Expand reasoning-failure examples and negative fixtures across ten families

[mdheller]

Context

Follow-up from #38 after #45 landed the first Reasoning Failure ontology tranche.

#45 added:

• Platform/ReasoningFailure/reasoning-failure.ttl
• shapes/reasoning_failure.shacl.ttl
• contexts/reasoning-failure.context.jsonld
• examples/reasoning_failure_demo.ttl
• catalog/reasoning_failure_registry.ttl
• docs/reasoning-failure-ontology-v0.1.md

The core semantic surface exists, but #38 also required positive and negative examples across at least ten failure families.

Scope

Add broader fixtures for reasoning failure coverage:

• positive examples across at least ten failure families
• negative fixtures for fail-closed behavior
• explicit reversal / inverse relation fixture
• explicit two-hop shortcut exploitation fixture
• missing invariant fixture
• missing verifier fixture
• missing evidence receipt fixture
• missing residual risk fixture
• ambiguous relation direction fixture

Candidate failure families

• cognitive state / stale context
• bias and framing
• formal relation reasoning
• exact string / tokenization / arithmetic
• impossible or contradictory task handling
• social reasoning and anthropomorphic trust
• code perturbation robustness
• causal or temporal reasoning
• multimodal / embodied / spatial reasoning
• scientific reasoning
• multi-agent misalignment / termination / obfuscation

Acceptance criteria

• Fixtures are placed so invalid examples do not break the repo-wide examples/ SHACL gate.
• A focused validator proves valid examples pass and invalid fixtures fail.
• make validate or a named make target includes the focused validator.
• make all remains green.
• Documentation states which failure families are covered.

Non-goals

• No runtime perturbation runner.
• No model evaluation service.
• No ledger storage.
• No UI or search implementation.