Add Tsubasa to the OpenAI-compatible provider catalog

[cenab]

Tsubasa users currently need to enter an OpenAI-compatible endpoint and model manually. Would a named Tsubasa preset fit Zero's provider catalog?

At 99721c7, openAICompat already supplies the transport and Chat Completions format. The proposed change would add a descriptor using that helper, with:

• Name and ID: Tsubasa / tsubasa
• Base URL: https://api.tsubasa.sh/v1
• Credential: TSUBASA_API_KEY, sent as bearer authentication
• Models: tsubasa-fast and tsubasa-pro; both have 32,768-token context windows
• Maximum output: 8,192 for Fast and 16,384 for Pro; an initial smaller request budget must leave room for Zero's prompt and tools

The existing custom-provider path is the alternative. No new transport, OAuth flow, default-provider change, or special routing is proposed. If the preset is welcome, the implementation would include catalog/factory regression checks and an actual configured-client check for request budgets, streaming, credential isolation, and HTTP errors. Tools would need explicit qualification rather than an inferred capability flag.

This is a Tsubasa-affiliated proposal prepared with OpenAI Codex. Source inspection is complete, but Zero-specific runtime and live Tsubasa inference have not been tested. The September 28 public model-list probe returned HTTP 404. I will wait for the required approved parent issue before implementation, as your contribution policy requests.