feat(check): 偏离栏直读、响应错位判断；复查复用和硬预算（C2）

- deviation.py：模型判已响应的条目，证据所在表格行的偏离格写着「偏离 / 部分满足」等
  就改判负偏离；照抄要求格、响应格写的是别的内容的改判「响应可能对不上」
- reuse.py：按「提示词 + 要求 + 候选句原文」算缓存键，存模型原始结论；复查时内容没变、
  证据找得回的条目直接沿用，不调模型
- 每次检查送给模型的字数有上限，超出的条目列为请人工核对，写进覆盖范围
- 开发集 17 份：召回 61% → 73%，负偏离未说明 64% → 90%，规矩投标文件问题数不变

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

server/tests/test_check_flow.py

@@ -339,3 +339,24 @@ async def test_manual_checks_are_kept_per_project(user: User) -> None:
     assert any("列入「需人工核对」2 条" in x for x in again["coverage"]["not_checked"])
     kept = (await user.get(f"/projects/{project_id}/manual-checks")).json()["items"]
     assert [i["checked"] for i in kept] == [True, False]
+
+
+async def test_recheck_reuses_unchanged_model_results(user: User) -> None:
+    """复查复用（C2）：同一份投标文件再查，送给模型的内容没变，模型那一层沿用上次的结论，不再调模型。"""
+    from bid_doctor.ai.models import AiCall
+    from bid_doctor.core.db import TenantContext, tenant_tx
+
+    project_id, bid, first = await checked_project(user)
+    second = await run_check(user, project_id, [bid["id"]])
+    assert second["counts"]["fatal"] == first["counts"]["fatal"]
+    reused = [x for x in second["coverage"]["checked"] if "沿用上次的结论" in x]
+    assert reused, second["coverage"]["checked"]
+
+    ctx = TenantContext(tenant_id=UUID(user.tenant_id), user_id=UUID(str(user.me["user"]["id"])))
+    async with tenant_tx(ctx) as s:
+        from sqlalchemy import func, select
+
+        n = (await s.execute(select(func.count()).where(AiCall.job_id == UUID(second["id"])))).scalar_one()
+    assert n == 0
+    results = (await user.get(f"/check-runs/{second['id']}/item-results")).json()["items"]
+    assert all(r["bid_citations"] for r in results if r["verdict"] == "responded")  # 沿用的结论证据也找回来了