feat: MCP audit middleware and human-in-the-loop approval gate

[Copilot]

Adds centralised auditing of every MCP tool call and an attribute-driven approval gate that pauses dangerous actions pending human confirmation via Signal poll.

Audit Middleware

• McpAuditMiddleware implements the AIAgentBuilder.Use(...) function invocation delegate — captures tool name, arguments, duration, success/failure for every call
• Emits structured log at Information level per invocation (plugs into existing Serilog/OTel pipeline)
• Auto-wired into the server-side agent pipeline via HausServiceCollectionExtensions.CreateAgent — no changes needed to individual McpQueryService classes

Human-in-the-Loop Approval

• [RequiresApproval] attribute marks methods that gate on human confirmation before execution
• Middleware sends a Signal poll ("⚠️ Agent wants to call '{ToolName}'. Approve?" → Yes/No), awaits vote via existing IPollTracker, then proceeds or denies
• Fail-safe: timeout defaults to deny
• Config overrides (AlwaysRequireApproval / NeverRequireApproval) allow ops control without redeployment

Decorated Tools

• UnlockHouseDoor — physical security
• SetSmartPlugPower — appliance power control
• SwitchOffAllHouseLights — bulk state change

Configuration

"AuditConfig": {
  "Enabled": true,
  "RedisStreamKey": "audit:mcp:calls",
  "MaxStreamLength": 10000,
  "ApprovalEnabled": true,
  "ApprovalTimeoutMs": 60000,
  "AlwaysRequireApproval": [],
  "NeverRequireApproval": []
}

Development config disables approval and exempts all current tools for unblocked local iteration.