[rmkr-dev]

Adds the llm-structured-output-guardrails skill and matching enforce-llm-structured-output prompt.

The LLM entries so far touch structured output only in passing: prompt injection says to validate tool arguments, routing skips targets without json_schema, and the eval gate lists JSON validity as a hard check. None of them says how a service actually gets a schema-valid document or a clean failure. This pair covers that.

• Native schema mode or strict tools where the target supports them, chosen from capability config. JSON mode and prompt-only targets are named explicitly and get the same checks.
• Schemas are versioned files in the repo: object root, additionalProperties: false, every property required (nullable instead of optional), enums with an escape value, no free-form unions. Rules a provider's strict subset rejects move into the validator instead of being deleted.
• Fixed check order: stop reason first (refusal, truncation, filter, unknown stop are typed outcomes), then extract, size cap, strict parse, JSON Schema validation, bind, business rules.
• Server-side validation even in native mode. networknt json-schema-validator is the default (3.x for Jackson 3 / Boot 4, 2.x for Jackson 2), compiled at startup, no remote $ref, format assertions on when a field depends on format.
• One repair with {path, keyword} feedback and no values, then fail closed with a problem+json code or a sealed result type. No guessing fixers, no relaxing the schema on retry.
• Model output is never evaluated. Tool calls need an allowlisted name, server-side validation, and end-user authorization; tenant and user never come from arguments.
• Counters by route, schema version, mode, outcome, and keyword. No payloads or values in logs, labels, or spans.
• Offline conformance fixtures on every PR plus a first-attempt conformance rate per schema version in the existing eval gate.

Cross-links to the prompt-injection, routing, PII redaction, eval gate, and cost/latency entries rather than repeating them.

Only skillbook/skills.json and promptbook/prompts.json change. Catalogs are 55 skills and 55 prompts. Local site check and script tests passed.