Add a pre-judge validation layer

[RGGH]

try code-based validation first before calling LLM judge (faster, deterministic, free!)

[RGGH]

add new .service() block for /judge-prompts with endpoints:

GET /judge-prompts - Get all
POST /judge-prompts - Create new
GET /judge-prompts/active - Get active
PUT /judge-prompts/active - Set active
GET /judge-prompts/{version} - Get by version

[RGGH]

eg

        ContainsKeyword("unfortunately"),
        MatchesRegex(r"\d+"),