[Feature] Add GPT-6 Astra to Azure OpenAI built-in support

[Xunzhuo]

Maintainer direction

I recommend accepting this as a bounded Data Plane & Networking feature. #3561 establishes GPT-6 Astra's canonical model card and direct OpenAI contract; Azure OpenAI needs its own provider binding because Azure deployment names, authentication, paths, and API-version handling are distinct runtime semantics.

Goal

Make openai/gpt-6-astra selectable through the built-in azure-openai provider with verified Chat Completions and Responses behavior, while keeping the operator's Azure deployment name explicit.

Current-main gap

• config/catalog/resources/providers/azure-openai.yaml currently advertises only openai/chat-completions@1#create and has no catalog model bindings.
• Microsoft's current reasoning-model documentation lists GPT-6 Astra version 2026-09-03 for both Chat Completions and Responses, with a 1,050,000-token context window and 128,000 maximum output tokens: https://learn.microsoft.com/en-ca/azure/foundry/openai/how-to/reasoning
• The direct OpenAI binding in [Feature] Add GPT-6 Astra built-in support #3561 cannot safely stand in for Azure's operator-defined deployment name or Azure-specific request construction.

Scope

• Add the verified openai/responses@1 operation and effective Azure request path/API-version behavior without regressing the existing Chat Completions path or api-key authentication.
• Add an openai/gpt-6-astra managed-cloud binding. Declare provider_model_id_kind: deployment_name and preserve the canonical catalog name separately; never invent or silently derive an operator's deployment name.
• Record only protocol capabilities and reasoning efforts verified for Azure. In particular, prove Chat xhigh and Responses max rather than copying the direct OpenAI binding without provider evidence.
• Add provider-boundary fixtures that assert the final Azure URL/query, headers, deployment identifier, model/request body, reasoning field, and tool preservation for Chat and Responses.
• Regenerate the catalog projections and update the Day-0 guide if Azure introduces a reusable managed-deployment step not already documented by [Feature] Add GPT-6 Astra built-in support #3561.

Completion criteria

• A catalog-backed user configuration with an explicit Azure deployment name materializes without hand-written model-card metadata.
• Chat Completions with xhigh reaches the Azure fixture with exactly one valid reasoning representation.
• Responses with max and a tool reaches the Azure fixture with the correct path, API version, deployment semantics, and preserved tool schema.
• Unsupported protocol/effort combinations fail before dispatch with a useful validation error.
• Catalog validation, materialization tests, provider conformance tests, generated projections, and public documentation pass their owning gates.

Non-goals

• Discover, create, or rename Azure deployments.
• Generalize every Azure OpenAI model or redesign Azure authentication.
• Change the canonical GPT-6 Astra card or direct OpenAI behavior delivered by [Feature] Add GPT-6 Astra built-in support #3561.

Reference implementation: #3561.

[Xunzhuo]

/accept