AI Boundary Coverage: Both Sides of Every Threshold — Derived From the Rules, Not Guessed

'ai_prompts::derive_boundaries' walks the stored rule definition and vocabulary, returning per input path: both sides of every comparison (649 and 650 for 'score >= 650'), every member of enumerated domains, and every declared input that no rule constrains — an unconstrained input being a coverage gap worth reporting.

**The derivation is free**: no token cost, no model call. It runs on the stored definition and vocabulary types before any AI is involved.

**The AI step** generates test cases for each boundary. Each proposed case is executed server-side: 'suggest-tests' sends the case to the engine's trace endpoint and checks whether 'matched' contains the target boundary's rule. A case that doesn't exercise its target boundary is discarded — not offered. "Verified by execution" is a measurement, not a model claim.

**Coverage Gap Closer**: reads two independent measurements — suite coverage ('/scenario-runs/:id/coverage') and production dead rules ('/monitoring/rules/dead'). A rule dead in both places is the highest-priority gap to close, and gets cases first.