fix: close adversarial review findings on document parsing

Malformed download URLs are rejected before the second request, error bodies are read only up to a bound, the telemetry summary keeps structural dispatch for duck-typed results, and the BMP magic now requires a valid DIB header. Docs and design are aligned with the implementation, and new test wording is neutral.

research-wiki/designs/2026-10-21-issue34-ocr-document-design.md

@@ -70,7 +70,8 @@ issue 把症状拆成两条;**两条同根**: 库把 OCR 输入建模成"一张
 | --- | --- | --- | --- |
 | 正常 PDF | `doc.pdf`,整份全文(E4 修复) | `doc.pdf`(上传分支变更,见 §8) | `doc.pdf`,完整文档结果 |
 | 前缀垃圾的 PDF | 同上 | 同上 | 同上 |
-| 已知图片 / 其它可解码图片 | `image.jpg`(不变) | `image.jpg`(不变) | `image.jpg`,1 页结果 |
+| 魔数表内的图片(JPEG/PNG/TIFF/BMP/GIF/WebP) | `image.jpg`(不变) | `image.jpg`(不变) | `image.jpg`,1 页结果 |
+| 魔数表之外的可解码图片 | `image.jpg`(不变,服务按内容解码) | `image.jpg`(不变) | **本地 `RequestRejectedError`**(请改走 `parse_layout`;表内格式的扩充另议) |
 | 无法识别的字节(HTML 页面、乱码、空) | `image.jpg`(不变,服务给幻觉文本 E13) | `image.jpg`(不变,服务 500) | **本地 `RequestRejectedError`,不发 HTTP** |
 
 ### Q3 文字内容从哪里来