feat(openai): optional vision gating to strip images for non-vision models

[JumpLink]

feat(openai): optional vision gating to strip images for non-vision models

Refs danny-avila/LibreChat#11418, danny-avila/LibreChat#14341

Problem

When image content reaches a model without vision support, OpenAI-compatible
providers reject the entire request:

• model is not a multimodal model (Scaleway/Mistral)
• No endpoints found that support image input (OpenRouter)

The caller often cannot prevent the image from being in the history:

• a tool returns an image (an image-generation tool feeding its artifact back),
• or an agent hands off from a vision-capable model to a text-only one, carrying
  the user's original image_url block along - reported independently in
  LibreChat#14341.

There is currently no way to tell the client "this model cannot take images".

Approach

An opt-in vision flag on the OpenAI-family chat classes. When vision: false,
image parts are stripped at a single choke point, immediately before the message
stream is delegated to the base class - so it does not depend on the internal
message-conversion path.

stripImagesFromMessages(messages, visionCapable):

• returns the input unchanged when visionCapable is true (the default), so
  existing behaviour is untouched;
• clones only the messages that actually carry an image;
• keeps a short text placeholder when stripping would leave a message empty;
• covers both OpenAI-style image_url parts and standard image data
  content blocks. The latter matters: they are converted to image_url further
  downstream, so dropping only image_url still lets an uploaded image through
  to a text-only model.

Wired into ChatOpenAI, AzureChatOpenAI, ChatDeepSeek and ChatXAI.
ChatOpenAI._streamRawResponseChunks is covered too, which is the path
ChatOpenRouter streams through.

API

new ChatOpenAI({ /* ...existing fields... */, vision: false });

Defaults to true everywhere - non-breaking.

Tests

src/llm/openai/utils/stripImages.test.ts: 7 cases over image_url parts, data
content blocks, the placeholder substitution, and the vision-capable pass-through.
Full src/llm/openai + src/messages suites pass (623), tsc and eslint clean.

Signed-off-by: JumpLink pascal@artandcode.studio