feat: add preferChatCompletions for older dual-mode models

[Archit-Kumar-16]

Summary

• Add catalog field preferChatCompletions so older dual-mode models keep Chat Completions on /chat/completions unless the existing header, reasoning, or Cursor trigger fires.
• Backfill the field on older dual-mode models across openai, azure-open-ai, microsoft-foundry, aws-bedrock-mantle, aws-bedrock, and databricks.
• Omit the field on gpt-5.4+ (and newer) so inbound Chat Completions is always served via Responses.

GPT audit: every older dual-mode GPT model that lists both chat and responses now has the field. gpt-5.4+ dual-mode GPT models do not. Chat-only GPT models (openrouter, azure-ai-foundry, cerebras, sambanova, and chat-only snapshots) do not need it because they are not Responses-capable.

Test plan

• Confirm CUE validation still accepts model YAML with preferChatCompletions: true
• Spot-check gpt-4o / gpt-5 / gpt-5.2 have the field; gpt-5.4 / gpt-5.6-sol / gpt-6 do not
• Confirm databricks hyphenated ids: databricks-gpt-5 has the field, databricks-gpt-5-4 does not
• Pair with the gateway PR that reads this field

Made with Cursor