[rmkr-dev]

Adds the llm-gateway-caching-guardrails skill and matching add-llm-gateway-caching prompt.

The cost/latency entry prices cached tokens and the routing entry tracks capabilities, but nothing says when a gateway may reuse an answer or how to keep a cache from leaking across tenants or serving the answer to a different question. This pair covers that.

• Three tiers rolled out in order of risk: provider prompt caching (no answer reuse), an opt-in exact response cache, and a semantic cache only on routes that qualify.
• Prompt caching: static content first and byte-stable, per-provider markers and limits kept in config with a checked-on date, write premiums only where the prefix repeats, cached read/write tokens recorded for the cost ledger.
• Exact cache key covers tenant, auth scope, resolved model, prompt and tool versions, normalized redacted messages, generation parameters, and corpus version. Only complete, valid, non-refused, non-tool responses are stored, with a TTL.
• Semantic tier searches only inside exact-match partitions, skips requests with IDs, amounts, dates, or code, and takes its threshold from a labeled near-miss pair set per route, with the similarity-vs-distance direction written down.
• Invalidation by version bump, source tag, and erasure request, plus a kill switch and purge per route and tier.
• X-Cache status header, poisoning and leakage controls, metrics without payloads, and CI tests for keys, eligibility, near misses, invalidation, and cache outages.

Cross-links to the cost/latency, routing, PII redaction, structured output, RAG retrieval, eval gate, and Spring Redis cache entries rather than repeating them.

Only skillbook/skills.json and promptbook/prompts.json change. Catalogs are 56 skills and 56 prompts. Local site check and script tests passed.