fix(gooddata-eval): fix create/execute pairing, score names, dataset default

- _extract_kda_calls picked the last create and last execute call
  independently, not as a pair. create_1 -> execute_1(success) ->
  create_2 (never executed) would wrongly report create_2's args as
  executed/succeeded using execute_1's stale result. Fixed by clearing
  execute_result whenever a new create call is seen.
- Score names executed/success/turn_completed now carry the kda_
  prefix (kda_executed/kda_success/kda_turn_completed), matching
  kda_triggered's existing convention -- gdc-nas's combo_report.py
  verdict() must read the same names, updated in the same change.
- dataset_name default corrected from "kda_skill" to "agent_kda_skill",
  matching the actual Langfuse dataset name used everywhere else.

JIRA: QA-28800
risk: nonprod

Author: FrankHuynh

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

@@ -118,17 +118,20 @@ def generate_simulated_kda_response(agent_message: str, measure_candidates: dict
 
 
 def _extract_kda_calls(tool_call_events: list[ToolCallEvent]) -> tuple[dict | None, dict | None]:
-    """Return (create_args, execute_result): the arguments of the LAST
-    `create_key_driver_analysis` call and the parsed result of the LAST
-    `execute_key_driver_analysis` call. Taking the last (not first) attempt matches
-    the observed retry-loop behaviour (kda_1 fails, kda_2 retries) -- the last
-    attempt is what actually determined the answer the chatbot gave.
+    """Return (create_args, execute_result) for the LAST create/execute *pair* -- not the
+    last create and last execute picked independently. Taking the last pair (not the
+    first) matches the observed retry-loop behaviour (kda_1 fails, kda_2 retries): the
+    last attempt is what actually determined the answer the chatbot gave. A new create
+    call clears any earlier execute_result, since that result belongs to the create it
+    followed, not to this one -- without that reset, `create_1 -> execute_1(success) ->
+    create_2 (never executed)` would wrongly pair create_2's args with execute_1's result.
     """
     create_args: dict | None = None
     execute_result: dict | None = None
     for tc in tool_call_events:
         if tc.function_name == "create_key_driver_analysis":
             create_args = tc.parsed_arguments()
+            execute_result = None
         elif tc.function_name == "execute_key_driver_analysis" and tc.result:
             execute_result = tc.parsed_result()
     return create_args, execute_result
@@ -430,7 +433,7 @@ def evaluate_agentic_kda_skill(
     initial_conversation_id: str | None = None,
     langfuse: object | None = None,
     dataset_item_id: str = "",
-    dataset_name: str = "kda_skill",
+    dataset_name: str = "agent_kda_skill",
     run_timestamp: str | None = None,
     model_version_override: str | None = None,
     run_metadata_extra: dict | None = None,
@@ -502,9 +505,9 @@ def _select_and_record(found: list[Any]) -> Any | None:
             # Gates strict_pass -- current scope is completion only (see KdaEvaluation docstring).
             strict_checks = {
                 "kda_triggered": ev.kda_triggered,
-                "executed": ev.executed,
-                "success": ev.success,
-                "turn_completed": ev.turn_completed,
+                "kda_executed": ev.executed,
+                "kda_success": ev.success,
+                "kda_turn_completed": ev.turn_completed,
             }
             # Informational only -- logged for visibility / a future correctness ticket,
             # NOT part of strict_checks/strict_pass. See KdaEvaluation docstring.

packages/gooddata-eval/tests/test_agentic_kda_skill.py

@@ -154,6 +154,24 @@ def test_extract_kda_calls_takes_last_execute_on_retry():
     assert execute_result == {"success": True, "data": {"summary": {}}}
 
 
+def test_extract_kda_calls_does_not_pair_a_new_create_with_an_earlier_execute():
+    # create_1 -> execute_1(success) -> create_2 (never executed): create_2's args must
+    # not get paired with execute_1's stale result -- that would wrongly report the run
+    # as executed/succeeded when the actual last attempt never ran.
+    events = ChatResult.model_validate(
+        {
+            "toolCallEvents": [
+                _tool_call("create_key_driver_analysis", arguments={"measure": {"type": "metric", "id": "a"}}),
+                _tool_call("execute_key_driver_analysis", result={"success": True, "data": {"summary": {}}}),
+                _tool_call("create_key_driver_analysis", arguments={"measure": {"type": "metric", "id": "b"}}),
+            ]
+        }
+    ).tool_call_events
+    create_args, execute_result = _extract_kda_calls(events)
+    assert create_args == {"measure": {"type": "metric", "id": "b"}}
+    assert execute_result is None
+
+
 def test_extract_kda_calls_none_when_no_tool_calls():
     create_args, execute_result = _extract_kda_calls([])
     assert create_args is None