feat: unify competency questions across ontology design and coverage evaluation

[rajnavakoti]

Description

Competency Questions (CQs) are a single artifact that serves both directions of ontology work: they scope the ontology at design time (requirements) AND validate it at eval time (test suite). This is the classic Grüninger & Fox (1995) framing — the questions you write to define an ontology are the same questions you use to verify it.

CB today does not use CQs. evals.py generates eval questions from four sources — seed, doc, persona, work-item — tagged with layer hints. That's source-driven coverage testing, related to but distinct from ontology-competency-driven testing.

Opportunity: make CQs a first-class per-block artifact used for both (a) scoping/generating the ontology and (b) measuring coverage — where coverage = % of the block's competency questions the agent can answer. One artifact = design tool + eval suite + the concrete definition of the "context coverage" thesis.

Background / references (so the lineage isn't lost)

• Grüninger & Fox (1995) — origin of competency questions; CQs as both requirements and tests.
• METHONTOLOGY, On-To-Knowledge, NeOn methodology — established ontology engineering methodologies that formalize CQ-driven design.
• Ontology learning (Maedche & Staab) — automatic ontology induction from text; the technique behind "AI generates the ontology when the user can't define one."
• Competency-question-driven ontology learning — the hybrid: seed → competency questions → ontology (more rigorous + publishable than seed → ontology).
• CQs as ontology unit tests — using the CQ set as a regression/coverage test.
• LLM-generated competency questions — emerging (2023-24); VERIFY current state before citing.

Acceptance Criteria

• Decide whether CQs become a first-class per-block artifact (e.g. competency-questions.yaml)
• Spec CQ authoring: human-defined AND/OR AI-generated from seed context (competency-question-driven ontology learning)
• Spec how the CQ set feeds ontology scoping at design time
• Spec how the CQ set feeds coverage eval (coverage = % answerable), reconciled with the current 4-source eval in evals.py
• Document the design↔eval unification in the architecture docs

Out of Scope

• Implementation of AI ontology generation itself (separate concern)
• The cache/router retrieval mechanics (separate concern)
• Choosing/citing specific recent LLM-CQ papers (research pass, do later)