test: 补齐规格纠偏安全场景 (#258)

2026-08-17 18:37:39 +08:00
parent be900b2411
commit 29cf66ce96
4 changed files with 138 additions and 11 deletions
@@ -220,6 +220,18 @@ func applyAPISpecContractMutation(mutation string, taskID *string, request map[s
 	case "attempt_id_too_long":
 		request["attempt_id"] = strings.Repeat("a", 192)
 		*key = apiSpecIdempotencyKey(*taskID, request)
+	case "candidate_count_101":
+		candidates := make([]any, 0, 101)
+		for index := 1; index <= 101; index++ {
+			value := "测试尺码" + strconv.Itoa(index)
+			candidates = append(candidates, map[string]any{
+				"candidate_id": "c" + strconv.Itoa(index),
+				"raw_text":     value,
+				"options":      map[string]any{"color": "测试黑", "size": value},
+			})
+		}
+		request["candidates"] = candidates
+		apiSpecRefreshHashes(*taskID, request, key)
 	case "body_too_large":
 		request["future_padding"] = strings.Repeat("x", 64<<10)
 	case "task_not_found":
@@ -46,12 +46,13 @@ type purchaseSpecContractCase struct {
 }

 type purchaseSpecContractModel struct {
-	Conclusion        string   `json:"conclusion"`
-	CandidateID       string   `json:"candidate_id"`
-	ConfidenceBPS     int      `json:"confidence_bps"`
-	Reason            string   `json:"reason"`
-	MissingDimensions []string `json:"missing_dimensions"`
-	Error             string   `json:"error"`
+	Conclusion         string   `json:"conclusion"`
+	CandidateID        string   `json:"candidate_id"`
+	ConfidenceBPS      int      `json:"confidence_bps"`
+	Reason             string   `json:"reason"`
+	ConflictDimensions []string `json:"conflict_dimensions"`
+	MissingDimensions  []string `json:"missing_dimensions"`
+	Error              string   `json:"error"`
 }

 type purchaseSpecContractExpected struct {
@@ -230,11 +231,12 @@ func contractAIClientForCase(testCase purchaseSpecContractCase) *contractAIModel
 		return client
 	}
 	client.response = AIModelMatchResponse{
-		Conclusion:        testCase.Model.Conclusion,
-		CandidateID:       testCase.Model.CandidateID,
-		ConfidenceBPS:     testCase.Model.ConfidenceBPS,
-		Reason:            testCase.Model.Reason,
-		MissingDimensions: append([]string(nil), testCase.Model.MissingDimensions...),
+		Conclusion:         testCase.Model.Conclusion,
+		CandidateID:        testCase.Model.CandidateID,
+		ConfidenceBPS:      testCase.Model.ConfidenceBPS,
+		Reason:             testCase.Model.Reason,
+		ConflictDimensions: append([]string(nil), testCase.Model.ConflictDimensions...),
+		MissingDimensions:  append([]string(nil), testCase.Model.MissingDimensions...),
 	}
 	return client
 }
@@ -375,6 +375,20 @@ class PurchaseSpecResolutionContractVectorTest(unittest.TestCase):
        elif mutation == "attempt_id_too_long":
            request["attempt_id"] = "a" * 192
            key = _idempotency_key(task_id, request)
+        elif mutation == "candidate_count_101":
+            request["candidates"] = [
+                {
+                    "candidate_id": f"c{index}",
+                    "raw_text": f"测试尺码{index}",
+                    "options": {
+                        "color": "测试黑",
+                        "size": f"测试尺码{index}",
+                    },
+                }
+                for index in range(1, 102)
+            ]
+            request["candidate_snapshot_hash"] = _snapshot_hash(request)
+            key = _idempotency_key(task_id, request)
        elif mutation == "body_too_large":
            request["future_padding"] = "x" * (64 * 1024)
        elif mutation == "task_not_found":
@@ -30,6 +30,40 @@
        "model_calls": 0
      }
    },
+    {
+      "name": "rule_traditional_simplified",
+      "task_id": "cg-contract-traditional",
+      "target_size": "藍 色 / XL 碼",
+      "candidates": ["蓝色／xl码"],
+      "candidate_snapshot_hash": "5433787fa5bc7cae41939504a221bb59cee26639e683f89506e524b0bf318202",
+      "idempotency_key": "spec-resolution-v1:f1c239574dcbf5c8bbb12a71f21dadf874d5d0b544e4fa18b957f896295801e4",
+      "mode": "rule",
+      "model": null,
+      "expected": {
+        "outcome": "matched",
+        "source": "rule",
+        "candidate_id": "c1",
+        "confidence_bps": 10000,
+        "model_calls": 0
+      }
+    },
+    {
+      "name": "rule_weight_range_unique",
+      "task_id": "cg-contract-range",
+      "target_size": "45-55公斤",
+      "candidates": ["70-90斤", "90-110斤", "110-130斤"],
+      "candidate_snapshot_hash": "1b81a16f0c1efcd457311e0f229859ad7df03453f9f9ee5946a9e78f353e95b6",
+      "idempotency_key": "spec-resolution-v1:e70666cb6141d09059c4bf9de0e755185fc1b5bd72e93c3b399e1c46507f63b6",
+      "mode": "rule",
+      "model": null,
+      "expected": {
+        "outcome": "matched",
+        "source": "rule",
+        "candidate_id": "c2",
+        "confidence_bps": 10000,
+        "model_calls": 0
+      }
+    },
    {
      "name": "rule_overlap_uncertain",
      "task_id": "cg-contract-overlap",
@@ -136,6 +170,70 @@
        "model_calls": 1
      }
    },
+    {
+      "name": "ai_uncertain",
+      "task_id": "cg-contract-ai-uncertain",
+      "target_size": "L码",
+      "candidates": ["M码", "XL码"],
+      "candidate_snapshot_hash": "f634515c8a7a0ebd688fd926520405051e12c622c87a4783a05936099ecfb2fa",
+      "idempotency_key": "spec-resolution-v1:7b531267d5ba16ed9d81a7d8d7e9e8e4553776bd35d3401e2b260609c6b022fb",
+      "mode": "ai",
+      "model": {
+        "conclusion": "uncertain",
+        "confidence_bps": 8500,
+        "reason": "测试信息不足"
+      },
+      "expected": {
+        "outcome": "uncertain",
+        "source": "ai",
+        "candidate_id": null,
+        "confidence_bps": 8500,
+        "model_calls": 1
+      }
+    },
+    {
+      "name": "ai_conflict",
+      "task_id": "cg-contract-ai-conflict",
+      "target_size": "L码",
+      "candidates": ["M码", "XL码"],
+      "candidate_snapshot_hash": "f634515c8a7a0ebd688fd926520405051e12c622c87a4783a05936099ecfb2fa",
+      "idempotency_key": "spec-resolution-v1:4d6581611767a1a56a305f1b769bb651be80d0113af1827c2fa8c66e72ea14b0",
+      "mode": "ai",
+      "model": {
+        "conclusion": "conflict",
+        "confidence_bps": 9500,
+        "reason": "测试规格维度冲突",
+        "conflict_dimensions": ["尺码"]
+      },
+      "expected": {
+        "outcome": "rejected",
+        "source": "ai",
+        "candidate_id": null,
+        "confidence_bps": 9500,
+        "model_calls": 1
+      }
+    },
+    {
+      "name": "ai_invalid_format",
+      "task_id": "cg-contract-ai-invalid",
+      "target_size": "L码",
+      "candidates": ["M码", "XL码"],
+      "candidate_snapshot_hash": "f634515c8a7a0ebd688fd926520405051e12c622c87a4783a05936099ecfb2fa",
+      "idempotency_key": "spec-resolution-v1:5a9dcc28c6dc6f8d877a93932533ca5f64e6e5d10153acf7ddb63cb6e06428de",
+      "mode": "ai",
+      "model": {
+        "conclusion": "other",
+        "confidence_bps": 9200,
+        "reason": "测试非法结论"
+      },
+      "expected": {
+        "outcome": "failed",
+        "source": "ai",
+        "candidate_id": null,
+        "confidence_bps": null,
+        "model_calls": 1
+      }
+    },
    {
      "name": "ai_timeout",
      "task_id": "cg-contract-ai-timeout",
@@ -162,6 +260,7 @@
    {"name": "target_size_not_claimed", "base_case": "rule_weight_unique", "mutation": "target_size_not_claimed", "expected_http_status": 422, "expected_error_code": "INVALID_SPEC_RESOLUTION_REQUEST"},
    {"name": "observed_at_without_timezone", "base_case": "rule_weight_unique", "mutation": "observed_at_without_timezone", "expected_http_status": 422, "expected_error_code": "INVALID_SPEC_RESOLUTION_REQUEST"},
    {"name": "attempt_id_too_long", "base_case": "rule_weight_unique", "mutation": "attempt_id_too_long", "expected_http_status": 422, "expected_error_code": "INVALID_SPEC_RESOLUTION_REQUEST"},
+    {"name": "candidate_count_101", "base_case": "rule_weight_unique", "mutation": "candidate_count_101", "expected_http_status": 422, "expected_error_code": "INVALID_SPEC_RESOLUTION_REQUEST"},
    {"name": "body_too_large", "base_case": "rule_weight_unique", "mutation": "body_too_large", "expected_http_status": 400, "expected_error_code": "INVALID_BODY"},
    {"name": "task_not_found", "base_case": "rule_weight_unique", "mutation": "task_not_found", "expected_http_status": 404, "expected_error_code": "TASK_NOT_FOUND"},
    {"name": "task_not_purchase", "base_case": "rule_weight_unique", "mutation": "task_not_purchase", "expected_http_status": 422, "expected_error_code": "TASK_NOT_PURCHASE"},