feat(tasks): add skill-templated task graph execution

app-instance/backend/beaver/skills/learning/eval.py

@@ -284,6 +284,9 @@ def _build_replay_case_reports(
         "side_effects": [*baseline.get("side_effects", []), *candidate_arm.get("side_effects", [])],
         "validator_notes": list(surrogate.get("notes") or []),
     }
+    historical_accepted_score = _historical_accepted_score(case)
+    if historical_accepted_score is not None:
+        case_report["historical_accepted_score"] = historical_accepted_score
     return case_report, {
         "run_id": case["run_id"],
         "session_id": case.get("session_id") or "",
@@ -293,6 +296,7 @@ def _build_replay_case_reports(
         "baseline_score": baseline_score,
         "candidate_score": candidate_score,
         "delta": round(candidate_score - baseline_score, 4),
+        **({"historical_accepted_score": historical_accepted_score} if historical_accepted_score is not None else {}),
     }
 
 
@@ -658,8 +662,11 @@ def _ability_score(*, case: dict[str, Any], arm: dict[str, Any], arm_name: str)
     if validator is not None:
         return _ability_from_validator(validator, arm)
     if not case.get("synthetic"):
-        score = _bounded_score(case.get("accepted_score"), default=0.75) if arm_name == "baseline" else _ability_from_output(arm)["final_score"]
-        return _ability_breakdown(score=score, source="user_feedback" if arm_name == "baseline" else "llm_judge")
+        result = _ability_from_output(arm, source="output_heuristic")
+        historical_accepted_score = _historical_accepted_score(case)
+        if historical_accepted_score is not None:
+            result["historical_accepted_score"] = historical_accepted_score
+        return result
     return _ability_breakdown(score=0.0, source="unscored", notes=["Synthetic cases require a validator."])
 
 
@@ -697,6 +704,12 @@ def _ability_from_output(arm: dict[str, Any], *, source: str = "llm_judge", note
     return _ability_breakdown(score=score, source=source, notes=notes)
 
 
+def _historical_accepted_score(case: dict[str, Any]) -> float | None:
+    if case.get("synthetic") or isinstance(case.get("validator"), dict) or "accepted_score" not in case:
+        return None
+    return _bounded_score(case.get("accepted_score"), default=0.75)
+
+
 def _ability_breakdown(*, score: float, source: str, notes: list[str] | None = None) -> dict[str, Any]:
     bounded = _bounded_score(score, default=0.0)
     return {