[CoR] Validate Azure AI model availability and quota before provisioning

[nturinski]

Problem

Current deployment quota validation covers several Azure hosting and data services but does not proactively verify Azure AI/OpenAI model availability and capacity. A plan can therefore provision infrastructure and only later discover that the selected subscription, region, model, version, or deployment SKU lacks quota or feature availability.

Proposed work

• Add structured AI deployment requirements to the prepare-plan schema, including model name, version, deployment SKU/type, requested capacity, and region candidates.
• For each candidate region, query the Cognitive Services model catalog to confirm that the requested model/version/SKU is offered.
• Resolve the SKU's usage metric and query Cognitive Services usage to confirm that remaining capacity is sufficient.
• Surface permission failures as unverified, not as successful validation. The usage query may require subscription-level Cognitive Services usage-reader permissions.
• Before scaffolding, offer actionable recovery choices such as selecting a valid region/model, reusing an existing account/deployment, requesting quota, or cancelling.
• Define a structured representation for reusing an existing Azure AI account and deployment rather than inferring it from prose.

Relevant APIs:

• GET /subscriptions/{subscriptionId}/providers/Microsoft.CognitiveServices/locations/{location}/models?api-version=2024-10-01
• GET /subscriptions/{subscriptionId}/providers/Microsoft.CognitiveServices/locations/{location}/usages?api-version=2024-10-01
• CLI equivalent: az cognitiveservices usage list --location <region>

Likely implementation surfaces include:

• resources/agents/azure-deploy/prepare/references/prepare-schemas.ts
• resources/agents/azure-deploy/prepare/references/sku-quota-validation.md
• resources/agents/azure-deploy/prepare/references/subagent-quota.md
• Deployment-plan UI and fixtures/tests that consume quota-validation results

Acceptance criteria

• A planned new AI deployment verifies model/version/SKU availability for its target region.
• Remaining quota is compared with requested capacity before provisioning.
• Unreadable quota is shown as unverified and requires explicit user approval or remediation.
• The user receives a viable alternative instead of discovering the failure during provisioning.
• Existing-account/deployment reuse is represented and validated explicitly.
• Tests cover unavailable models, insufficient quota, sufficient quota, permission failure, and reuse.