[FEATURE] Add eval scoring support to all exporters

[epinzur]

Problem Statement

When I eval a trace or span in mastra, I want the score to show up on my langfuse/braintrust/etc... traces and spans.

Proposed Solution

Design and build a new system for passing Eval data through the observability event stack. Update the exporters to consume this data, and save scores to traces & spans.

Component

Other

Alternatives Considered

No response

Example Use Case

I want to see my mastra evals in braintrust/langfuse/arize.

Additional Context

We already do this for langfuse, but we are circumventing the observability pipeline. We should refactor and upgrade observability to natively handle evals and scores.

Also consider if the design should include dataset/experimentation support.

Verification

• I have searched the existing issues to make sure this is not a duplicate
• I have provided sufficient context for the team to understand the request

[zijin1994]

@astout55555 and I will work on this!

[zijin1994]

@epinzur We made a initial attempt of this. Now the default exporter + playground would see scores on spans, and they are persistent for the default exporter as well. If this is the right direction, we will start implementing exporters logics (maybe langfuse first and then datadog once that gets merged?)

ps: this is how it looks atm (inside span details):

[epinzur]

@zijin1994 I can't comment on the UI/UX look and feel. Please work with @greglobinski or @smthomas on that.

I'm much more interested in the backend implementation. Could you share a PR with your progress for me to review and comment on?

[epinzur]

@zijin1994 Actually, I made a goof in my last comment.

We already have a UI for Scores on spans. You don't need to create any new UI or experience in studio/playground for scores on spans. The existing system is sufficient.

But we do want to change the backend path to send the scores through the observability system when going out to mastra storage (via DefaultExporter) or to mastra API (via CloudExporter).

We also may need a way to save a score on a span vs on a trace. I think we currently allow both of these scenarios in the system. You can confirm with @YujohnNattrass