rustlavel-otel: OpenTelemetry traces and metrics, over OTLP.

App::new()?
    .routes(routes::web::routes)
    .plugin(OpenTelemetry::default())
    .serve()
    .await

Nothing is instrumented twice. Every package in the framework already reports what it does on rustlavel_core::events; this listens, the way Telescope and the Prometheus exporter do, and turns what it hears into OTLP. Adding a span to a new package therefore means dispatching an event, not reaching for this crate.

What it produces

• A server span per request, named for the route pattern, carrying the method, path, status and route. Everything the request awaits — a query, a model call, an MCP tool, an outbound request — becomes a child span.
• Metrics for the same work: request, query, model and job durations as histograms, token usage and job outcomes as counters, and the exporter's own queue depth and drop count so a broken telemetry pipeline is visible from inside the telemetry.

Propagation

W3C traceparent is read from the request and written to the response, so a trace that began upstream continues here instead of starting again. For an outbound call, attach [trace::traceparent] to carry it onwards.

Encoding

Protobuf by default, because every collector must accept application/x-protobuf while JSON is optional. The encoder is written here rather than generated ([protobuf]), which rule one asks for and which the narrow, fixed schema makes reasonable. application/json is a configuration switch away when a payload needs to be readable in a proxy log.

Failure

Telemetry must not be able to take down what it observes. Recording a span is a mutex and a push; the queue is bounded and drops with a counter and a log line; an unreachable collector costs spans and never a request. See [exporter] for the drop and retry policy in full.