[skills] Sync ai-gateway with current docs

[vercel-gh-bot-5]

AI Gateway renamed evaluation models to decision models (docs updated 2026-10-07, ai 7.0.128 / @ai-sdk/gateway 4.0.104 changelog "rename evaluate to decide") and added an OpenAI-compatible Decisions API. The skill still said evaluation is unavailable through OpenAI-compatible endpoints and recommended the now-deprecated experimental_evaluate / gateway.evaluationModel().

• Stale: "it is not available through the OpenAI-, Anthropic-, or Cohere-compatible endpoints" (SKILL.md) and "Evaluation is not supported through the OpenAI-compatible, Anthropic-compatible, or Cohere-compatible endpoints. Use one of the three surfaces above instead." (references/evaluation.md). Correction: decisions are also served by the OpenAI-compatible POST /v1/decisions; they are not supported through Chat Completions, Responses, Anthropic- or Cohere-compatible endpoints. Added an OpenAI Decisions SDK row to the surface table and the OpenAI-compatible surface to the SKILL.md routing row. Sources: https://vercel.com/docs/ai-gateway/modalities/decision (last_updated 2026-10-07): "They are not supported through Chat Completions, Responses, or the Anthropic-compatible or Cohere-compatible endpoints."; https://vercel.com/docs/ai-gateway/sdks-and-apis/openai-decisions (last_updated 2026-10-07): "Ask typed questions about shared input and get back probabilities, choices, and scores using the OpenAI-compatible"; https://vercel.com/docs/ai-gateway/sdks-and-apis (last_updated 2026-10-07).
• Stale: import { experimental_evaluate as evaluate } from 'ai'; and "experimental_evaluate from ai". Correction: experimental_decide (ai 7.0.128+); experimental_evaluate noted as the deprecated alias for earlier AI SDK 7 releases. Source: https://vercel.com/docs/ai-gateway/getting-started/decision (last_updated 2026-10-07): "Earlier AI SDK 7 releases call it"; "Use ai 7.0.128 or later for this quickstart". Confirmed in ai@7.0.133 dist/index.d.ts: /** @deprecated Use \experimental_decide` instead. */ export declare const experimental_evaluate`.
• Stale: "use gateway.evaluationModel(<model-id>)" as the current method. Correction: gateway.decisionModel(<model-id>) with a version note (@ai-sdk/gateway 4.0.104+, shipped with ai 7.0.128+; earlier releases use the deprecated gateway.evaluationModel). Source: https://vercel.com/docs/ai-gateway/modalities/decision (2026-10-07): "Decision models were previously called evaluation models." Confirmed in @ai-sdk/gateway@4.0.108 CHANGELOG (4.0.104: "rename evaluate to decide").
• Redirected links updated: /docs/ai-gateway/modalities/evaluation → /modalities/decision, /getting-started/evaluation → /getting-started/decision, /models-and-providers/evaluation-fallbacks → /models-and-providers/decision-fallbacks (https://vercel.com/docs/ai-gateway/models-and-providers/decision-fallbacks, 2026-10-07: "Decision fallbacks were previously called evaluation fallbacks."). Model list filter ?capabilities=evaluation → ?capabilities=decision, as linked from every decision page.

Verification

run_plugin_checks: build, typecheck, validate, and test all pass (1079 tests, 0 failures).

Test updates

None.

Not changed

• Frontmatter metadata.docs still cites https://vercel.com/docs/ai-gateway/modalities/evaluation, which redirects to /modalities/decision (same product); frontmatter left byte-identical. Prompt signals and entities (experimental_evaluate, v1/evaluate, "evaluation model") left as matching fields.
• "Evaluation" wording, headings, and the references/evaluation.md file name kept; only facts and links changed.
• POST /v1/evaluate remains the documented HTTP endpoint, so it was kept.
• references/evaluation.md "Gateway state can be a string, object, or array": the ai 7.0.133 changelog says arrays passed directly as AI SDK state now contain decision state parts; left for a follow-up since the Gateway HTTP format is unchanged.
• Decision fallbacks stored in virtual models (live-page-only passage on the Decision Fallbacks page) not added.

Opened automatically by the vercel-plugin drift agent. Review before merge.