Issue 003: OpenTelemetry Distributed Tracing

Summary

Adopt OpenTelemetry tracing end-to-end. Wire micrometer-tracing-bridge-otel
into the Spring backend, propagate trace context to the Python AI
sidecar, deploy a local OpenTelemetry Collector + Jaeger via
docker-compose, and add a regression IT that captures the expected
span topology of a typical analyze request.

Motivation

• P95 latency investigations today rely on logs; we have no causal
  chain showing which dependency is the bottleneck.
• Micrometer counters (commit 6447917) answer "did it time out?"
  but not "where in the pipeline?".
• springdoc-openapi already shows that instrumentation hooks
  exist; the tracing side is the missing half.

Tasks

• Add micrometer-tracing-bridge-otel and
  opentelemetry-exporter-otlp to backend/pom.xml.
• Configure management.tracing.sampling.probability=0.1,
  management.tracing.propagation.type=w3c, and
  management.otlp.tracing.endpoint (env-overridable).
• Annotate every *Controller method with @Observed and the
  WorkflowOrchestrator stage methods.
• Add @NewSpan("analysis.computeContextHash") to the hash
  helper and @ContinueSpan to the sidecar call.
• Add OtelMiddleware to ai-service/app/observability.py;
  register in ai-service/app/main.py; pass traceparent through
  in the Python call chain.
• Add otel-collector and jaeger services to
  docker-compose.yml and scripts/otel-collector-config.yaml.
• Add the InMemorySpanExporter-based regression IT in
  src/test/java/com/finsight/it/observability/.
• Emit traceparent on 4xx/5xx responses (decision from
  the RFC).
• docs/operations.md: how to read traces in Jaeger, how to
  copy a trace ID from a 500 response.

Acceptance criteria

• A full analyze request produces 6+ spans, including
  GET /api/research/stock/{symbol}/analyze,
  analysis.cacheLookup, analysis.leaseWait,
  ai.sidecar.analyze-stock, document.keywordSearch,
  document.vectorSearch.
• Traces from Java and the Python sidecar share the same
  traceparent.
• docker compose up exposes Jaeger on http://localhost:16686
  and shows traces for any analyze request issued against the
  compose stack.
• The regression IT fails if any of the expected spans go missing
  in a refactor.

Out of scope

• Tail-based sampling.
• Replacing Micrometer metrics.
• Continuous profiling.

References

• docs/rfcs/RFC-003-opentelemetry-distributed-tracing.md
• Spring Boot 3.3 tracing reference
  (https://docs.spring.io/spring-boot/docs/3.3.x/reference/html/actuator.html#actuator.micrometer-tracing)
• OpenTelemetry semantic conventions
  (https://opentelemetry.io/docs/specs/semconv/)

Estimate

3 weeks. Split into 4 PRs:

1. Backend dependencies + auto-instrumentation (~400 LoC, 1 PR)
2. Custom spans + W3C propagation to sidecar (~300 LoC, 1 PR)
3. Collector + Jaeger compose profile (~120 LoC, 1 PR)
4. Regression IT + docs (~250 LoC, 1 PR)