Add xAI language model parameters

[guillaumegay13]

Summary

Adds xAI API-key parameter coverage for the current xAI language models Manifest can route to xAI:

• grok-4.3
• grok-4.20-0309-reasoning
• grok-4.20-0309-non-reasoning
• grok-4.20-multi-agent-0309 via /v1/responses
• grok-build-0.1

Also adds the xai provider display label and includes the xAI structured-output format selector for Chat Completions (response_format.type) and Responses (text.format.type).

Sources

Official xAI docs used:

• https://docs.x.ai/developers/models
• https://docs.x.ai/developers/models/grok-4.3
• https://docs.x.ai/developers/models/grok-4.20-0309-reasoning
• https://docs.x.ai/developers/models/grok-4.20-0309-non-reasoning
• https://docs.x.ai/developers/models/grok-4.20-multi-agent-0309
• https://docs.x.ai/developers/models/grok-build-0.1
• https://docs.x.ai/developers/rest-api-reference/inference/chat
• https://docs.x.ai/developers/model-capabilities/text/reasoning
• https://docs.x.ai/developers/model-capabilities/text/structured-outputs
• https://docs.x.ai/developers/model-capabilities/text/multi-agent

Media, voice, image input, streaming, tools, tool choice, and related tool orchestration controls are intentionally not modeled here because they are endpoint/request-shape capabilities rather than standalone MPS parameters. This PR keeps only scalar/enumerated parameters that map cleanly into the current schema.

Applicability notes

• grok-4.3 keeps reasoning_effort values none, low, medium, and high. Live smoke confirmed reasoning_effort: none is accepted, but presence_penalty, frequency_penalty, and stop are still rejected by the API, so those fields are omitted instead of modeled as conditional.
• grok-4.20-0309-reasoning and grok-build-0.1 work through Chat Completions, but live smoke showed reasoning_effort is rejected on that route, so it is omitted.
• grok-4.20-0309-non-reasoning accepts stop, but live smoke showed presence_penalty and frequency_penalty are rejected, so only stop is included from that set.
• grok-4.20-multi-agent-0309 is Responses API-only. It now includes max_output_tokens, temperature, top_p, reasoning.effort, and text.format.type; Chat Completions-only paths such as max_completion_tokens and response_format.type are intentionally omitted.

Validation

• npm run validate
• npm test
• npm run guard:params
• npm run typecheck
• npm run lint
• npm run build
• git diff --check

Live xAI smoke

Ran with the local xAI key from .env.keys without printing or committing the key.

Positive smoke:

• grok-4.3 with reasoning_effort: none, temperature, top_p, seed, and max_completion_tokens returned 200.
• grok-4.20-0309-non-reasoning with temperature, top_p, seed, max_completion_tokens, and separately stop, returned 200.
• grok-4.20-0309-reasoning with temperature, top_p, seed, and max_completion_tokens returned 200.
• grok-build-0.1 with temperature, top_p, seed, and max_completion_tokens returned 200.
• grok-4.20-multi-agent-0309 on /v1/responses with max_output_tokens and reasoning.effort returned 200.
• grok-4.20-multi-agent-0309 on /v1/responses with temperature and top_p returned 200.
• grok-4.20-multi-agent-0309 on /v1/responses with text.format.type=json_object returned 200.

Expected negative smoke:

• grok-4.3 with stop returned 400 as expected.
• grok-4.3 with penalties returned 400 as expected.
• grok-4.20-0309-non-reasoning with penalties returned 400 as expected.
• grok-4.20-0309-reasoning and grok-build-0.1 with reasoning_effort returned 400 as expected.
• grok-4.20-multi-agent-0309 on Chat Completions returned 400 as expected.
• grok-4.20-multi-agent-0309 on /v1/responses with Chat Completions-style response_format returned 400 as expected; the API says to use text.format on Responses.

[vercel]

@guillaumegay13 is attempting to deploy a commit to the Manifest Team on Vercel.

A member of the Team first needs to authorize it.