Add persistent background alert investigation and native-channel report delivery #88

MOSAIC currently needs an open browser to trigger alert investigations, and native messaging configuration alone does not publish completed reports.

Propose an optional persistent worker that:

• polls an existing Alertmanager using explicit operator-selected label filters;
• records each fingerprint/start-time incident in durable state and mirrors the alert into MOSAIC;
• invokes the existing headless API in a dedicated read-only automation session, collects actual tool history, consults IRAOP, and generates a complete report;
• publishes through the already configured OpenClaw native channel with a fixed operator-selected recipient, separate from the read-only agent's tools;
• preserves all stage requests, responses, tool results, source references and delivery receipts, and avoids duplicate processing across restarts;
• installs as a user systemd service on an existing deployment, with no browser dependency or changes to model serving, cluster write permissions, or default Helm behavior.

The initial deployment and validation target is Spark with OpenClaw 2026.6.10 and an existing Feishu application. View-mode tool policy remains enforced; native publication uses the existing operator CLI adapter rather than enabling mutating tools in the agent session. Unknown delivery outcomes are recorded for reconciliation rather than blindly retried. A new controlled test alert will be used to validate the full path, with complete evidence retained.

Related: #74 (native Feishu), #78 (authenticated push receiver), #86 (research evidence/protocol compatibility). This worker can poll an existing Alertmanager independently of the push receiver.