture prevents this by default).

The Fundamental Rule

Normal test  →  disable_fallbacks autouse fixture applies automatically
                  Do NOT request enable_fallbacks
                  If a fallback fires anyway, fix the underlying issue

Fallback test → request enable_fallbacks fixture explicitly
                  Assert pytest.warns(UserWarning) is raised
                  Assert the operation still produced correct output

Minimal Template

import pytest

def test_should_<succeed_via_fallback>_when_<trigger_condition>(enable_fallbacks):
    """Validates the <X> → <Y> fallback path fires and emits a UserWarning."""
    # Arrange: set up the condition that triggers the fallback
    # (e.g., make a plugin unavailable, force a backend error, etc.)

    # Act + Assert: warn AND correct output
    with pytest.warns(UserWarning, match="<expected fragment from warning message>"):
        result = <operation_under_test>()

    # Assert the output is still correct (not just that it completed)
    assert result is not None          # execution-only — add semantic assertion too
    assert <domain_invariant_holds>    # e.g., interval invariant

Known Fallback Warning Messages (match= fragments)

Fallback	Match fragment
Parallel pool start-up failure → sequential	"Failed to initialize parallel pool"
Parallel force_serial recovery → sequential	"Parallel execution failed"
Joblib requested but missing → threads	"Joblib is not available"
Visualization → simplified bar	"Visualization fallback"
Plugin execution error → legacy path	"legacy" or "fallback"
Perturbation fallback	"perturbation"

Use match= with a substring, not the full message, to avoid brittleness:

with pytest.warns(UserWarning, match="fallback"):
    ...

INFO Log Side-Effect (mandatory per Fallback Visibility Policy)

Every fallback must also emit an INFO log. Assert it when testable:

import logging

def test_should_log_fallback_decision_when_plugin_fails(enable_fallbacks, caplog):
    with caplog.at_level(logging.INFO, logger="calibrated_explanations"):
        with pytest.warns(UserWarning):
            result = explainer.explain_factual(X_query)

    # Assert both the warning (user-visible) and the log (observability)
    assert any("fallback" in record.message.lower() for record in caplog.records)
    assert result is not None

Full Example: Parallel → Sequential Fallback

def test_should_explain_factual_when_parallel_backend_unavailable(
    enable_fallbacks, binary_explainer, monkeypatch
):
    """Parallel execution failure falls back to sequential and emits UserWarning."""
    # Arrange: force the parallel backend to raise on submission
    def broken_submit(*args, **kwargs):
        raise RuntimeError("Mock parallel backend failure")

    monkeypatch.setattr(
        "calibrated_explanations.parallel.backend.submit",
        broken_submit
    )

    # Act + Assert
    with pytest.warns(UserWarning, match="fallback"):
        explanations = binary_explainer.explain_factual(X_query[:2])

    # Semantic assertion: output is still valid
    for exp in explanations:
        assert exp.prediction['low'] <= exp.prediction['predict']
        assert exp.prediction['predict'] <= exp.prediction['high']

Not a UserWarning: cache backend and failed explicit requests

• The minimal in-package cache backend (used when cachetools is not installed) is a registered warning-policy exemption. It logs one WARNING per process when an enabled cache is built and emits no UserWarning, so assert it with caplog, not pytest.warns.
• An explicitly requested cache or parallel executor that cannot be configured raises ConfigurationError (fail closed) instead of falling back. Test it with pytest.raises(ConfigurationError) and assert details["capability"].

What NOT to Do

# ❌ WRONG: accessing enable_fallbacks in a test that has no fallback code path
def test_explain_factual_basic(enable_fallbacks):
    explainer.explain_factual(X_query)   # no fallback triggered; fixture is wasteful

# ❌ WRONG: asserting a fallback fires but not checking the output is still correct
def test_fallback_fires(enable_fallbacks):
    with pytest.warns(UserWarning):
        pass   # empty act — what is being tested?

# ❌ WRONG: swallowing the fallback warning silently
def test_fallback_fires(enable_fallbacks):
    result = explainer.explain_factual(X_query)
    # No pytest.warns; warning is not asserted — this is an incomplete fallback test

Out of Scope

• Writing standard (non-fallback) tests (see ce-test-author).
• Implementing a new fallback (see ce-fallback-impl).
• Auditing whether existing tests wrongly trigger fallbacks (see ce-test-audit).

Evaluation Checklist

• enable_fallbacks fixture requested.
• pytest.warns(UserWarning, match=...) wraps the act section.
• At least one semantic domain assertion on the output.
• INFO log side-effect asserted via caplog when feasible.
• match= uses a substring, not the full warning message.
• Test name follows test_should_<succeed_via_fallback>_when_<trigger>.