Problem

There is no visibility into provider health during a session. The cooldown tracker (if merged) records failures for circuit-breaking, but there is no success/error rate tracking, latency measurement, or health summary for diagnostics.

Proposed Solution

Extend the provider tracking to record per-provider stats: success count, error count, average latency, last error reason. Expose via /status slash command and hermes status CLI. Lightweight in-memory counters, no persistence needed. Pairs with the cooldown tracker as the observability layer.