fix(owlgen): warn on single-child covering axiom equivalence

[jdsika]

Summary

Emit a warning when an abstract class has only one direct is_a child, since the covering axiom degenerates to an equivalence (Parent ≡ Child). The warning recommends --skip-abstract-class-as-unionof-subclasses to suppress covering axioms if the equivalence is undesired.

Motivation

The covering axiom feature (linkml#3219) correctly expresses that every instance of an abstract class must belong to one of its subclasses. With a single child, this is semantically equivalent to Parent ≡ Child per OWL 2 Primer §4.2. While OWL-correct, this may surprise users building extensible ontologies where more children are added downstream — the equivalence causes new siblings to inherit constraints from the first child via RDFS transitivity.

Rather than silently skipping the axiom (which would change the OWL semantics), a warning alerts users to the implication and points them to the existing escape hatch.

Changes

• owlgen.py: Add logger.warning() when an abstract class has exactly 1 child before the covering axiom is emitted. Update field docstring and CLI help text to document the warning.
• test_owlgen.py: Add test_abstract_class_with_single_child_emits_warning (validates warning is logged with flag recommendation) and test_single_child_warning_suppressed_by_skip_flag (validates no warning when skip flag is set).

Design Decision

Per review feedback from @mgskjaeveland (linkml#3309): the single-child equivalence is not a generator error — it is the expected OWL semantics. The proper response is a warning, not a silent skip. Users who need to avoid the equivalence should use --skip-abstract-class-as-unionof-subclasses.

References

• W3C OWL 2 Primer §4.2 — bidirectional rdfs:subClassOf = owl:equivalentClass
• W3C OWL 2 Syntax §9.1.4 — DisjointUnion as proper covering construct
• feat(owlgen): add covering axiom for abstract classes  linkml/linkml#3219 — Original covering axiom feature
• fix(owlgen): warn on covering axiom edge cases for abstract classes linkml/linkml#3309 — Upstream discussion

[jdsika]

🔍 Adversarial Review — PR #5

Summary

Solid, well-motivated PR. The OWL 2 semantics analysis is correct and the warning-not-skip approach aligns with upstream reviewer feedback. No actual bugs found — but there are test coverage gaps and a couple of design concerns worth discussing before merge.

---

🐛 Bugs & Issues

No code bugs found. The control flow is correct:

if not children → warn (no axiom)
elif len == 1  → warn (equivalence)
if children    → emit axiom (for 1+ children)

The if/elif warning block and the separate if children: axiom block are intentionally decoupled — 1-child case fires both the elif warning AND the if children: axiom. This is correct per the design decision.

One correctness note on the test assertion:

test_abstract_class_with_single_child_emits_warning asserts (EX.Parent, RDFS.subClassOf, EX.Child) in g. This is correct because _boolean_expression returns exprs[0] directly for single-element lists (line ~1325), so _union_of([child_uri]) → child_uri, producing the raw Parent rdfs:subClassOf Child triple (no owl:unionOf wrapper). Combined with Child rdfs:subClassOf Parent from is_a, this is indeed bidirectional subClassOf = equivalence per OWL 2 §4.2. ✅

---

⚠️ Concerns

1. No-children warning suppressed by skip flag — intentional?
   When skip_abstract_class_as_unionof_subclasses=True, the no-children warning is also suppressed (tested by test_no_children_warning_suppressed_by_skip_flag). An orphan abstract class (no subclasses at all) might be worth warning about regardless of the skip flag — it suggests an incomplete class hierarchy, not just a covering axiom concern. Consider whether this warning should live outside the if not self.skip_abstract_class_as_unionof_subclasses: guard.

2. CLI help text ambiguity:
   The added text reads: "Note: warnings are emitted for abstract classes with zero children (no axiom) or one child (equivalence)."
   This doesn't clarify that warnings fire when the flag is NOT set. A reader might wonder if warnings still appear with --skip-abstract-class-as-unionof-subclasses. Suggest: "Note: without this flag, warnings are emitted for abstract classes with zero children (no axiom) or one child (equivalence)."

3. Unicode ≡ in warning message:
   "The covering axiom makes them equivalent (%s ≡ %s)." — the ≡ symbol could cause UnicodeEncodeError in non-UTF-8 terminals or CI log sinks. Consider (%s == %s) or (%s ≡ %s [equivalentClass]) with an ASCII fallback, or just spell it out: "making them equivalent".

4. _union_of single-element optimization is the root cause:
   The real issue is that _boolean_expression silently collapses single-element unions into the bare element. For covering axioms, this transforms Parent ⊑ unionOf(Child) (which is just Parent ⊑ Child) into a raw rdfs:subClassOf triple that's indistinguishable from a manually-asserted subsumption. A future PR might consider explicitly emitting owl:equivalentClass when len(children) == 1, making the intent machine-readable. (Out of scope for this PR, but worth a tracking issue.)

---

🧪 Test Coverage Assessment

Covered well:

• ✅ 0 children → warning + no axiom
• ✅ 0 children + skip flag → no warning
• ✅ 1 child → warning + axiom emitted (with bidirectional subClassOf check)
• ✅ 1 child + skip flag → no warning + no axiom

Missing scenarios:

1. 2+ children → no warning emitted (negative test). test_abstract_class_gets_union_of_subclasses_by_default exists on main but doesn't use caplog to verify absence of warnings. A regression could add spurious warnings for the normal multi-child case without any test catching it.

2. Non-abstract class with 1 child → no warning. test_non_abstract_class_does_not_get_union_of_axiom exists but doesn't check warnings. A concrete class with a single child should never trigger the equivalence warning.

3. Abstract class with mixins (not is_a) children. The code uses mixins=False, is_a=True — a class with mixin-only "children" should get 0 is_a children and trigger the no-children warning. This edge case is untested.

Style nit: All 4 tests do import logging inside the function body. logging is a stdlib module — just add it to the module-level imports alongside pytest, rdflib, etc.

---

📋 Fix Plan

1. Add negative warning test for 2+ children:

   def test_abstract_class_with_multiple_children_no_warning(caplog):
       import logging
       with caplog.at_level(logging.WARNING, logger="linkml.generators.owlgen"):
           g = _owl_graph(_build_abstract_schema())  # Animal -> Dog, Cat
       assert _union_members(g, EX.Animal) is not None
       assert not any("equivalent" in msg.lower() for msg in caplog.messages)
       assert not any("has no children" in msg for msg in caplog.messages)

2. Clarify CLI help text — prepend "without this flag" to the Note sentence.

3. Move import logging to module-level in the test file.

4. Consider decoupling no-children warning from skip flag (design discussion — not blocking).

5. Open tracking issue for explicit owl:equivalentClass emission when single-child covering axiom is detected (future improvement).

---

✅ What's Good

• OWL 2 semantics are correct — the equivalence claim (Parent ≡ Child from bidirectional rdfs:subClassOf) is well-grounded in OWL 2 Primer §4.2 and OWL 2 Syntax §9.1.4.
• Warn-don't-skip design respects @mgskjaeveland's upstream feedback on fix(owlgen): warn on covering axiom edge cases for abstract classes linkml/linkml#3309 — the axiom IS semantically correct, so suppressing it silently would be wrong.
• Lazy %-formatting for logger.warning() — correct Python logging style (no eager f-string evaluation).
• logger = logging.getLogger(__name__) — properly configured, matches the caplog logger name in tests.
• Thorough inline comments explaining the equivalence derivation.
• PR description is excellent — cites specs, references upstream issues, documents the design decision.