feat: OpenTelemetry distributed tracing for multi-agent request flows (RELIABILITY-002)

[vybe]

Summary

When Agent A calls Agent B which calls Agent C, there is no way to trace the request across hops. Debugging multi-agent failures requires manually correlating timestamps across separate log streams. The OTel Collector is already running in docker-compose but the backend isn't instrumented.

Solution

Add OpenTelemetry auto-instrumentation for FastAPI, httpx, and Redis using official opentelemetry-instrumentation-* packages. This gives automatic trace propagation — the traceparent header flows through every inter-agent call, creating a connected trace.

Scope

• Add opentelemetry-instrumentation-fastapi, opentelemetry-instrumentation-httpx, opentelemetry-instrumentation-redis to backend dependencies
• Initialize OTel SDK in main.py startup (exporter pointed at existing collector on :4317)
• Auto-instrument FastAPI app, httpx client, and Redis connections
• Propagate trace context through agent-to-agent calls (traceparent header via httpx instrumentation)
• Add trace ID to structured log output (logging_config.py) for log-trace correlation

Acceptance Criteria

• Every API request gets a trace ID automatically
• Agent A → backend → Agent B calls share the same trace ID
• Trace ID appears in structured JSON logs
• Traces export to OTel Collector (already running on otel-collector:4317)
• No performance regression (sampling at 10% for high-volume endpoints)
• ~30 lines in main.py, 3 new pip packages, zero new infrastructure

Key Files

• src/backend/main.py — OTel initialization
• src/backend/logging_config.py — trace ID in log format
• src/backend/services/agent_client.py — httpx auto-instrumented (no code change needed)
• docker-compose.yml — collector already configured

Dependencies

None — collector infrastructure already exists.