fix: VLM image analysis for Responses-protocol providers (input_text injection)

[Floatend]

Problem

Configuring VLM image analysis (图片分析路由) for a Responses-protocol provider breaks every image request. The relay analyzes the image with the configured vision model, but then injects the description using a Chat-Completions style block:

{"type": "text", "text": "..."}

Upstream Responses APIs reject this block type, e.g. DeepSeek:

Failed to deserialize the JSON body into the target type: input: unknown variant `text`,
expected one of `input_text`, `output_text`, `input_image`, `input_file`

This is presumably why the manager UI hard-disables VLM/Strip for Responses-protocol providers (vlmUnsupportedProtocol). The rest of the VLM pipeline already handles the input array (URL collection recognizes input_image, stripping removes input_image blocks), so only the injection format is broken.

Changes

• vision.rs: inject_text_into_message / inject_analysis / strip_image_blocks now carry a responses: bool flag and emit {"type": "input_text"} blocks for Responses requests instead of {"type": "text"}. Chat behavior unchanged.
• protocol_proxy.rs: passes relay.protocol == RelayProtocol::Responses into strip_image_blocks.
• App.tsx: removes the Responses-protocol disable so VLM/Strip can be configured for those providers; adds friendlier option labels (原样发送图片 / 移除图片 / 视觉辅助分析).

Tests

• New: inject_analysis_uses_input_text_block_for_responses_protocol (input_text for responses)
• New: inject_text_into_message_keeps_chat_text_block_when_not_responses (text preserved for chat)
• Full suite on current main: 252 passed, 0 failed. tsc --noEmit clean.

Verification

Tested end-to-end with a DeepSeek (responses) provider + Qwen VLM: sending an image no longer errors; the image is analyzed by the VLM and the description reaches DeepSeek as text.

Note: the UI unlock (vlmUnsupportedProtocol = false) is intentional — VLM/Strip now work for Responses providers thanks to the injection fix. Chat Completions providers are unaffected.

[BigPizzaV3]

维护者审查：修复方向正确。Responses 请求中的文本内容块应使用 input_text，当前改动也保留了 Chat Completions 的 text 行为。已批准来自 fork 的三平台 CI。

合并前建议再补一项完整链路回归测试：构造含 input_image 的 Responses 请求，经 strip_image_blocks 处理后断言图片块被移除、VLM 描述以 input_text 注入，并且最终请求体不再包含 Chat 格式的新增 text 块。当前新增测试只直接调用 inject_analysis，尚未覆盖 protocol_proxy -> strip_image_blocks 的协议标志传递。CI 通过且该链路测试补齐后可以合并。

[Floatend]

已补齐完整链路回归测试,见新提交 4f291cd:

strip_image_blocks_injects_input_text_for_responses_protocol(wiremock 真实 HTTP):

1. 构造含 input_image 块的 Responses 格式请求(与 Codex 发往中继的格式一致)
2. 经 strip_image_blocks(..., responses=true) 完整链路处理(mock VLM 返回描述)
3. 断言:图片块被移除 / 描述以 input_text 注入 / 请求体不出现 Chat 格式的 text 块

该测试覆盖了 protocol_proxy → strip_image_blocks 的协议标志传递对应的行为。测试结果:

• lib 测试 217 passed, 0 failed(新增 1 个)
• 全量 cargo test:lib 217 + integration 36,全部通过(唯一失败项为 Windows 环境 symlink 权限测试,与本次改动无关)
• tsc --noEmit 通过

CI 结果出来后可随时合并,多谢 review!