fix(go/tracing): mark only originating span as failure source

[ktsoator]

Summary

• mark only the span where an error originates with genkit:isFailureSource
• propagate an internal handled marker through nested Genkit spans while keeping parent spans in the error state
• preserve the application error type in OpenTelemetry events and restore the original error identity at the outermost Genkit span

Problem

RunInNewSpan currently sets IsFailureSource on every span that observes a returned error. As the same error propagates through nested actions and flows, each parent span therefore exports genkit:isFailureSource=true, which obscures the actual source and can overcount failures in monitoring.

The tracing package already contains markedError, markErrorAsHandled, and isErrorAlreadyMarked, but the marker was never applied in RunInNewSpan.

Implementation

The originating span now marks and wraps an unhandled error before returning it. Parent spans recognize the marker and omit genkit:isFailureSource, while still recording their normal OpenTelemetry error event and error status.

The internal wrapper is removed when recording the application error in OpenTelemetry and when the error exits the outermost Genkit span. This keeps exception.type, direct sentinel comparisons, and errors.Is/errors.As behavior compatible with the original error.

Testing

Added regression coverage that verifies:

• the child/originating span exports genkit:isFailureSource=true
• the parent span does not export the failure-source attribute
• both child and parent spans remain in the error state
• the original error identity and errors.Is behavior are preserved
• the internal marker is removed without stripping application-added wrappers

Commands run:

go vet ./core/tracing
go test -race ./core/tracing
go test ./...

All passed.

Fixes #5153

Checklist

• PR title follows Conventional Commits
• Tested with unit, race, vet, and full Go test suite
• Docs not applicable: no public API or user-facing behavior changed

[gemini-code-assist]

Code Review

This pull request introduces an unwrapMarkedError helper function to preserve the original error identity in telemetry, updates RunInNewSpan to correctly mark and unwrap errors, and enhances testing around failure sources and error unwrapping. Feedback on these changes includes adding a nil check during type assertion in unwrapMarkedError to prevent potential nil pointer dereferences, and checking if err != nil before unwrapping in RunInNewSpan to avoid unnecessary operations.