docs(gooddata-eval): drop the kda_skill cross-reference from _extract_metric_result

Confusing to a reviewer with no KDA context; the relevant fact is that
conversation.py now shares this implementation instead of its own copy.

JIRA: QA-29053
risk: nonprod

Author: FrankHuynh

packages/gooddata-eval/src/gooddata_eval/core/agentic/metric_skill.py

@@ -166,8 +166,8 @@ def _extract_metric_result(tool_call_events: list[ToolCallEvent]) -> dict | None
     """Result payload of the create_metric tool call.
 
     Prefers the most recent successful call in this turn -- when the agent retries
-    after a validation error, the earlier failed attempt must not shadow it. Mirrors
-    the "last wins" pairing in ``kda_skill._extract_kda_calls``.
+    after a validation error, the earlier failed attempt must not shadow it. Shared
+    with ``conversation.py``, which imports this instead of keeping its own copy.
     """
     for tc in reversed(tool_call_events):
         if tc.function_name != "create_metric" or not tc.result:

packages/gooddata-eval/tests/test_agentic_metric_skill.py

@@ -51,7 +51,7 @@ def test_extract_metric_result_skips_a_failed_call_after_an_earlier_success():
 
 def test_extract_metric_result_prefers_the_most_recent_successful_call():
     """Two distinct successful create_metric calls in one turn (not a retry after a
-    failure) -- the later one wins, consistent with kda_skill's 'last pair' rule."""
+    failure) -- the later one wins."""
     calls = [
         _create_metric_call('{"data": {"metric_id": "m1"}}'),
         _create_metric_call('{"data": {"metric_id": "m2"}}'),