provider proposal: Cohere integration

[stackbilt-admin]

Summary

Preliminary provider proposal: evaluate adding Cohere as a first-class provider to @stackbilt/llm-providers.

Why this came up

During OSS imagery generation for this repo, the text-to-image model hallucinated Cohere as an integrated provider. That is not currently true. The package currently supports OpenAI, Anthropic, Cloudflare Workers AI, Cerebras, Groq, and NVIDIA NIM.

Initial scope questions

• Should the provider target Cohere Chat/Command models only, or include embeddings/rerank APIs in a separate surface?
• Which capabilities map cleanly to the existing LLMProvider contract: streaming, tool/function calling, structured output, citations/search, usage accounting, cache metadata?
• Which models should enter the catalog initially, and what use cases should they be tagged for?
• Should LLMProviders.fromEnv() discover COHERE_API_KEY?

Preliminary acceptance criteria

• Verify current Cohere request/response/stream shapes against primary documentation or live fixtures before implementation.
• Add provider config type, factory wiring, fromEnv() detection, catalog entries, capability metadata, and model drift tests.
• Add schema drift fixtures/tests for non-streaming and streaming responses.
• Document capability boundaries clearly, especially if embeddings/rerank are out of scope for this text-generation package.
• Keep imagery/marketing copy accurate: do not list Cohere 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.