feat(core): OpenTelemetry metrics, Prometheus scrape, and HTTP RED #89

Status

The tracing and probe foundations are implemented, but the metrics path described here is still outstanding.

Implemented

• OpenTelemetry tracing for tool calls with Kubeflow and MCP attributes.
• Unauthenticated HTTP health and readiness endpoints.
• An optional OpenTelemetry dependency group in pyproject.toml.
• Safe behavior when OpenTelemetry dependencies are not installed.

Remaining

• Add OpenTelemetry metrics initialization with a Prometheus reader/exporter.
• Add an unauthenticated GET /metrics endpoint on the HTTP edge.
• Instrument tool call count and duration, including tool and success dimensions.
• Export a readiness gauge aligned with /ready.
• Add HTTP rate, error, and duration metrics.
• Add rate-limit and circuit-breaker metrics.
• Support OTLP metrics export when configured.
• Install the OpenTelemetry group in the published image; the current Dockerfile does not install it.
• Document Prometheus scraping, OTLP metrics, and NetworkPolicy guidance.
• Add focused unit tests and a stable metric-name contract.
• Preserve safe behavior when the optional metrics dependencies are unavailable.

Out of scope

• W3C trace-context propagation into Kubeflow SDK child spans.
• Langfuse or MLflow exporters.
• Helm ServiceMonitor resources.
• Multi-process metric aggregation.

This issue remains open until the remaining metrics implementation and documentation are merged.