enable image input for vision-capable prime inference models

[kevinjosethomas]

• prime inference models were all marked text-only, so images got dropped with a "model does not support images" message even on claude, gpt-5, and other vision models.
• prime inference's models api reports no modality data, so vision support can't be derived like other providers and is now curated via a flag on the model metadata.
• verified each flagged model actually reads images by probing the live api with a test image; minimax-m3 correctly stays text-only.

---

Note

Low Risk
Metadata-only change to declared input modalities; main risk is a wrong vision flag advertising image support for a model that does not accept images.

Overview
Prime Inference models were all registered as text-only, so attached images were rejected even for Claude, GPT-5, Grok, and other vision backends. Because Prime Inference’s models API does not expose modality data, vision is now curated via an optional vision flag on PRIME_INFERENCE_MODEL_METADATA, and createPrimeInferenceModel sets input to ["text", "image"] when that flag is set (otherwise ["text"]).

Flagged models include Anthropic Claude variants, OpenAI GPT-5.x entries, Grok, Kimi K2.7 Code, and Qwen3-VL; text-only models such as MiniMax M3 stay unchanged. models.generated.ts and the Prime Inference model test are updated to match.

^{Reviewed by Cursor Bugbot for commit fc5a65d. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Enable image input for vision-capable Prime Inference models

Adds a vision flag to PrimeInferenceModelMetadata in generate-models.ts and sets it to true for several Anthropic Claude, OpenAI GPT 5.x, Kimi, Qwen, and Grok model variants. The createPrimeInferenceModel factory now sets input to ["text", "image"] for vision-flagged models instead of ["text"] only. The generated models.generated.ts reflects these updated input declarations.

^{Macroscope summarized fc5a65d.}