[Bug]: 使用内置 MiniMax 供应商且模型选择MiniMax-M3时，图片无法被模型正确识别

## QwenPaw Version

QwenPaw, version 2.0.0.post4

## Description

使用 QwenPaw 内置的 MiniMax 供应商（默认走 Anthropic 兼容协议）时，模型完全无法识别上传图片的内容，会返回与图片无关的乱编答案。

## Component(s) Affected

- [x] Core / Backend (app, agents, config, providers, utils, local_models)
- [ ] Console (frontend web UI)
- [ ] Channels (DingTalk, Feishu, QQ, Discord, iMessage, etc.)
- [ ] Skills
- [ ] CLI
- [ ] Documentation (website)
- [ ] Tests
- [ ] CI/CD
- [ ] Scripts / Deploy

## Environment

- **QwenPaw version:** 2.0.0.post4
- **OS:** Windows 11
- **Install method:** 桌面版安装
- **Model:** MiniMax-M3

## Steps to Reproduce

1.打开 QwenPaw，选择内置的 MiniMax 供应商。
2.上传任意一张显示时间日期的截图。
3.发送问题："图上的年日月时分秒是多少？"。

## Actual vs Expected

- **Actual:** 模型完全无法识别图片内容，返回与图片毫无关系的乱码/编造内容。
- **Expected:** 模型应能正确识别图片中日期。

## Logs / Screenshots

上传的图片：

对话：

Start by tracing the built-in MiniMax provider and its Anthropic-compatible request path in the core/backend components, focusing on how uploaded images are represented and sent for MiniMax-M3. Reproduce the reported date-recognition failure on Windows 11, then verify that the model receives the image content and correctly identifies it without breaking text-only requests.

Vom Indexierungsmodell aus dem Issue-Text verfasst.