feat: let document calls treat deterministic parse failures as rejected input

The service wraps every failure, including its own 400s, as a 500 with a parsing-failed detail. Only a fixed set of known bad-input markers is narrowed to a rejected request, and only when the caller passes markers; the shared default classification is unchanged.
2026-10-21 23:43:02 -04:00
parent 3db7d9979c
commit 0357d5afb3
2 changed files with 149 additions and 4 deletions
@@ -20,10 +20,12 @@ from polygateway.errors import (
    TransientError,
 )
 from polygateway.transports.monkey_ocr import (
+    _DEFAULT_REJECTED_MARKERS,
    MonkeyOcrTransport,
    _classify_status,
    _parse_middle_json,
    _sniff_media,
+    _translate_http_errors,
    _upload_part,
 )
 from polygateway.types import SourceConfig
@@ -455,6 +457,112 @@ class TestErrorTranslation:
        assert isinstance(ei.value.__cause__, httpx.TimeoutException)


+def _http_error(status, body):
+    request = httpx.Request("POST", "http://ocr.example/parse")
+    return httpx.HTTPStatusError(
+        str(status),
+        request=request,
+        response=httpx.Response(status, content=body.encode(), request=request),
+    )
+
+
+def _parsing_failed(reason):
+    # 真实形态(2026-10-21 实测): 服务把一切异常包成 500 + {"detail": "Parsing failed: …"}
+    return json.dumps({"detail": f"Parsing failed: {reason}"})
+
+
+class TestRejectedMarkers:
+    """#34 设计 §6: 确定性坏输入的 500 在文档路径归「请求被拒」,其余分类一概不动。
+
+    共享的 `_classify_status` 默认行为字节级不变——医院线对坏图的重试/熔断行为属公共行为,
+    推广须另行设计(issue #35)。
+    """
+
+    @pytest.mark.parametrize("marker", _DEFAULT_REJECTED_MARKERS)
+    def test_known_deterministic_failures_become_rejected_with_status_500(self, marker):
+        exc = _http_error(500, _parsing_failed(marker.capitalize()))
+        err = _classify_status(exc, "monkey_1", "parse", _DEFAULT_REJECTED_MARKERS)
+        assert isinstance(err, RequestRejectedError)
+        assert err.status_code == 500  # 服务有响应: 熔断按「响应即健康」记成功
+        assert marker.capitalize() in err.body_text  # 响应体摘要仍随该分支
+
+    def test_marker_matching_is_case_insensitive(self):
+        exc = _http_error(500, '{"detail": "PARSING FAILED: FAILED TO OPEN STREAM"}')
+        err = _classify_status(exc, "monkey_1", "parse", _DEFAULT_REJECTED_MARKERS)
+        assert isinstance(err, RequestRejectedError)
+
+    @pytest.mark.parametrize(
+        "body",
+        [
+            pytest.param('{"detail": "Failed to open stream"}', id="no-parsing-failed-prefix"),
+            pytest.param(
+                _parsing_failed("500: Model not initialized"), id="real-fault-same-prefix"
+            ),
+            pytest.param(
+                _parsing_failed("400: Unsupported file type: .webp"), id="unsupported-type"
+            ),
+            pytest.param(_parsing_failed("CUDA out of memory"), id="oom"),
+            pytest.param("", id="empty-body"),
+        ],
+    )
+    def test_everything_else_stays_transient(self, body):
+        """`Parsing failed:` 前缀不能单独当判据: 服务把真故障也包成同一前缀。
+        `Unsupported file type` 有意不在白名单——嗅探后它只可能是契约变化或我方标签 bug。"""
+        err = _classify_status(
+            _http_error(500, body), "monkey_1", "parse", _DEFAULT_REJECTED_MARKERS
+        )
+        assert isinstance(err, TransientError)
+
+    @pytest.mark.parametrize("body", [_parsing_failed(m) for m in _DEFAULT_REJECTED_MARKERS])
+    def test_default_call_without_markers_is_unchanged(self, body):
+        err = _classify_status(_http_error(500, body), "monkey_1", "parse")
+        assert isinstance(err, TransientError)
+
+    @pytest.mark.parametrize(
+        ("status", "exc_type"),
+        [
+            (429, TransientError),
+            (502, TransientError),
+            (503, TransientError),
+            (401, SourceDeadError),
+            (403, SourceDeadError),
+            (404, RequestRejectedError),
+        ],
+    )
+    def test_other_statuses_are_never_narrowed(self, status, exc_type):
+        body = _parsing_failed(_DEFAULT_REJECTED_MARKERS[0])
+        err = _classify_status(_http_error(status, body), "m", "parse", _DEFAULT_REJECTED_MARKERS)
+        assert type(err) is exc_type
+
+    def test_translate_http_errors_passes_markers_through(self):
+        with pytest.raises(RequestRejectedError), _translate_http_errors(
+            "monkey_1", "parse", rejected_markers=_DEFAULT_REJECTED_MARKERS
+        ):
+            raise _http_error(500, _parsing_failed(_DEFAULT_REJECTED_MARKERS[0]))
+        # 默认不传: 行为不变
+        with pytest.raises(TransientError), _translate_http_errors("monkey_1", "parse"):
+            raise _http_error(500, _parsing_failed(_DEFAULT_REJECTED_MARKERS[0]))
+
+    def test_markers_are_configurable_per_transport(self):
+        assert MonkeyOcrTransport()._rejected_markers == _DEFAULT_REJECTED_MARKERS
+        custom = ("synthetic marker",)
+        assert MonkeyOcrTransport(deterministic_failure_markers=custom)._rejected_markers == custom
+
+    async def test_layout_and_text_paths_do_not_apply_the_markers(self):
+        """传输实例持有标记,但只有文档路径使用;医院线的两个方法行为不变。"""
+        body = _parsing_failed(_DEFAULT_REJECTED_MARKERS[0]).encode()
+        t = _transport_for(
+            _routes(
+                text_resp=httpx.Response(500, content=body),
+                parse_resp=httpx.Response(500, content=body),
+            )
+        )
+        with pytest.raises(TransientError):
+            await t.recognize_text(image=b"jpg", source=_source(), call_id="c1")
+        with pytest.raises(TransientError):
+            await t.parse_layout(image=b"jpg", source=_source(), call_id="c1")
+
+
 class TestCheckHealth:
    async def test_healthy(self):
        t = _transport_for(_routes(health_resp=httpx.Response(200, json={"status": "healthy"})))