docs: add Tsubasa setup with bounded context and output

[cenab]

CoreCoder's existing OpenAI client can load Tsubasa's endpoint and aliases. Add a setup row to both READMEs with a 32,768-token context setting and a 4,096-token output request, plus the endpoint/key disclosure and the requirement for enabled tool calls.

No provider implementation, SDK behavior, pricing, or capability metadata changes.

Validation on CoreCoder 0.7.0 (b37645f075f8f82f0f063b2088bea2854813dce5), Python 3.13.15 and OpenAI Python 3.19.2:

• python -m pytest tests -q: 185 passed.
• ruff check corecoder tests, python -m compileall -q corecoder tests, and git diff --check: passed.
• Ten actual installed CLI requests against a controlled loopback server: both aliases, streamed text, a fragmented tool call followed by an actual fixture-file read and serialized tool-result history, a single 401 with exit 1, and SIGINT cancellation with exit 130.
• Every actual body passed Tsubasa's shared strict Chat Completions schema and complete input/output budget check before the server returned a controlled response. HOME was preserved; unrelated personal hooks/MCP configuration was denied by the local test sandbox.

These checks establish client configuration and protocol plumbing. They do not establish live model quality or enable server-side tool capability. Additional MCP tools and long histories can still reach an endpoint's context limit.

Provider-affiliated contribution for Tsubasa; prepared with AI assistance.