feat: 结构化输出按官方规范 + 两段式重试 + DeepSeek 复检开关

- 豆包: response_format.json_schema(strict)；DeepSeek: 切 Responses API text.format.json_schema
  （官方 chat/completions 通道不支持 json_schema），不支持时自动降级 json_object（[output] schema）
- 重试两段式：传输错误 cfg.retries 次；解析失败/空正文有独立 3 次专用重试，仍失败落 parse_fail
- 移除三票复核（decide_final/_verify_clean/verify_clean），DeepSeek 单次判定即终稿
- 新增 [run] deepseek_recheck 开关（默认 no=仅豆包初筛；yes=追加 DeepSeek 复检）
- GUI 不再覆盖 config 提示词；prompt 相对/前导斜杠路径按 exe 目录解析
- 报表去「票型/复核」列并同步说明；README/config.example.ini 同步
- 测试新增输出格式、重试、开关与提示词路径用例（50 passed）

tests/test_config.py

@@ -1,7 +1,9 @@
 # -*- coding: utf-8 -*-
 """config 模块测试：模板生成、缺项检测、读写回环、提示词路径。"""
+import violation_detector.config as config_mod
 from violation_detector.config import (
-    AppConfig, ensure_config, load_config, missing_fields, save_config,
+    AppConfig, effective_mode, ensure_config, load_config, missing_fields,
+    save_config,
 )
 
 
@@ -63,18 +65,57 @@ def test_mode_from_ini(tmp_path):
     assert load_config(str(ini)).mode == "cascade"
 
 
-def test_verify_clean_from_ini(tmp_path):
+def test_output_schema_from_ini(tmp_path):
     ini = tmp_path / "config.ini"
-    ini.write_text("[run]\nverify_clean = no\n", encoding="utf-8")
-    assert load_config(str(ini)).verify_clean is False
-    ini.write_text("[run]\nverify_clean = yes\n", encoding="utf-8")
-    assert load_config(str(ini)).verify_clean is True
-    # 非法值保持默认开启
-    ini.write_text("[run]\nverify_clean = 也许\n", encoding="utf-8")
-    assert load_config(str(ini)).verify_clean is True
-    # 未配置保持默认开启
+    ini.write_text("[output]\nschema = on\n", encoding="utf-8")
+    assert load_config(str(ini)).output_schema_mode == "on"
+    ini.write_text("[output]\nschema = off\n", encoding="utf-8")
+    assert load_config(str(ini)).output_schema_mode == "off"
+    ini.write_text("[output]\nschema = auto\n", encoding="utf-8")
+    assert load_config(str(ini)).output_schema_mode == "auto"
+    # 非法值回退 auto；未配置默认 auto
+    ini.write_text("[output]\nschema = 乱来\n", encoding="utf-8")
+    assert load_config(str(ini)).output_schema_mode == "auto"
     ini.write_text("[ark]\n", encoding="utf-8")
-    assert load_config(str(ini)).verify_clean is True
+    assert load_config(str(ini)).output_schema_mode == "auto"
+
+
+def test_output_schema_roundtrip(tmp_path):
+    ini = tmp_path / "config.ini"
+    save_config(AppConfig(output_schema_mode="off"), str(ini))
+    assert load_config(str(ini)).output_schema_mode == "off"
+    assert AppConfig().output_schema_mode == "auto"
+
+
+def test_deepseek_recheck_default_and_ini(tmp_path):
+    # 默认关
+    assert AppConfig().deepseek_recheck is False
+    ini = tmp_path / "config.ini"
+    ini.write_text("[run]\ndeepseek_recheck = yes\n", encoding="utf-8")
+    assert load_config(str(ini)).deepseek_recheck is True
+    ini.write_text("[run]\ndeepseek_recheck = no\n", encoding="utf-8")
+    assert load_config(str(ini)).deepseek_recheck is False
+    # 非法值/未配置保持默认关
+    ini.write_text("[run]\ndeepseek_recheck = 也许\n", encoding="utf-8")
+    assert load_config(str(ini)).deepseek_recheck is False
+    ini.write_text("[ark]\n", encoding="utf-8")
+    assert load_config(str(ini)).deepseek_recheck is False
+
+
+def test_effective_mode_respects_recheck_switch():
+    on = AppConfig(mode="cascade", deepseek_recheck=True)
+    off = AppConfig(mode="cascade", deepseek_recheck=False)
+    assert effective_mode(on, "cascade") == "cascade"
+    assert effective_mode(off, "cascade") == "doubao"
+    # 开关只影响 cascade；doubao/deepseek 不受影响
+    assert effective_mode(off, "doubao") == "doubao"
+    assert effective_mode(off, "deepseek") == "deepseek"
+
+
+def test_deepseek_recheck_roundtrip(tmp_path):
+    ini = tmp_path / "config.ini"
+    save_config(AppConfig(deepseek_recheck=True), str(ini))
+    assert load_config(str(ini)).deepseek_recheck is True
 
 
 def test_save_and_load_roundtrip(tmp_path):
@@ -104,3 +145,21 @@ def test_prompt_path_custom(tmp_path):
     p.write_text("x", encoding="utf-8")
     cfg = AppConfig(prompt_file=str(p))
     assert cfg.prompt_path == p
+
+
+def test_prompt_path_relative_prefers_app_dir(tmp_path, monkeypatch):
+    (tmp_path / "prompts.txt").write_text("x", encoding="utf-8")
+    monkeypatch.setattr(config_mod, "app_dir", lambda: tmp_path)
+    monkeypatch.setattr(config_mod, "resource_dir", lambda: tmp_path)
+    cfg = AppConfig(prompt_file="prompts.txt")
+    assert cfg.prompt_path == tmp_path / "prompts.txt"
+
+
+def test_prompt_path_leading_slash_treated_relative(tmp_path, monkeypatch):
+    # 无盘符的 /prompts.txt 应视为 exe 同目录的相对文件（Windows 容错）
+    (tmp_path / "prompts.txt").write_text("x", encoding="utf-8")
+    monkeypatch.setattr(config_mod, "app_dir", lambda: tmp_path)
+    monkeypatch.setattr(config_mod, "resource_dir", lambda: tmp_path)
+    for raw in ("/prompts.txt", "\\prompts.txt"):
+        cfg = AppConfig(prompt_file=raw)
+        assert cfg.prompt_path == tmp_path / "prompts.txt", raw

tests/test_pipeline.py

@@ -1,7 +1,7 @@
 # -*- coding: utf-8 -*-
-"""pipeline 模块测试：排序、投票裁决、成本计算、缓存路径。"""
+"""pipeline 模块测试：排序、成本计算、缓存路径。"""
 from violation_detector.pipeline import (
-    _cache_path, _cat_num, deepseek_cost, decide_final, natural_key,
+    _cache_path, _cat_num, deepseek_cost, natural_key,
 )
 
 
@@ -19,42 +19,11 @@ def test_natural_key_sort():
 def test_cat_num():
     assert _cat_num("12. 侵权 - 其他") == 12
     assert _cat_num("1. 烟草") == 1
-    # 无编号的结论统一排到最后（无违规/违规不变动分类，不影响裁决）
+    # 无编号的结论统一排到最后（无违规/违规不明不变动分类）
     assert _cat_num("无违规") == 99
     assert _cat_num("违规不明") == 99
 
 
-def test_decide_final_unanimous_clean():
-    final, note = decide_final([R("无违规"), R("无违规"), R("无违规")])
-    assert final["category"] == "无违规"
-    assert "3票一致" in note
-
-
-def test_decide_final_clean_majority_keeps_but_flags():
-    final, note = decide_final([R("无违规"), R("无违规"), R("12. 侵权 - 除人物外的其他侵权")])
-    assert final["category"] == "无违规"
-    assert "2:1" in note and "人工复核" in note
-
-
-def test_decide_final_violation_majority_wins():
-    final, note = decide_final(
-        [R("无违规"), R("12. 侵权 - 除人物外的其他侵权"), R("12. 侵权 - 除人物外的其他侵权")])
-    assert final["category"].startswith("12.")
-    assert "人工复核" in note
-
-
-def test_decide_final_three_way_split_takes_lowest_number():
-    final, note = decide_final([R("无违规"), R("13. 侵权 - 人物相关"), R("5. 负向敏感")])
-    assert final["category"].startswith("5.")
-    assert "从严" in note
-
-
-def test_decide_final_two_violations_differ_majority_none():
-    # 两张违规票不同、一张无违规：违规方 2 票 > 1，并列取序号最小
-    final, _ = decide_final([R("无违规"), R("13. 侵权 - 人物相关"), R("12. 侵权 - 其他")])
-    assert final["category"].startswith("12.")
-
-
 def test_deepseek_cost_known_usage():
     # miss 407 + cached 1792 + completion 2282
     results = {"a": R("无违规")}

tests/test_providers.py

@@ -1,10 +1,17 @@
 # -*- coding: utf-8 -*-
-"""providers 模块测试：判定 JSON 解析、分类归一化、图片编码、usage 规整。"""
+"""providers 模块测试：判定 JSON 解析、分类归一化、图片编码、usage 规整、output_schema、重试。"""
+import asyncio
 import base64
 
 import pytest
 
-from violation_detector.providers import CANON, _norm_usage, encode_image, norm_category, parse_judgment
+from violation_detector.config import AppConfig
+from violation_detector.providers import (
+    CANON, OUTPUT_SCHEMA, SCHEMA_CATEGORIES, SCHEMA_NAME, _chat_detect,
+    _extract_output_text, _norm_usage, _result_responses,
+    _schema_explicitly_unsupported, build_chat_payload,
+    build_responses_payload, encode_image, norm_category, parse_judgment,
+)
 
 GOOD = ('{"文字/图形属性": "黑色T恤印字母", "侵权/违规逻辑": "未授权使用商标", '
         '"违规分类": "12. 侵权 - 除人物外的其他侵权"}')
@@ -67,3 +74,196 @@ def test_norm_usage_handles_missing_details():
         "prompt": 100, "completion": 50, "cached": 0}
     assert _norm_usage({}) == {"prompt": None, "completion": None, "cached": 0}
     assert _norm_usage({"prompt_tokens": 100, "prompt_tokens_details": {"cached_tokens": 80}})["cached"] == 80
+
+
+# ---------- output_schema ----------
+
+def test_output_schema_shape():
+    assert OUTPUT_SCHEMA["type"] == "object"
+    assert set(OUTPUT_SCHEMA["required"]) == {"文字/图形属性", "侵权/违规逻辑", "违规分类"}
+    assert OUTPUT_SCHEMA["additionalProperties"] is False
+    assert set(OUTPUT_SCHEMA["properties"]) == set(OUTPUT_SCHEMA["required"])
+
+
+def test_schema_categories_covers_canon_and_specials():
+    expect = [CANON[n] for n in sorted(CANON)] + ["无违规", "违规不明"]
+    assert SCHEMA_CATEGORIES == expect
+
+
+def test_build_chat_payload_schema_vs_json():
+    # 豆包：结构化输出 = response_format.json_schema(strict)；降级 = json_object
+    messages = [{"role": "user", "content": []}]
+    s = build_chat_payload("m", messages, {"temperature": 0}, True)
+    assert s["model"] == "m" and s["temperature"] == 0
+    rf = s["response_format"]
+    assert rf["type"] == "json_schema"
+    assert rf["json_schema"]["name"] == SCHEMA_NAME
+    assert rf["json_schema"]["strict"] is True
+    assert rf["json_schema"]["schema"] is OUTPUT_SCHEMA
+    j = build_chat_payload("m", messages, {}, False)
+    assert j["response_format"] == {"type": "json_object"}
+
+
+def test_build_responses_payload_schema():
+    # DeepSeek：结构化输出 = text.format.json_schema（Responses API）
+    items = [{"role": "user", "content": [{"type": "input_text", "text": "p"}]}]
+    p = build_responses_payload("m", items, {"max_output_tokens": 5000}, True)
+    assert p["model"] == "m" and p["max_output_tokens"] == 5000
+    fmt = p["text"]["format"]
+    assert fmt["type"] == "json_schema"
+    assert fmt["name"] == SCHEMA_NAME
+    assert fmt["schema"] is OUTPUT_SCHEMA
+
+
+def test_extract_output_text_concats_output_text():
+    data = {"output": [
+        {"type": "reasoning", "role": "assistant",
+         "content": [{"type": "reasoning_text", "text": "思考"}],
+         "status": "completed"},
+        {"type": "message", "role": "assistant",
+         "content": [{"type": "output_text", "text": '{"a": 1}'}],
+         "status": "completed"},
+    ]}
+    assert _extract_output_text(data) == '{"a": 1}'
+
+
+def test_result_responses_ok():
+    data = {
+        "status": "completed",
+        "usage": {"input_tokens": 10, "output_tokens": 20,
+                  "input_tokens_details": {"cached_tokens": 4}},
+        "output": [{"type": "message", "role": "assistant", "content": [
+            {"type": "output_text", "text": '{"文字/图形属性": "T恤", '
+                                            '"侵权/违规逻辑": "无", "违规分类": "无违规"}'}]}],
+    }
+    r = _result_responses("a.jpg", data)
+    assert r["status"] == "ok" and r["category"] == "无违规"
+    assert r["usage"] == {"prompt": 10, "completion": 20, "cached": 4}
+    assert r["finish_reason"] == "completed"
+
+
+def test_schema_explicitly_unsupported():
+    assert _schema_explicitly_unsupported(
+        "Unrecognized request argument supplied: response_format.json_schema")
+    assert _schema_explicitly_unsupported("model not support json_schema")
+    assert _schema_explicitly_unsupported("不支持的参数 json_schema")
+    assert not _schema_explicitly_unsupported("HTTP 400 some other problem")
+
+
+# ---------- 重试机制（错误码 / 空正文 / 解析失败） ----------
+
+RETRY_GOOD = '{"文字/图形属性": "T恤", "侵权/违规逻辑": "无", "违规分类": "无违规"}'
+RETRY_BAD = '{"文字/图形属性": "T恤"}'  # 缺字段，解析失败
+
+
+def _chat_ok(content):
+    return {"choices": [{"message": {"content": content}, "finish_reason": "stop"}],
+            "usage": {}}
+
+
+class _FakeResp:
+    def __init__(self, status, data=None, body=""):
+        self.status = status
+        self._data = data
+        self._body = body
+
+    async def json(self):
+        return self._data
+
+    async def text(self):
+        return self._body
+
+
+class _FakeCM:
+    def __init__(self, resp):
+        self._resp = resp
+
+    async def __aenter__(self):
+        return self._resp
+
+    async def __aexit__(self, *exc):
+        return False
+
+
+class _FakeSession:
+    def __init__(self, responses):
+        self.responses = list(responses)
+        self.posts = 0
+
+    def post(self, url, json, headers):
+        resp = self.responses[min(self.posts, len(self.responses) - 1)]
+        self.posts += 1
+        return _FakeCM(resp)
+
+
+def _run_chat(session, retries=3):
+    cfg = AppConfig(output_schema_mode="off", retries=retries)
+    return asyncio.run(_chat_detect(
+        session, asyncio.Semaphore(1), cfg, "a.jpg", stop=lambda: False,
+        provider="deepseek", label="DeepSeek",
+        base_url="https://fake.deepseek.com", api_key="k", model="m",
+        messages=[{"role": "user", "content": "p"}],
+        extras={"temperature": 0, "max_tokens": 50}))
+
+
+async def _no_sleep(_):
+    return None
+
+
+def test_parse_fail_retries_then_recovers(monkeypatch):
+    # 第 1 次返回不可解析 → 触发重试；第 2 次成功
+    monkeypatch.setattr(asyncio, "sleep", _no_sleep)
+    sess = _FakeSession([_FakeResp(200, _chat_ok(RETRY_BAD)),
+                         _FakeResp(200, _chat_ok(RETRY_GOOD))])
+    res = _run_chat(sess)
+    assert res["status"] == "ok" and res["category"] == "无违规"
+    assert sess.posts == 2
+
+
+def test_parse_fail_exhausts_to_parse_fail(monkeypatch):
+    # 一直解析失败 → 第 1 次进入解析段，独立再重试 3 次后落 parse_fail（违规不明）
+    monkeypatch.setattr(asyncio, "sleep", _no_sleep)
+    sess = _FakeSession([_FakeResp(200, _chat_ok(RETRY_BAD))])
+    res = _run_chat(sess, retries=3)
+    assert res["status"] == "parse_fail" and res["category"] == "违规不明"
+    assert sess.posts == 1 + 3  # 首判 + 3 次独立解析重试
+
+
+def test_parse_fail_dedicated_budget_ignores_http_errors(monkeypatch):
+    # 先 1 次 HTTP 500，再解析失败：HTTP 错误不消耗解析段次数，
+    # 解析仍独立获得 cfg.retries 次专用重试
+    monkeypatch.setattr(asyncio, "sleep", _no_sleep)
+    sess = _FakeSession([_FakeResp(500, body="boom"),
+                         _FakeResp(200, _chat_ok(RETRY_BAD))])
+    res = _run_chat(sess, retries=3)
+    assert res["status"] == "parse_fail" and res["category"] == "违规不明"
+    assert sess.posts == 2 + 3  # 1×HTTP500 + 首判解析失败 + 3×解析专用重试
+
+
+def test_error_code_retries_then_recovers(monkeypatch):
+    # HTTP 500 → 退避重试；随后 200 成功
+    monkeypatch.setattr(asyncio, "sleep", _no_sleep)
+    sess = _FakeSession([_FakeResp(500, body="boom"),
+                         _FakeResp(200, _chat_ok(RETRY_GOOD))])
+    res = _run_chat(sess)
+    assert res["status"] == "ok" and res["category"] == "无违规"
+    assert sess.posts == 2
+
+
+def test_error_code_exhausts_to_error(monkeypatch):
+    # 一直 HTTP 500 → 重试耗尽落 error（检测异常）
+    monkeypatch.setattr(asyncio, "sleep", _no_sleep)
+    sess = _FakeSession([_FakeResp(500, body="boom")])
+    res = _run_chat(sess, retries=3)
+    assert res["status"] == "error" and res["category"] == "检测异常"
+    assert sess.posts == 3
+
+
+def test_empty_content_retries(monkeypatch):
+    # 空正文 → 重试；随后成功
+    monkeypatch.setattr(asyncio, "sleep", _no_sleep)
+    sess = _FakeSession([_FakeResp(200, _chat_ok("")),
+                         _FakeResp(200, _chat_ok(RETRY_GOOD))])
+    res = _run_chat(sess)
+    assert res["status"] == "ok" and res["category"] == "无违规"
+    assert sess.posts == 2

tests/test_report.py

@@ -7,7 +7,7 @@ from violation_detector.report import build_report, organize_output, sanitize_di
 
 def ROW(cat, fname="img.jpg", channel="豆包初筛"):
     return {"file": fname, "category": cat, "channel": channel, "attr": "属性",
-            "logic": "逻辑", "status": "ok", "vote_note": "", "remark": ""}
+            "logic": "逻辑", "status": "ok", "remark": ""}
 
 
 def test_sanitize_dirname():
@@ -48,7 +48,6 @@ def test_build_report_structure(tmp_path):
     rows = [ROW("12. 侵权 - 除人物外的其他侵权", "a.jpg"),
             ROW("无违规", "b.jpg", channel="DeepSeek复检"),
             ROW("无违规", "b.jpg")]
-    rows[2]["vote_note"] = "票型2:1（无违规 / 无违规 / 12. 侵权），建议人工复核"
     out = build_report(rows, str(tmp_path), {"mode": "cascade", "ds_calls": 3,
                                              "ds_peak_cost": 0.1, "ds_idle_cost": 0.05})
     assert out.endswith(".xlsx")
@@ -56,13 +55,12 @@ def test_build_report_structure(tmp_path):
     wb = load_workbook(out)
     ws = wb["检测结果"]
     assert ws["B2"].value.startswith("商品图合规检测工具")
-    headers = [ws.cell(row=4, column=c).value for c in range(2, 10)]
+    headers = [ws.cell(row=4, column=c).value for c in range(2, 9)]
     assert headers == ["序号", "文件名", "违规分类", "判定通道", "文字/图形属性",
-                       "侵权/违规逻辑", "票型/复核", "备注"]
+                       "侵权/违规逻辑", "备注"]
     assert ws.cell(row=5, column=2).value == 1
     assert ws.cell(row=6, column=5).value == "DeepSeek复检"
     assert ws.cell(row=5, column=5).value == "豆包初筛"
-    assert "人工复核" in ws.cell(row=7, column=8).value
 
     ws2 = wb["统计汇总"]
     cats = [ws2.cell(row=r, column=2).value for r in (5, 6)]