[043][Phase 9][US7] EvalsTelemetry + runner span emission

[jwesleye]

Scope

US7 — emit OTel spans for the eval runner itself so regressions in eval metrics correlate with deployment/observability events in the same dashboard.

Priority: P2

Tasks

• T135 [P] [US7] Tests in eval/tests/telemetry_test.rs using opentelemetry-sdk::testing::trace::InMemorySpanExporter — span tree (run_set → case → evaluator), error-recording on failed case, parent-span inheritance
• T136 [US7] EvalsTelemetry + EvalsTelemetryBuilder in eval/src/telemetry.rs (feature telemetry)
• T137 [US7] Wire EvalsTelemetry into EvalRunner::run_set — emit three-level span tree with standardized attributes (FR-035); honor existing parent span
• T138 [US7] eval/tests/us7_end_to_end_test.rs — full run produces expected span tree; failing case surfaces as errored span (US7 scenario 3)

Acceptance

• Span names: swink.eval.run_set → swink.eval.case → swink.eval.evaluator.
• Attributes include eval_set.id, case.id, case.name, evaluator.name, prompt.version, score.value, score.threshold, verdict, duration_ms, session_id.
• Evaluator failure records OTel status-error + exception event rather than silently succeeding.
• Spans nest under any already-active parent span.

References

• Spec FR-035, US7 scenarios 1–4
• Research R-005

Depends on

#753 (runner).

[jwesleye]

🤖 Automated agent picking this up. Scope: T135–T138 (EvalsTelemetry + three-level span tree wired into EvalRunner::run_set). Branch will stack on PR #819 (US2 runner). Auto-merge enabled.

[jwesleye]

PR #820 shipped all 4 tasks (T135–T138). Worker correctly detected that PR #819 had already merged to integration and branched off integration directly instead of the stacked base. Design note: with num_runs > 1, only the first run emits per-evaluator spans (keeps span count = cases × evaluators, not × num_runs) — documented inline per FR-035.

[jwesleye]

All tasks from this issue are on integration. Closing as complete.