feat: add P7 OCR soak scenario and scoreboard

tests/unit/test_ocr_client.py

@@ -350,7 +350,7 @@ class TestTelemetry:
         for row in recorder.rows:
             assert "<ocr:text image_bytes=15>" in row["messages"]
             assert "RAW-IMAGE-BYTES" not in row["messages"]  # 图像字节绝不入库
-        assert recorder.rows[0]["error"] is not None
+        assert recorder.rows[0]["error"].startswith("TransientError:")  # 类名前缀口径
         assert recorder.rows[1]["error"] is None
         assert recorder.rows[1]["prompt_tokens"] == 0
 

tests/unit/test_soak_scoreboard.py

@@ -0,0 +1,54 @@
+"""P7 记分板新增纯函数测试(M3 计划 T8): 成功率/坏源占比/错误分类。"""
+
+import pytest
+
+from tools.soak.scoreboard import (
+    inv_errors_classified,
+    inv_fault_share,
+    inv_success_rate,
+)
+
+_KNOWN = ("TransientError", "AllSourcesExhausted", "cancelled")
+
+
+def _row(source="monkey_1", error=None, cache_hit=0):
+    return {"source_name": source, "error": error, "cache_hit": cache_hit}
+
+
+class TestSuccessRate:
+    def test_pass_at_threshold(self):
+        inv_success_rate(98, 100, min_rate=0.98)
+
+    def test_fail_below(self):
+        with pytest.raises(AssertionError, match="成功率"):
+            inv_success_rate(97, 100, min_rate=0.98)
+
+    def test_zero_total_rejected(self):
+        with pytest.raises(AssertionError):
+            inv_success_rate(0, 0, min_rate=0.98)
+
+
+class TestFaultShare:
+    def test_share_within_threshold(self):
+        rows = [_row()] * 9 + [_row("monkey_3")]
+        inv_fault_share(rows, ["monkey_3", "monkey_4"], max_share=0.15)
+
+    def test_share_exceeds(self):
+        rows = [_row()] * 4 + [_row("monkey_3")] * 2
+        with pytest.raises(AssertionError, match="坏源尝试占比"):
+            inv_fault_share(rows, ["monkey_3"], max_share=0.15)
+
+    def test_cache_rows_excluded(self):
+        rows = [_row(cache_hit=1)] * 100 + [_row("monkey_3")] + [_row()] * 9
+        inv_fault_share(rows, ["monkey_3"], max_share=0.15)  # 缓存行不入分母
+
+
+class TestErrorsClassified:
+    def test_known_prefixes_pass(self):
+        rows = [_row(error="TransientError: m1 超时"), _row(error="cancelled"), _row()]
+        inv_errors_classified(rows, _KNOWN)
+
+    def test_unknown_prefix_fails(self):
+        rows = [_row(error="KeyError: 'oops'")]
+        with pytest.raises(AssertionError, match="未分类"):
+            inv_errors_classified(rows, _KNOWN)