[Bug]: Assistant_code trajectory step logs raw LLM content instead of extracted code

[sami-marreed]

What happened

When code is extracted from an assistant response, Assistant_code is written with the full raw LLM content instead of the extracted code. That makes Assistant_code a duplicate of Raw_Assistant_Response.

Fix requirement: keep both steps — do not collapse to one.

• Raw_Assistant_Response → full raw LLM content (unchanged purpose)
• Assistant_code → extracted code only (when code was extracted)
• Assistant_nl → raw content when there is no code (existing branch)

Validated on origin/main @ 2538f1c2.

How to reproduce

1. Run a CugaLite turn where the model returns prose + a python fence.
2. Inspect trajectory steps.

Example model output:

Sure, here is the code:
```python
print(await find_tools("orders"))

Code path:

```python
self._tracker.collect_step(Step(name="Raw_Assistant_Response", data=content))
if code:
    self._tracker.collect_step(Step(name="Assistant_code", data=content))  # same content!

Expected for a code turn:

{"name": "Raw_Assistant_Response", "data": "Sure, here is the code:\n```python\nprint(await find_tools(\"orders\"))\n```"},
{"name": "Assistant_code", "data": "print(await find_tools(\"orders\"))"}

Actual: both steps have the same raw content string.

Non-goal: removing Raw_Assistant_Response or only logging extracted code.

Environment

• OS: n/a (code-path bug)
• CUGA: origin/main @ 2538f1c2
• Component: CugaLite graph adapter / activity tracker

Logs, screenshots, or config

src/cuga/backend/cuga_graph/nodes/cuga_lite/adapter/graph_adapter.py — on_response_processed

Checklist

• I searched existing issues and this doesn’t look like a duplicate