Benchmark coverage: close the gaps found by the consistency-type audit #94

The coverage audit (docs/internals/verified-completions/docs/COVERAGE-AUDIT.md, PR #90) checked the corpus against the full space of logic-resolvable internal inconsistency. Verdict: the measured core is solid (12 families, n=30 each) but NOT exhaustive. Two workstreams close the gap:

1. Corpus-only - rules already exist, no engine change:

• disjoint class family (entity in two mutually exclusive classes) - rule V3
• domain violation family (attribute on impossible entity type) - rule V4
• identity families (distinct-yet-merged, shared unique identifier) - rules I1/I2
• self-causation flavors (acyclic rule already catches self-edges)

2. New rules + corpus families - pure logic, engine capabilities sufficient (Datalog + comparisons + aggregation, all shipped):

• sum/partition overflow: budget 2000 = flights 1500 + hotel 800 (sum aggregation exists; the benchmark doc wrongly filed this as out of reach - corrected in the audit)
• percentage partitions exceeding 100
• act-after-death / before-birth (activity date vs lifespan)
• age-date arithmetic (born 1990, 'was 20 in 2005')
• duration mismatch (3-hour flight, 10:00 to 12:00)
• point-in-interval (meeting 'during' a dated conference, outside its dates)
• symmetric-relation polarity ('X married Y' + 'Y not married to X')
• same-time vs strict order
• relation functionality (single direct manager - needs a seed decision)
• direction 3-cycles and opposite-direction pairs, distance symmetry
• effect-before-cause via event dates
• taxonomic subclass violations (subclass_of as extractor DATA + one fixed rule)

Explicitly excluded (documented in the audit, revisit deliberately): general quantifiers, counterfactuals, sarcasm, world-falsity without a ground-truth KG, unit conversion, closed-world absence, conditional policies, calendar semantics, threshold spatial reasoning.

Every new rule follows the one-clause-per-detection-relation discipline (#91) and every new family ships at n=30 with clean twins and controls, wired through full_bench.py.