fix(tracing): don't record an empty streaming model as llm.model_name

[jeremypng]

Description

On streaming chat completions, convertSSEToJSON took the model from the first SSE chunk only. Azure OpenAI's leading content-filter chunk carries model:"", so the span's llm.model_name was overwritten with an empty string (the request-time value is last-write-lost), breaking cost attribution in OpenInference consumers.

This fixes it by capturing the first non-empty model across all SSE chunks in the stream, so a leading empty-model chunk (or any later chunk reporting a different value) can't clobber it. It also adds the same resp.Model != "" guard to buildResponseAttributes that buildResponsesResponseAttributes already has for the Responses API path, so an empty response model can never overwrite the request-time llm.model_name attribute.

This intentionally differs from the translator's streamingResponseModel (used for metrics), which takes the last non-empty value across the stream. For tracing, taking the first non-empty value is a defensive choice: it means a later, possibly-anomalous chunk (e.g. a routed fallback reporting a different model) can't silently override the model that was already correctly captured earlier in the stream.

Two behavioral guarantees after this change:

• Streamed spans carry the upstream-resolved model as soon as any chunk in the stream reports one, regardless of chunk order.
• An empty model value never overwrites the request-time llm.model_name value, on either the streaming or non-streaming Chat Completions path.

Related Issues/PRs (if applicable)

Fixes #2443

Special notes for reviewers (if applicable)

• Test coverage added (internal/tracing/openinference/openai):
  1. TestBuildResponseAttributesEmptyModelOmitsModelName (response_attrs_test.go) — an empty response model must not emit llm.model_name at all.
  2. sse_converter_test.go, case "model from first non-empty chunk (Azure content-filter prologue)" — model comes from the first chunk that actually carries one, skipping the empty-model prologue chunk.
  3. sse_converter_test.go, case "model retains first non-empty value even if a later chunk differs" — a later chunk reporting a different model does not override the first non-empty value.
• Commit is signed off (DCO) per CONTRIBUTING.md.
• Generative AI disclosure: this fix and its tests were prepared with AI assistance (Claude); I've reviewed and understand the change and take full ownership of it, per CONTRIBUTING.md's generative AI policy.

[nacx]

Thanks!

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.86%. Comparing base (35bde0c) to head (ecaa537).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2444      +/-   ##
==========================================
- Coverage   84.87%   84.86%   -0.02%     
==========================================
  Files         154      154              
  Lines       22419    22423       +4     
==========================================
+ Hits        19028    19029       +1     
- Misses       2236     2238       +2     
- Partials     1155     1156       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.