feat(langfuse): propagate observation metadata for evals (#14634)

# Pull Request Template

## Description

We need to pass on trace level attributes down to the spans inside them
like tool calls, observations, etc.
This way, we can filter observations based on trace level attributes.


## Type of change

- [x] Bug fix (non-breaking change which fixes an issue)

## How Has This Been Tested?

Please describe the tests that you ran to verify your changes. Provide
instructions so we can reproduce. Please also list any relevant details
for your test configuration.

Attributes added to observation metadata for easy filtering
<img width="1327" height="708" alt="image"
src="https://github.com/user-attachments/assets/8f1d1bf8-cde4-481d-a2c2-7920ad2fc52e"
/>

added a `generation_stage` to differentiate llm_calls that call tools vs
those that generate a `final_response`
<img width="1806" height="968" alt="CleanShot 2026-06-03 at 15 11 09@2x"
src="https://github.com/user-attachments/assets/db1fa8e0-7f2d-404b-a719-27a16d400442"
/>


propagated attributes to tool calls for future use
<img width="903" height="517" alt="image"
src="https://github.com/user-attachments/assets/edc61ce8-93db-465c-a66e-043138e2dc15"
/>



## Checklist:

- [x] My code follows the style guidelines of this project
- [x] I have performed a self-review of my code
- [x] I have commented on my code, particularly in hard-to-understand
areas
- [ ] I have made corresponding changes to the documentation
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective or that my
feature works
- [x] New and existing unit tests pass locally with my changes
- [x] Any dependent changes have been merged and published in downstream
modules

enterprise/app/helpers/captain/chat_generation_recorder.rb

@@ -10,6 +10,7 @@ module Captain::ChatGenerationRecorder
     # Create a generation span with model and token info for Langfuse cost calculation.
     # Note: span duration will be near-zero since we create and end it immediately, but token counts are what Langfuse uses for cost calculation.
     tracer.in_span("llm.captain.#{feature_name}.generation") do |span|
+      apply_current_langfuse_attributes(span)
       set_generation_span_attributes(span, chat, message)
     end
   rescue StandardError => e
@@ -37,11 +38,23 @@ module Captain::ChatGenerationRecorder
       ATTR_GEN_AI_USAGE_INPUT_TOKENS => message.input_tokens,
       ATTR_GEN_AI_USAGE_OUTPUT_TOKENS => message.respond_to?(:output_tokens) ? message.output_tokens : nil,
       ATTR_LANGFUSE_OBSERVATION_INPUT => format_input_messages(chat),
-      ATTR_LANGFUSE_OBSERVATION_OUTPUT => message.respond_to?(:content) ? message.content.to_s : nil
+      ATTR_LANGFUSE_OBSERVATION_OUTPUT => message.respond_to?(:content) ? message.content.to_s : nil,
+      format(ATTR_LANGFUSE_OBSERVATION_METADATA, 'generation_stage') => generation_stage(message)
     }
   end
 
   def format_input_messages(chat)
     chat.messages[0...-1].map { |m| { role: m.role.to_s, content: m.content.to_s } }.to_json
   end
+
+  def generation_stage(message)
+    message_has_tool_calls?(message) ? 'tool_call' : 'final_response'
+  end
+
+  def message_has_tool_calls?(message)
+    return false unless message.respond_to?(:tool_calls)
+
+    tool_calls = message.tool_calls
+    tool_calls.respond_to?(:any?) && tool_calls.any?
+  end
 end