Adopt @ai-sdk/openai-compatible in microservice-accessability to use dotCMS as a model provider

Description

Use a dotCMS instance as the model provider for the accessibility agent, via @ai-sdk/openai-compatible.

import { createOpenAICompatible } from '@ai-sdk/openai-compatible';

const dotcms = createOpenAICompatible({
  name: 'dotcms',
  baseURL: `${DOTCMS_HOST}/api/inference/v1`,
  apiKey: DOTCMS_API_TOKEN,
});

model: dotcms('claude-sonnet-4-6')

Depends on #37431, which adds the /api/inference/v1 endpoints.

Where the code lands. dotCMS/microservice-accessability. Provider selection is already isolated in src/agents/runtime/model.ts behind an AgentProvider union and a PROVIDER_DEFAULT_MODEL record, so this is a new case plus config validation in src/config.ts. Tracked here because it is gated on #37431.

Constraints to build against, verified against a running instance. Chat, streaming, tool calling, stream_options.include_usage, embeddings and image generation all work.

• baseURL must end in /api/inference/v1. The package appends /chat/completions, /embeddings and /images/generations; a base URL missing /v1 returns a 404 that reads like a server fault.
• Vision is unsupported. A plain string and an array of text parts are both accepted; any non-text part returns a 400 naming the part type.
• frequency_penalty, presence_penalty, seed and user are accepted and ignored. Only temperature, max_tokens, top_p and stop reach a provider, so seed will not give reproducible output.

Acceptance Criteria

• dotcms added to the AgentProvider union in src/agents/runtime/model.ts, backed by createOpenAICompatible
• PROVIDER_DEFAULT_MODEL gains a dotCMS default
• Config validation in src/config.ts for the host and API token, with the base URL either required in full or derived, so a missing /v1 cannot reach a request
• The API token is sent as Authorization: Bearer and is never logged
• dotCMS error responses map to AI SDK error types where the package's defaults are not already correct
• Existing agent tests still green with the new provider case
• The agent completes a multi-step tool-calling run (src/agents/runtime/model-call.ts, stepCountIs(24)) against a dotCMS instance using only base URL + token
• Setup documented: base URL with /v1, token as apiKey, unsupported vision, ignored sampling parameters

Priority

Medium — blocked on #37431 landing first.

Additional Context

• Server-side counterpart: #37431 (/api/inference/v1 endpoints)
• Implementation repo: https://github.com/dotCMS/microservice-accessability
• @ai-sdk/openai-compatible docs: https://ai-sdk.dev/providers/openai-compatible-providers

Work in dotCMS/microservice-accessability, starting with src/agents/runtime/model.ts and src/config.ts; review the existing AgentProvider and PROVIDER_DEFAULT_MODEL patterns, then inspect src/agents/runtime/model-call.ts for the stepCountIs(24) run. Validate the provider against a dotCMS instance after #37431 lands, run the existing agent tests, and document the /v1 base URL, token, unsupported vision, and ignored sampling parameters.