breakpilot-compliance/backend-compliance/tests/test_capability.py

"""Tests for Master Capability Registry v0 (Phase 2C, Compliance Execution domain).

Acceptance: a registry mints stable MCAP ids, stores typed relations + versioned
policy + lifecycle events + provenance, and DERIVES confidence/status on demand
(never stored). Hard rule: a certification alone can never yield CONFIRMED.

No real ISO/cert mappings here — only synthetic relations (mappings are not part of
v0; they are Execution data, injected later).
"""

from __future__ import annotations

import pytest

from compliance.capability import (
    DEFAULT_POLICY,
    AssertionStatus,
    CapabilityCandidate,
    CapabilityRegistry,
    CapabilityRelation,
    Confidence,
    EvidenceKind,
    LifecycleEventType,
    PolicyRule,
    PolicyVersion,
    RelationType,
    assert_no_certification_confirms,
    deprecate_capability,
    evaluate_relation,
    merge_capabilities,
    mint_capability,
    resolve,
)


def _rel(rt, ek, target="MCAP-00001", src="x"):
    return CapabilityRelation(
        relation_id="r", source=src, target_capability_id=target, relationship_type=rt, evidence_kind=ek
    )


# 1. minting assigns stable, incrementing MCAP ids.
def test_mint_stable_mcap_id():
    reg = CapabilityRegistry()
    a = mint_capability(reg, CapabilityCandidate(raw_term="Patch Management"))
    b = mint_capability(reg, CapabilityCandidate(raw_term="Incident Response"))
    assert a.capability_id == "MCAP-00001"
    assert b.capability_id == "MCAP-00002"
    assert reg.capabilities["MCAP-00001"].name == "Patch Management"


# 2. minted capability carries provenance.
def test_mint_records_provenance():
    reg = CapabilityRegistry()
    cap = mint_capability(reg, CapabilityCandidate(raw_term="Secure Development"))
    assert cap.provenance.author == "system"
    assert "Secure Development" in cap.provenance.basis


# 3. all seven relationship types exist.
def test_relation_types_complete():
    assert {t.value for t in RelationType} == {
        "equivalent", "supports", "requires", "confirms", "broader_than", "narrower_than", "related_to",
    }


# 4. confidence is COMPUTED from relation_type + evidence + policy_version, not stored.
def test_confidence_computed_and_policy_referenced():
    rel = _rel(RelationType.SUPPORTS, EvidenceKind.CERTIFICATION)
    a = evaluate_relation(rel, DEFAULT_POLICY)
    assert a.status == AssertionStatus.INFERRED and a.confidence == Confidence.LOW
    assert a.policy_version == "capability-policy-v0"
    # the registry itself stores NO confidence/coverage
    assert "confidence" not in CapabilityRegistry.model_fields
    assert "coverage" not in CapabilityRegistry.model_fields


# 4b. a DIFFERENT policy version yields a different result for the SAME relation
#     -> "why did you say X last year" needs the policy-as-of-then.
def test_policy_versioning_changes_outcome():
    rel = _rel(RelationType.SUPPORTS, EvidenceKind.CERTIFICATION)
    v0b = PolicyVersion(
        policy_version="capability-policy-v0b",
        rules=[PolicyRule(
            relationship_type=RelationType.SUPPORTS, evidence_kind=EvidenceKind.CERTIFICATION,
            status=AssertionStatus.INFERRED, confidence=Confidence.MEDIUM,
        )],
    )
    assert evaluate_relation(rel, DEFAULT_POLICY).confidence == Confidence.LOW
    assert evaluate_relation(rel, v0b).confidence == Confidence.MEDIUM
    assert evaluate_relation(rel, v0b).policy_version == "capability-policy-v0b"


# 5 + 8. HARD RULE: a certification alone can NEVER be CONFIRMED.
def test_certification_never_confirmed():
    for rt in (RelationType.SUPPORTS, RelationType.EQUIVALENT, RelationType.BROADER_THAN, RelationType.RELATED_TO):
        a = evaluate_relation(_rel(rt, EvidenceKind.CERTIFICATION), DEFAULT_POLICY)
        assert a.status != AssertionStatus.CONFIRMED
    # only a concrete artifact via a CONFIRMS relation reaches CONFIRMED
    assert evaluate_relation(_rel(RelationType.CONFIRMS, EvidenceKind.ARTIFACT)).status == AssertionStatus.CONFIRMED
    # the shipped policy structurally satisfies the hard rule
    assert_no_certification_confirms(DEFAULT_POLICY)


# 8b. a policy that maps a certification to CONFIRMED is rejected.
def test_bad_policy_rejected():
    bad = PolicyVersion(
        policy_version="bad",
        rules=[PolicyRule(
            relationship_type=RelationType.EQUIVALENT, evidence_kind=EvidenceKind.CERTIFICATION,
            status=AssertionStatus.CONFIRMED, confidence=Confidence.HIGH,
        )],
    )
    with pytest.raises(ValueError):
        assert_no_certification_confirms(bad)


# 6. provenance on a curated relation atom.
def test_relation_carries_provenance():
    rel = CapabilityRelation(
        relation_id="r1", source="certification:ISO27001", target_capability_id="MCAP-00001",
        relationship_type=RelationType.SUPPORTS, evidence_kind=EvidenceKind.CERTIFICATION,
    )
    assert rel.relationship_type == RelationType.SUPPORTS
    assert hasattr(rel, "provenance")


# 9. merge keeps a redirect: resolve(old) follows it to the new capability.
def test_merge_keeps_redirect():
    reg = CapabilityRegistry()
    old = mint_capability(reg, CapabilityCandidate(raw_term="Update Management"))
    new = mint_capability(reg, CapabilityCandidate(raw_term="Software Update Management"))
    event = merge_capabilities(reg, old.capability_id, new.capability_id)
    assert event.event_type == LifecycleEventType.MERGE
    assert reg.capabilities[old.capability_id].redirect_to == new.capability_id
    # resolve follows the redirect to the canonical capability
    assert resolve(reg, old.capability_id).capability_id == new.capability_id
    assert reg.lifecycle_events[-1].from_ids == [old.capability_id]


# 9b. deprecate without a redirect resolves to None (no canonical target).
def test_deprecate_without_redirect_resolves_none():
    reg = CapabilityRegistry()
    cap = mint_capability(reg, CapabilityCandidate(raw_term="Legacy Capability"))
    deprecate_capability(reg, cap.capability_id)
    assert resolve(reg, cap.capability_id) is None
    assert reg.capabilities[cap.capability_id].state.value == "deprecated"


# requires = an obligation NEEDS a capability (relevance), not possession -> unknown.
def test_requires_is_relevance_not_possession():
    a = evaluate_relation(_rel(RelationType.REQUIRES, EvidenceKind.NONE), DEFAULT_POLICY)
    assert a.status == AssertionStatus.UNKNOWN