ai_abstracted

feat: forward TextRequest.image on Mistral and Ollama, cap Ollama out…

• OllamaTextClient and MistralTextClient now forward TextRequest.image, so vision works on every text provider rather than only Claude and Gemini. Ollama sends bare base64 in its per-message images array; Mistral sends an OpenAI-style image_url content part with a data: URI. Pick a multimodal model — a text-only one rejects the request with AiInvalidRequestException.
• Behavior change OllamaTextClient now maps TextRequest.maxTokens onto options.num_predict. Ollama previously ran unbounded regardless of the field; generations are now capped at the requested budget (default 4096). Raise maxTokens if you relied on the old unbounded behavior.
• A Mistral request without an image keeps its content a plain string rather than a single-element parts array, so text-only models are unaffected.

First public release.

• Provider-agnostic contracts for text, image, video, speech, sound-effect, and music generation, each a single async method that takes a typed request and returns a GenerationResult (bytes plus normalized metadata).
• A shared HTTP transport with exponential backoff, jittered retries, and async-job polling for the providers that run long jobs.
• Typed error hierarchy (AiException and friends) that maps HTTP status codes to auth, rate-limit, invalid-request, transient, and timeout failures.
• Clients for Google Gemini (text and image), Google Veo (video, with Veo 3 audio), OpenAI (image), Black Forest Labs FLUX (image), ElevenLabs (speech and sound effects), Suno (music), Anthropic Claude (text), Mistral (text), and Ollama (local text).
• Multi-turn conversations (TextRequest.history) and an optional image on the current turn (TextRequest.image) for vision-capable text models.
• An environment credential loader, a provider registry, and an in-memory fake for every capability so downstream code stays testable without a network.