feat(ai): Multimodal callAI — image content blocks for Claude/Gemini/OpenAI/Mistral

Size: M — enabling infra (prerequisite for onboarding image attachments #65).

Problem

callAI(provider, apiKey, systemPrompt, userMessage) (src/lib/ai.ts:278-333) is text-only — it passes a plain string as the sole user message for all four providers. All the models actually configured are vision-capable (Claude claude-sonnet-4, Gemini gemini-2.0-flash, Mistral mistral-large-latest, OpenAI gpt-4o-mini), so nothing but the wrapper prevents multimodal input. Facet extraction (facet-extract.ts:84) and the onboarding reply (profile/chat/route.ts:67) both call this text-only path.

Scope

Add a multimodal variant (e.g. callAIVision(provider, apiKey, system, text, images: {mime, dataOrUrl}[])) that builds provider-specific image content blocks:

• Claude: content: [{type:"text"}, {type:"image", source:{type:"base64"|"url", ...}}]
• OpenAI: content: [{type:"text"}, {type:"image_url", image_url:{url}}]
• Gemini: contents parts with inlineData/fileData
• Mistral: image_url content parts (pixtral-capable model — may need a model bump)

Keep callAI as the text-only fast path; the vision call is opt-in. Cap image count/size and downscale server-side to control token cost. Handle providers/models that can't do vision by falling back to text + a note.

Blocks: the onboarding attachment feature.

Start in src/lib/ai.ts:278-333 and inspect the existing provider-specific callAI handling, then trace callers at facet-extract.ts:84 and profile/chat/route.ts:67. Done means an opt-in multimodal path supports the four listed providers, preserves the text-only path, limits image size/count, and falls back to text with a note when vision is unavailable.