"""真实S02引文纠正、失败回执和恢复次数；合成HTTP不认证文学质量。"""

import json
from decimal import Decimal

import pytest
import test_评测执行与失败收敛 as 运行测试

from muse.效果评测.接口 import 评测错误

pytestmark = pytest.mark.数据库
执行环境 = 运行测试.执行环境


def _准备评委(env):
    运行测试._启动执行(env)
    运行测试._运行就绪(env)
    env["app"].要求评测().推进实验(env["actor"], env["exp"]["experiment_id"])


def _工作面(env):
    return env["app"].要求评测().读取执行工作面(env["actor"], env["exp"]["experiment_id"])


def _报告(env):
    return env["app"].要求评测().读取实验报告(env["actor"], env["exp"]["experiment_id"])


@pytest.mark.case_id(
    "TC-70f0ffe8cb5f",
    environment="隔离PG与合成HTTP；不认证真实外部模型效果",
    given="隔离PG中固定单元、两侧真实生成交付及受控合成评委HTTP",
    when="注入原文中不存在的引文并按原固定上限纠正",
    then=[
        "实际两次评委回合；首轮无纠正信息，次轮原字材料不变且带本评委原始草稿、固定错误和原字提示。旧提示ID改为left/right，实际回执和费用另查S02。"
    ],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_坏引文只纠正一次且保留原稿错误及原字提示__70f0ff(执行环境):
    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "ok"])
    运行测试._运行就绪(env)
    requests = [json.loads(r["input"]) for r in env["received"][2:]]
    assert len(requests) == 2 and "correction" not in requests[0]
    correction = requests[1].pop("correction")
    assert requests[1] == requests[0]
    assert (
        correction["previousDraft"]["dimensions"][0]["evidence"][0]["quote"]
        == "候选正文中不存在的引文"
    )
    assert "引文未在对应候选中出现" in correction["error"]
    assert correction["validQuoteHints"] == {
        side: [requests[0][side]["text"]] for side in ("left", "right")
    }
    judge = next(u for u in _工作面(env)["units"] if u["kind"] == "comparison")
    assert judge["state"] == "completed" and len(judge["rejections"]) == 1
    assert len(judge["cost"]["calls"]) == 2
    assert _报告(env)["coverage"]["compared"] == 1
    assert Decimal(_报告(env)["cost"]["total_usd"]) == Decimal(".5")


@pytest.mark.case_id(
    "TC-883b95d3365f",
    environment="隔离PG与合成HTTP；不认证真实外部模型效果",
    given="隔离PG中固定单元、两侧真实生成交付及受控合成评委HTTP",
    when="注入原文中不存在的引文并按原固定上限纠正",
    then=[
        "一次纠正耗尽后没有第三次调用，状态失败和完整样本分母保留；PAIRWISE_QUOTE_NOT_FOUND承接旧错误码，摘要保留两侧/维度计数与实际回执，不回显错误引文。"
    ],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_一次纠正耗尽后保留失败诊断且不发第三次__883b95(执行环境):
    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "bad_quote", "ok"])
    运行测试._运行就绪(env, allow_failure=True)
    work = _工作面(env)
    judge = next(u for u in work["units"] if u["kind"] == "comparison")
    assert judge["state"] == "failed" and judge["output"] is None
    assert len(env["received"]) == 4 and env["scripted"] == ["ok"]
    assert len(judge["rejections"]) == 2
    assert all(r["failure_code"] == "PAIRWISE_QUOTE_NOT_FOUND" for r in judge["rejections"])
    assert "候选正文中不存在的引文" not in json.dumps(judge["rejections"], ensure_ascii=False)
    report = _报告(env)
    assert report["outcomes"]["incomplete"] == 1
    assert report["coverage"] == {"samples": 1, "generated": 1, "compared": 0}
    assert Decimal(report["cost"]["total_usd"]) == Decimal(".5")


@pytest.mark.case_id(
    "TC-5edc5e65298b",
    environment="隔离PG与合成HTTP；不认证真实外部模型效果",
    given="隔离PG中固定单元、两侧真实生成交付及受控合成评委HTTP",
    when="注入原文中不存在的引文并按原固定上限纠正",
    then=[
        "两名反向独立评委各有预注册纠正上限；首位两次纠正后完成，次位首轮不含前一位草稿，总计四次评委调用。"
    ],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
@pytest.mark.parametrize("执行环境", [{"judges": 2, "corrections": 2, "calls": 12}], indirect=True)
def test_每名评委独立最多两次纠正且不共享前一报告__5edc5e(执行环境):
    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "bad_quote", "ok", "ok"])
    运行测试._运行就绪(env)
    requests = [json.loads(r["input"]) for r in env["received"][2:]]
    assert len(requests) == 4
    assert ["correction" in r for r in requests] == [False, True, True, False]
    assert requests[0]["left"] == requests[3]["right"]
    judges = [u for u in _工作面(env)["units"] if u["kind"] == "comparison"]
    assert len(judges) == 2 and all(u["state"] == "completed" for u in judges)
    assert sorted(len(u["rejections"]) for u in judges) == [0, 2]
    assert _报告(env)["outcomes"]["tie"] == 1


@pytest.mark.case_id(
    "NC-w25-25b001",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["零纠正或原调用配额不足不多发一次"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
@pytest.mark.parametrize(
    "执行环境", [{"corrections": 0}, {"corrections": 2, "calls": 3}], indirect=True
)
def test_零纠正或调用配额不足不新增一次调用__25b001(执行环境):
    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "ok"])
    运行测试._运行就绪(env, allow_failure=True)
    judge = next(u for u in _工作面(env)["units"] if u["kind"] == "comparison")
    assert len(judge["rejections"]) == 1 and judge["state"] == "failed"
    assert len(env["received"]) == 3 and env["scripted"] == ["ok"]


@pytest.mark.case_id(
    "NC-w25-25b002",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["拒绝提交后中断恢复只用余下次数且保原尝试"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_拒绝提交后中断恢复只继续余下纠正且保留原尝试__25b002(执行环境, monkeypatch):
    from muse.共享.错误 import Muse错误

    env = 执行环境
    svc = env["app"].要求评测()
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "ok"])
    save = type(svc).保存执行交付

    def 中断(self, *args, **kwargs):
        result = save(self, *args, **kwargs)
        if result.get("outcome") == "comparison_rejected":
            raise Muse错误("合成拒绝提交后中断")
        return result

    monkeypatch.setattr(type(svc), "保存执行交付", 中断)
    运行测试._运行就绪(env, allow_failure=True)
    before = next(u for u in _工作面(env)["units"] if u["kind"] == "comparison")
    assert len(env["received"]) == 3 and before["state"] == "failed"
    monkeypatch.setattr(type(svc), "保存执行交付", save)
    运行测试._恢复失败单元(env)
    运行测试._运行就绪(env)
    after = next(u for u in _工作面(env)["units"] if u["kind"] == "comparison")
    assert after["state"] == "completed" and after["rejections"] == before["rejections"]
    assert after["evidence"]["runtime"]["attempt_id"] != before["rejections"][0]["attempt_id"]
    assert len(env["received"]) == 4 and "correction" in json.loads(env["received"][-1]["input"])
    assert Decimal(_报告(env)["cost"]["total_usd"]) == Decimal(".5")


@pytest.mark.case_id(
    "NC-w25-25b003",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["耗尽后恢复被拒，UPDATE/DELETE不能重置次数"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_纠正耗尽不能经恢复或删除拒绝重置次数__25b003(执行环境):
    import psycopg

    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "bad_quote", "ok"])
    运行测试._运行就绪(env, allow_failure=True)
    before = _工作面(env)
    with pytest.raises(评测错误, match="上限"):
        运行测试._恢复失败单元(env)
    for statement in [
        "DELETE FROM evaluation.muse_comparison_rejection",
        "UPDATE evaluation.muse_comparison_rejection SET sequence=1",
    ]:
        with env["pool"].连接() as conn, pytest.raises(psycopg.Error), conn.transaction():
            conn.execute(statement)
    assert _工作面(env) == before
    assert len(env["received"]) == 4 and env["scripted"] == ["ok"]


@pytest.mark.case_id(
    "NC-w25-25b004",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["重签拒绝内容或前序顺序仍过不了S02原回合核验"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
@pytest.mark.parametrize("changed", ["raw_output", "previous_hash", "sequence"])
def test_拒绝记录篡改重签仍不能冒原回合或重置链__25b004(执行环境, monkeypatch, changed):
    from copy import deepcopy

    from muse.共享.错误 import Muse错误
    from muse.效果评测.存储 import 评测存储
    from muse.正式变更.接口 import 固定哈希

    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "ok"])
    运行测试._运行就绪(env)
    read = 评测存储.读取引文拒绝

    def 篡改(self, unit_id):
        rows = deepcopy(read(self, unit_id))
        if rows:
            row = rows[0]
            if changed == "raw_output":
                row["output"]["rationale"] = "不是原调用返回的文字"
            elif changed == "previous_hash":
                row["previous_hash"] = "f" * 64
            else:
                row["sequence"] = 2
            row["record_hash"] = 固定哈希({k: v for k, v in row.items() if k != "record_hash"})
        return rows

    monkeypatch.setattr(评测存储, "读取引文拒绝", 篡改)
    with pytest.raises(Muse错误):
        _报告(env)
    assert len(env["received"]) == 4


@pytest.mark.case_id(
    "NC-w25-25b005",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["拒绝后停止不发纠正，保留拒绝和全部费用"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_拒绝后停止实验不派发纠正且保留拒绝费用__25b005(执行环境, monkeypatch):
    env = 执行环境
    svc = env["app"].要求评测()
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "ok"])
    save = type(svc).保存执行交付

    def 停止(self, *args, **kwargs):
        result = save(self, *args, **kwargs)
        if result.get("outcome") == "comparison_rejected":
            svc.取消实验(env["actor"], env["exp"]["experiment_id"], "stop-after-rejection")
        return result

    monkeypatch.setattr(type(svc), "保存执行交付", 停止)
    运行测试._运行就绪(env, allow_failure=True)
    assert len(env["received"]) == 3 and env["scripted"] == ["ok"]
    work = _工作面(env)
    assert work["state"] == "stopped"
    judge = next(u for u in work["units"] if u["kind"] == "comparison")
    assert len(judge["rejections"]) == 1 and judge["output"] is None
    assert Decimal(_报告(env)["cost"]["total_usd"]) == Decimal(".375")


@pytest.mark.case_id(
    "NC-w25-25b006",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["协议、非引文语义错误及未知成本不自动纠正"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
@pytest.mark.parametrize("script", ["bad_output", "bad_mixed", "unknown_cost"])
def test_协议其他语义错误及未知费用不自动纠正__25b006(执行环境, script):
    env = 执行环境
    _准备评委(env)
    env["scripted"].extend([script, "ok"])
    运行测试._运行就绪(env, allow_failure=True)
    assert len(env["received"]) == 3 and env["scripted"] == ["ok"]
    judge = next(u for u in _工作面(env)["units"] if u["kind"] == "comparison")
    assert judge["output"] is None
    if script == "bad_mixed":
        assert [r["failure_code"] for r in judge["rejections"]] == ["PAIRWISE_NOT_EXECUTED"]
        assert judge["unregistered_deliveries"] == []
    else:
        assert judge["rejections"] == []
    if script == "unknown_cost":
        assert _报告(env)["cost"]["total_usd"] is None
        assert _工作面(env)["state"] == "reconciling"
    else:
        assert judge["state"] == "failed"


@pytest.mark.case_id(
    "NC-w25-25b007",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["纠正前稿偏离原回合在发送前拒绝"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_纠正的上次草稿被改写在发送前拒绝__25b007(执行环境, monkeypatch):
    from dataclasses import replace

    import muse.编排.执行评测 as flow

    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "ok"])
    create = flow.模型请求

    def 改写(*args, **kwargs):
        request = create(*args, **kwargs)
        content = json.loads(request.用户输入)
        if "correction" in content:
            content["correction"]["previousDraft"]["rationale"] = "其他回合的报告"
            return replace(
                request, 用户输入=json.dumps(content, ensure_ascii=False, sort_keys=True)
            )
        return request

    monkeypatch.setattr(flow, "模型请求", 改写)
    运行测试._运行就绪(env, allow_failure=True)
    assert len(env["received"]) == 3 and env["scripted"] == ["ok"]
    judge = next(u for u in _工作面(env)["units"] if u["kind"] == "comparison")
    assert judge["state"] == "failed" and len(judge["rejections"]) == 1


@pytest.mark.case_id(
    "NC-w25-25b008",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["S02已保存但B10事务失败时，不得绕过未登记回合重发"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_拒绝事务失败不能直接恢复重置次数__25b008(执行环境, monkeypatch):
    from muse.效果评测.存储 import 评测存储

    env = 执行环境
    _准备评委(env)
    env["scripted"].extend(["bad_quote", "ok"])
    save = 评测存储.保存引文拒绝

    def 写入失败(self, record):
        raise 评测错误("合成拒绝事务失败")

    monkeypatch.setattr(评测存储, "保存引文拒绝", 写入失败)
    运行测试._运行就绪(env, allow_failure=True)
    monkeypatch.setattr(评测存储, "保存引文拒绝", save)
    with pytest.raises(评测错误, match="未登记"):
        运行测试._恢复失败单元(env)
    assert len(env["received"]) == 3 and env["scripted"] == ["ok"]
    judge = next(u for u in _工作面(env)["units"] if u["kind"] == "comparison")
    assert len(judge["unregistered_deliveries"]) == 1 and not judge["rejections"]


def _漏登记(env, monkeypatch, kind):
    from copy import deepcopy

    if kind == "generation":
        运行测试._启动执行(env)
    else:
        _准备评委(env)
        env["scripted"].append("bad_quote" if kind == "comparison_bad" else "ok")
    service = type(env["app"].要求评测())
    save = service.保存执行交付
    captured = []

    def 中断(self, ctx, text, output, call):
        if not captured:
            captured.append(
                {
                    "unit_id": ctx.任务.冻结输入["输入"]["unit_id"],
                    "call_id": call,
                    "output": deepcopy(output),
                }
            )
            raise 评测错误("合成业务登记前中断")
        return save(self, ctx, text, output, call)

    monkeypatch.setattr(service, "保存执行交付", 中断)
    运行测试._运行就绪(env, allow_failure=True)
    monkeypatch.setattr(service, "保存执行交付", save)
    assert len(captured) == 1
    return captured[0]


@pytest.mark.case_id(
    "NC-w25-25b009",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["生成和评委原输出经S02复核后补登，错作者/伪造拒绝、原样幂等、恢复不重发旧回合"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
@pytest.mark.parametrize("kind", ["generation", "comparison_valid", "comparison_bad"])
def test_按S02原交付补登后恢复不重发已发生回合__25b009(执行环境, monkeypatch, kind):
    from copy import deepcopy
    from dataclasses import replace

    from muse.共享.错误 import Muse错误
    from muse.效果评测.接口 import 交付补登请求

    env = 执行环境
    svc = env["app"].要求评测()
    original = _漏登记(env, monkeypatch, kind)
    count = len(env["received"])
    with pytest.raises(评测错误, match="未登记"):
        运行测试._恢复失败单元(env)
    request = 交付补登请求(**original)
    with pytest.raises(评测错误, match="作者"):
        svc.补登记交付(replace(env["actor"], 作者="foreign"), request)
    forged = deepcopy(original)
    forged["output"]["fabricated"] = True
    with pytest.raises(Muse错误):
        svc.补登记交付(env["actor"], 交付补登请求(**forged))
    receipt = svc.补登记交付(env["actor"], request)
    assert svc.补登记交付(env["actor"], request) == receipt
    assert len(env["received"]) == count
    unit = next(u for u in _工作面(env)["units"] if u["unit_id"] == original["unit_id"])
    assert unit["unregistered_deliveries"] == []
    assert bool(unit["rejections"]) == (kind == "comparison_bad")
    运行测试._恢复失败单元(env)
    运行测试._运行就绪(env)
    assert len(env["received"]) == count + (kind == "comparison_bad")
    unit = next(u for u in _工作面(env)["units"] if u["unit_id"] == original["unit_id"])
    assert unit["state"] == "completed"


@pytest.mark.case_id(
    "NC-w25-25b00a",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["实际CLI补登与HTTP重放同一回执，零新增模型调用"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_实际CLI补登与HTTP重放不调用模型__25b00a(执行环境, monkeypatch, tmp_path):
    import subprocess
    import sys

    import test_生产评测权限隔离 as 数据测试
    from fastapi.testclient import TestClient

    from muse.接入.http.应用 import 创建应用
    from muse.配置 import 读取配置

    env = 执行环境
    original = _漏登记(env, monkeypatch, "comparison_bad")
    config = 数据测试._配置文件(env["pool"], tmp_path)
    request_file = tmp_path / "restore.json"
    request_file.write_text(json.dumps(original, ensure_ascii=False))
    result = subprocess.run(
        [sys.executable, "-I", "-m", "muse", "评测", str(config), "补登记交付", str(request_file)],
        cwd=tmp_path,
        capture_output=True,
        text=True,
        timeout=30,
    )
    assert result.returncode == 0, result.stderr
    receipt = json.loads(result.stdout)
    http = 创建应用(读取配置(config))
    with TestClient(http, headers={"origin": "http://testserver"}) as client:
        http.state.装配 = env["app"]
        assert (
            client.post(
                "/api/v1/session", json={"password": "synthetic-evaluation-only"}
            ).status_code
            == 200
        )
        replay = client.post("/api/v1/evaluation/deliveries/restore", json=original)
        assert replay.status_code == 200 and replay.json() == receipt
        report = client.get(
            "/api/v1/evaluation/experiments/" + env["exp"]["experiment_id"] + "/report"
        )
        assert report.status_code == 200
        units = report.json()["samples"][0]["units"]
        assert sum(len(u["rejections"]) for u in units) == 1
    assert len(env["received"]) == 3


@pytest.mark.case_id(
    "NC-w25-25b00b",
    environment="隔离PG与受控合成HTTP",
    given="隔离PG、固定任务与实际模型回合",
    when="触发本用例的纠正、登记失败、重签、恢复或停止边界",
    then=["停止后的未登记交付不能补写为新业务结果"],
    contract="docs/系统架构/新版设计/模块设计/B10-效果评测.md",
)
def test_实验停止后拒绝补登新产物__25b00b(执行环境, monkeypatch):
    from muse.效果评测.接口 import 交付补登请求

    env = 执行环境
    original = _漏登记(env, monkeypatch, "comparison_valid")
    svc = env["app"].要求评测()
    svc.取消实验(env["actor"], env["exp"]["experiment_id"], "stop-before-restore")
    with pytest.raises(评测错误, match="停止"):
        svc.补登记交付(env["actor"], 交付补登请求(**original))
    assert len(env["received"]) == 3
    assert sum(len(u["unregistered_deliveries"]) for u in _工作面(env)["units"]) == 1