Tool spans already carry duration_ms and status, but the metrics layer
only counted them. Expose that data:
- GetTopTools now returns avg/p95 duration and error count per tool.
- Timeseries buckets gain tool_avg_ms / tool_p95_ms (filtered
percentile_cont over tool spans).
- Dashboard Top Tools shows avg latency per tool; the Latency panel,
previously always empty (it read run-level duration that is never
emitted), now plots real tool-span latency (min/avg/p95).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
William Valentin
5014d89258