Add permanent AgentLink hot reload

[Oryntai]

What changed

• install one permanent MCP against .agent-link/active.json and hot-reload room changes into existing GUI tasks
• add correlated peer_ask, peer_listen, and peer_respond tools so a responder can leave the MCP call, perform local read-only work, answer the exact request, and listen again
• keep the encrypted WebSocket connected after completing an individual structured goal
• isolate protocol state by room, preserve identities, and make repeated GUI installation idempotent
• update setup, protocol, security, operations, and coding-agent documentation for v0.3.0

Why

Previously every host or join wrote a different MCP config path, requiring a full GUI restart. Goal completion also closed the bridge, and uncorrelated exchanges were awkward for ad-hoc questions during normal feature work.

Verification

• npm run smoke
• five consecutive runs of node test/mcp-integration-test.js
• npm audit --omit=dev
• syntax checks for all source, script, and test JavaScript
• git diff --check
• staged secret scan

A local npm ci could not remove the ngrok native DLL because the owner's current test tunnel is still running. npm install restored dependencies and all local checks passed; GitHub Actions performs clean installs on fresh Windows and Ubuntu runners.