fix(agents): keep Langfuse completion content as assistant text

[Ordinary (OrdinarySF)]

What

LangfuseSpanAdapter no longer writes MessagePart.Reasoning into gen_ai.completion.{i}.content (or the matching prompt attribute) when a visible assistant reply exists.

• gen_ai.completion.{i}.content / gen_ai.prompt.{i}.content stay tool calls or MessagePart.Text
• reasoning is emitted on gen_ai.completion.{i}.reasoning / gen_ai.prompt.{i}.reasoning
• finish_reason is kept on text completions even when reasoning parts are present

gen_ai.output.messages is unchanged and still contains both part types.

Why

Langfuse Preview maps gen_ai.completion* to observation output. The adapter preferred reasoning over text:

when {
    toolCalls.isNotEmpty() -> /* tool JSON */
    reasoningParts.isNotEmpty() -> /* thinking, as content */
    else -> /* actual text */
}

Reasoning models (Grok, o-series, DeepSeek, …) return thinking as MessagePart.Reasoning and the short final answer as MessagePart.Text. Preview then showed the thinking (often an English paraphrase of a non-English prompt) while gen_ai.usage.output_tokens still counted only the visible completion — for example 3 tokens for RELEVANT next to a long translation.

The agent itself is fine: Message.textContent() already reads only MessagePart.Text. This is an export mapping bug, not a model or Langfuse UI bug.

How it was verified

• Regression tests in LangfuseSpanAdapterTest:
  • testCompletionAttributesPreferTextOverReasoning
  • testCompletionAttributesReasoningOnlyLeavesContentEmpty
  • testPromptAttributesPreferTextOverReasoning
• ./gradlew :agents:agents-features:agents-features-opentelemetry:jvmTest passes locally.

Notes

• Scope is the Langfuse adapter only.
• WeaveSpanAdapter still prefers reasoning over text in the same when. Happy to follow up if you want the same split there.

closes #2216

[Ordinary (OrdinarySF)]

Rebased onto develop after #2248.

#2248 filters out a reasoning part whose content is empty (Gemini signature-only parts). That filter stays, on both prompt and completion attributes.

This PR covers the remaining case: a reasoning part that has text. Langfuse Preview reads .content, so the visible reply stays on gen_ai.completion.{i}.content / gen_ai.prompt.{i}.content. The thinking moves to gen_ai.*.reasoning. finish_reason still comes from the text (or tool-call) branch.

LangfuseSpanAdapterTest keeps the #2248 empty-part test and adds the non-empty cases. Local :agents:agents-features:agents-features-opentelemetry:jvmTest --tests ai.koog.agents.features.opentelemetry.integration.langfuse.LangfuseSpanAdapterTest passed.

Oleksandr Shylenko (@shilenkoalexander) #2248 landed in this same when. This branch is the follow-up for reasoning parts that actually contain text.