Принять agent replay маржинальности номенклатуры

scripts/domain_truth_harness.py

@@ -386,6 +386,85 @@ def normalize_optional_bool(value: Any) -> bool | None:
     return None
 
 
+LEGACY_DEBUG_RESULT_MODES = {"confirmed_balance", "heuristic_candidates"}
+BUSINESS_EXPECTED_RESULT_MODES = {
+    "clarification_required",
+    "limited_accounting_answer",
+    "evidence_or_honest_boundary",
+    "ranking_or_limited_accounting_answer",
+    "same_inventory_margin_context_or_clarification",
+}
+
+
+def _business_review_is_clean(step_state: dict[str, Any]) -> bool:
+    business_review = step_state.get("business_first_review")
+    if not isinstance(business_review, dict):
+        return True
+    return len(dcl.normalize_string_list(business_review.get("issue_codes"))) == 0
+
+
+def business_expected_result_mode_matches(expected_result_mode: str, step_state: dict[str, Any]) -> bool:
+    reply_type = str(step_state.get("reply_type") or "").strip()
+    response_type = str(step_state.get("response_type") or "").strip()
+    truth_mode = str(step_state.get("truth_mode") or "").strip()
+    answer_shape = str(step_state.get("answer_shape") or "").strip()
+    detected_intent = str(step_state.get("detected_intent") or "").strip()
+    capability_id = str(step_state.get("capability_id") or "").strip()
+    assistant_text = str(step_state.get("assistant_text") or "").strip()
+    clean_business_review = _business_review_is_clean(step_state)
+    in_margin_context = (
+        detected_intent == "inventory_margin_ranking_for_nomenclature"
+        or capability_id == "inventory_inventory_margin_ranking_for_nomenclature"
+    )
+
+    if expected_result_mode == "clarification_required":
+        return (
+            clean_business_review
+            and (
+                truth_mode == "clarification_required"
+                or answer_shape == "clarification_required"
+                or (reply_type == "partial_coverage" and response_type == "LIMITED_WITH_REASON")
+            )
+        )
+
+    if expected_result_mode == "limited_accounting_answer":
+        return (
+            clean_business_review
+            and in_margin_context
+            and bool(assistant_text)
+            and reply_type in {"partial_coverage", "factual", "factual_with_explanation"}
+        )
+
+    if expected_result_mode == "evidence_or_honest_boundary":
+        return (
+            clean_business_review
+            and bool(assistant_text)
+            and reply_type in {"partial_coverage", "factual", "factual_with_explanation"}
+        )
+
+    if expected_result_mode == "ranking_or_limited_accounting_answer":
+        return (
+            clean_business_review
+            and in_margin_context
+            and bool(assistant_text)
+            and reply_type in {"partial_coverage", "factual", "factual_with_explanation"}
+        )
+
+    if expected_result_mode == "same_inventory_margin_context_or_clarification":
+        return (
+            clean_business_review
+            and bool(assistant_text)
+            and (
+                in_margin_context
+                or truth_mode == "clarification_required"
+                or answer_shape == "clarification_required"
+            )
+            and reply_type in {"partial_coverage", "factual", "factual_with_explanation"}
+        )
+
+    return False
+
+
 def evaluate_truth_step(
     *,
     step: dict[str, Any],
@@ -627,7 +706,24 @@ def evaluate_truth_step(
         resolve_nested_placeholders(step.get("expected_result_mode"), step_results, bindings, runtime_bindings) or ""
     ).strip()
     actual_result_mode = str(step_state.get("result_mode") or "").strip()
-    if expected_result_mode and actual_result_mode and not dcl.identifiers_match(actual_result_mode, expected_result_mode):
+    if expected_result_mode in BUSINESS_EXPECTED_RESULT_MODES:
+        if not business_expected_result_mode_matches(expected_result_mode, step_state):
+            append_finding(
+                findings,
+                step,
+                "wrong_result_mode",
+                "Business answer mode does not match the expected semantic answer contract.",
+                actual={
+                    "result_mode": actual_result_mode or None,
+                    "reply_type": reply_type or None,
+                    "truth_mode": step_state.get("truth_mode"),
+                    "answer_shape": step_state.get("answer_shape"),
+                    "intent": detected_intent or None,
+                    "capability": capability_id or None,
+                },
+                expected=expected_result_mode,
+            )
+    elif expected_result_mode and actual_result_mode and not dcl.identifiers_match(actual_result_mode, expected_result_mode):
         append_finding(
             findings,
             step,