fix: close independent verification findings on document parsing

Zip construction and member reads now share one narrow boundary so corrupt BZIP2, LZMA and version fields cannot leak native exceptions. Non-success statuses are read before classification, text-bearing blocks require lines, the size cap and reasoning contract are pinned by tests that fail under mutation, and the per-page timing claim is corrected to the observed range.

tests/unit/test_monkey_ocr.py

@@ -704,27 +704,83 @@ class TestParseDocument:
                 document=_PDF_BODY, source=_source(), call_id="c1"
             )
 
-    async def test_oversized_declared_content_length_is_rejected(self, monkeypatch):
-        monkeypatch.setattr(monkey_ocr, "_MAX_PACKAGE_BYTES", 100)
-        resp = httpx.Response(200, content=b"x" * 10, headers={"content-length": "101"})
-        with pytest.raises(ResultInvalidError):
+    @staticmethod
+    def _oversized_valid_package():
+        """**合法**的 ZIP 但整体超过上限: 去掉上限检查时它能被正常解析,用例才会转红。
+
+        (独立验证发现: 旧用例用非 ZIP 垃圾字节,去掉上限后同样抛 ResultInvalidError——
+        走的是 BadZipFile 路径,上限从未被测到。)
+        """
+        return _doc_package(
+            [_doc_page([_text_block("SYNTH-A")])], extra=[("images/big.jpg", b"x" * 5000)]
+        )
+
+    async def _assert_cap_rejects_before_parsing(self, monkeypatch, resp):
+        monkeypatch.setattr(monkey_ocr, "_MAX_PACKAGE_BYTES", 2000)
+        parsed = []
+
+        def spy(package):
+            parsed.append(len(package))
+            return [], [(1.0, 1.0)], ""
+
+        monkeypatch.setattr(monkey_ocr, "parse_document_package", spy)
+        with pytest.raises(ResultInvalidError, match="体积超过上限"):
             await self._transport(zip_resp=resp).parse_document(
                 document=_PDF_BODY, source=_source(), call_id="c1"
             )
+        assert parsed == []  # 解析函数根本没被调用
 
-    async def test_oversized_streamed_body_is_rejected_by_cumulative_count(self, monkeypatch):
-        """没有 Content-Length(分块传输)时靠累计字节兜底: 多个小块合起来超限。"""
-        monkeypatch.setattr(monkey_ocr, "_MAX_PACKAGE_BYTES", 100)
+    async def test_oversized_valid_package_is_rejected_by_declared_content_length(
+        self, monkeypatch
+    ):
+        resp = httpx.Response(200, content=self._oversized_valid_package())  # 自动带 Content-Length
+        await self._assert_cap_rejects_before_parsing(monkeypatch, resp)
+
+    async def test_oversized_valid_package_is_rejected_by_cumulative_count(self, monkeypatch):
+        """没有 Content-Length(分块传输)时靠累计字节兜底。"""
+        package = self._oversized_valid_package()
 
         async def chunks():
-            for _ in range(3):
-                yield b"x" * 60
+            for offset in range(0, len(package), 700):
+                yield package[offset : offset + 700]
 
-        resp = httpx.Response(200, content=chunks())
-        with pytest.raises(ResultInvalidError):
-            await self._transport(zip_resp=resp).parse_document(
-                document=_PDF_BODY, source=_source(), call_id="c1"
-            )
+        await self._assert_cap_rejects_before_parsing(
+            monkeypatch, httpx.Response(200, content=chunks())
+        )
+
+    async def test_package_under_the_cap_still_parses(self, monkeypatch):
+        """对照: 上限只拦超限的,不误伤正常包。"""
+        monkeypatch.setattr(monkey_ocr, "_MAX_PACKAGE_BYTES", 10_000)
+        result = await self._transport(
+            zip_resp=httpx.Response(200, content=self._oversized_valid_package())
+        ).parse_document(document=_PDF_BODY, source=_source(), call_id="c1")
+        assert [e.text for e in result.elements] == ["SYNTH-A"]
+
+    @pytest.mark.parametrize(
+        ("status", "exc_type"),
+        [
+            (401, SourceDeadError),
+            (404, RequestRejectedError),
+            (500, TransientError),
+            (302, RequestRejectedError),  # 3xx 不是 `is_error`,但 raise_for_status 会抛
+        ],
+    )
+    async def test_streamed_error_body_is_read_before_classification(self, status, exc_type):
+        """真正的流式响应(异步迭代器 body)不会被预缓冲: 不先 `aread()`,摘要读到的是空串。
+
+        (独立验证发现: `httpx.Response(content=bytes)` 已预缓冲,删掉 `aread()` 后旧用例仍绿;
+        且旧条件 `is_error` 漏掉 3xx。)
+        """
+        body = '{"detail": "synthetic upstream reason"}'
+
+        async def streamed():
+            yield body.encode()
+
+        t = self._transport(zip_resp=httpx.Response(status, content=streamed()))
+        with pytest.raises(exc_type) as ei:
+            await t.parse_document(document=_PDF_BODY, source=_source(), call_id="c1")
+        assert ei.value.status_code == status
+        assert ei.value.body_text == body
 
     async def test_bad_package_is_result_invalid(self):
         t = self._transport(zip_resp=httpx.Response(200, content=b"not a zip"))