Canonical secrets policy for agents + best-effort redaction

- /security.md is the single canonical policy page (control plane principle,
  SECRET_TRANSFER_REQUIRED, out-of-band transfer via SSH)
- short version with link embedded in /create.md, room .md, llms.txt, landing,
  MCP create tool response; docs updated
- redactSecrets() applied on input to messages, questions, resolutions,
  contracts and room brief/goal/participant instructions (best-effort: PEM
  keys, JWTs, common token prefixes, password/token/secret assignments)

packages/mcp/src/index.ts

@@ -56,6 +56,7 @@ server.tool(
               .map((u, i) => `Participant "${created.participants[i].role}" invite URL (secret):\n${u}\nAgent-readable state: ${u}.md`)
               .join('\n\n') +
             `\n\nObserver URL (read-only, for the human to watch whose turn it is):\n${created.observer_url}` +
+            `\n\nSecurity: never put secret VALUES (passwords, API keys, tokens, private keys, cookies, credentials, .env contents) into this room — not in messages, questions, contracts or URLs. Discuss only a secret's name/source/purpose/transfer method; move values out-of-band (SSH/SCP) or record SECRET_TRANSFER_REQUIRED. Full policy: ${BASE_URL}/security.md` +
             `\n\nNext: YOU are participant "${created.participants[0].role}" — keep that token. ` +
             `Reply to the human with BOTH the other participant's invite URL and the observer URL. ` +
             `Then start by stating verified facts from your side.`,