fix(gooddata-eval): don't feed fallback-trace latency into KDA value_score

_select_kda_trace falls back to the max-latency candidate only so scores
still attach to a real trace instead of being orphaned when KDA never
triggered -- that fallback trace's latency/cost are not meaningful KDA
numbers (previous comment said so but the caller still passed pt.latency
into log_quality_and_value_scores regardless, contaminating value_score
with an unrelated turn's duration).

Gate on ev.kda_triggered (our own tool-call extraction, not Langfuse) before
trusting pt.latency/pt.total_cost. When kda_triggered is False, pass None so
log_quality_and_value_scores' own worst-case default (speed=0) applies,
which is the correct outcome for a case that never triggered KDA at all.

Verified offline: a never-triggered case with a mocked fallback trace
(latency=999.0, totalCost=5.0) now logs latency_sec=None/cost_usd=None
instead of leaking those numbers into value_score.

JIRA: QA-28800
risk: nonprod

Author: FrankHuynh

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

@@ -156,8 +156,9 @@ def _trace_has_kda_call(langfuse: Any, trace_id: str) -> bool:
 
 
 def _select_kda_trace(langfuse: Any, candidates: list[Any]) -> Any | None:
-    """Pick the candidate trace that made the KDA tool call; fall back to max-latency
-    if none did (KDA never triggered -- nothing meaningful to time anyway)."""
+    """Pick the candidate trace that made the KDA tool call. Falls back to max-latency
+    only so scores still attach to some real trace instead of being orphaned -- callers
+    must not treat that fallback trace's latency as a real KDA duration."""
     for candidate in candidates:
         if _trace_has_kda_call(langfuse, candidate.id):
             return candidate
@@ -433,12 +434,14 @@ def evaluate_agentic_kda_skill(
             with observe(langfuse, pt.id if pt else None, dataset_item_id, run_name, run_metadata) as tid:
                 for score_name, value in {**strict_checks, **informational_checks}.items():
                     score_safe(langfuse, tid, name=score_name, value=float(value), data_type="BOOLEAN")
+                # pt can be a fallback (non-KDA) trace when kda_triggered is False -- its
+                # latency/cost aren't real KDA numbers, so don't feed them into value_score.
                 log_quality_and_value_scores(
                     langfuse,
                     tid,
                     strict_checks=strict_checks,
-                    latency_sec=pt.latency if pt else None,
-                    cost_usd=pt.total_cost if pt else None,
+                    latency_sec=pt.latency if pt and ev.kda_triggered else None,
+                    cost_usd=pt.total_cost if pt and ev.kda_triggered else None,
                 )
 
     if not summary.pass_at_k: