feature: Robustness suite — semantic tier with human sign-off

Parent

#2087

What to build

Extend the Robustness suite with the semantic tier per docs/specs/eval-framework.md §4: full paraphrase/rewording perturbations, for both the invariance and sensitivity families. Semantic perturbations may be LLM-drafted, but every one requires human sign-off confirming it actually preserves (invariance) or changes as intended (sensitivity) the meaning before it enters the corpus — the human review is the real ground-truth authority; the LLM only saves drafting time.

The sensitivity family's edit-type coverage carries over from the mechanical tier (#2466) — including the restriction-word case ("only", "just" narrowing or broadening a grant) — but expressed as natural paraphrase instead of literal word insertion (e.g. rewording a grant to "solely" or "exclusively" cover a subset, rather than mechanically inserting "only").

Acceptance criteria

• Semantic-tier invariance and sensitivity perturbations exist for the primary scenario set, drafted with LLM assistance.
• The semantic-tier sensitivity set includes at least one paraphrased restriction case (e.g. "solely"/"exclusively"/equivalent narrowing language), not just the mechanical-tier's literal "only"/"just" insertion.
• A documented sign-off step/record exists and is required before a semantic perturbation is added to the corpus — no semantic perturbation ships without it.
• The suite runs both tiers (mechanical + semantic) and both families, still scored as the same two separate metrics.

Blocked by

#2466 (needs the mechanical-tier harness and the invariance/sensitivity scoring it introduces).