provider proposal: native Mistral AI integration

[stackbilt-admin]

Summary

Preliminary provider proposal: evaluate adding a native Mistral AI provider to @stackbilt/llm-providers.

Why this came up

During OSS imagery generation for this repo, the text-to-image model hallucinated Mistral as an integrated provider. That is not currently true: the package has Mistral model IDs through other providers, but no first-class MistralProvider.

Current related support:

• Cloudflare Workers AI includes @cf/mistral/mistral-7b-instruct-v0.1.
• NVIDIA NIM includes mistralai/mistral-large-2-instruct.

Those are provider-hosted models, not native Mistral API integration.

Initial scope questions

• Should the repo support the native Mistral API as a first-class provider?
• Which models should enter the catalog initially?
• Does native Mistral support the same package surfaces we expect from providers: streaming, tool/function calling, structured output, usage accounting, prompt/cache metadata, and balance/health semantics?
• Should LLMProviders.fromEnv() discover MISTRAL_API_KEY?

Preliminary acceptance criteria

• Add MistralProvider only after verifying current native API request/response/stream shapes against primary documentation or live fixtures.
• Add provider config type, factory wiring, fromEnv() detection, model catalog entries, and model drift tests.
• Add schema drift fixtures/tests for non-streaming and streaming responses.
• Document capabilities accurately in README and CHANGELOG.
• Keep imagery/marketing copy accurate: do not list Mistral as integrated until this ships.

Notes

This is not a bug in current runtime behavior. It is a product/package roadmap issue created to convert a hallucinated visual claim into an explicit tracking primitive.