[rmkr-dev]

Adds the llm-pii-redaction-guardrails skill and matching add-llm-pii-redaction prompt.

Nothing in the catalogs already covered redaction for an LLM gateway. Logging, prompt-injection, eval, and gateway routing mention the problem and leave the pipeline to another entry. This pair is that pipeline.

• Redact once, before the provider call, and reuse that view for logs, traces, prompt caches, and eval fixtures.
• Deterministic detectors (email, phone, Luhn-checked cards, named national IDs, API keys, JWTs), an optional recognizer, and an allowlist.
• Per-route mode: mask, typed placeholder, or reversible tokens with a request-scoped vault. Secrets are never restored. Restore only on the response to that caller.
• Overlapping spans redact the union of every surviving finding. A slice that meets an API key, JWT, or other secret stays non-restorable, including when a longer personal-data span contains it or the spans only partly overlap. Reversible-mode tests cover that nested case and a partial overlap. Combining spans does not mean the detectors found every secret.
• Stream hold-back so a value or placeholder split across chunks is still caught.
• Fail-closed by default, with fail-open only where a route opts in.
• Counters by entity type and action, with no raw values. Latency sits in the gateway budget.
• Tests include golden fixtures, false positives and negatives, a leak test against a stub provider and the log appender, and the nested and partial-overlap cases above.

Only skillbook/skills.json and promptbook/prompts.json change. Catalogs are 54 skills and 54 prompts. Local site check passed.