fix(core): cap Alibaba thinking budget variants at 64k - #51154

Issue for this PR

N/A. Found while testing default output limits for #51021.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Alibaba thinking budget variants now stop at 64k. Anthropic's already stop at 32k.

models.dev lists budget_tokens support for these models without a maximum, so variant.ts derived the budgets from the catalog output limit: max was the output limit minus one, and high was half of it.

• Scope: Alibaba Chat, Alibaba Messages, and Vercel's alibaba/* models.
• With fix(ai): fit thinking budgets to the output limit #51157: at a 32k compaction limit, fix(ai): fit thinking budgets to the output limit #51157 halves a 63,999 budget to 16,000.
• Before these changes, once feat(core): fit output limits to the context window #51021 sends the catalog output limit, a max budget of 131,071 against 131,072 cut the answer off after 8 tokens, and Alibaba rejected the 32k summary ("max_completion_tokens [32000] must be greater than thinking_budget [131071]").
• Kimi's old max of 262,143 was above the 256k default output limit in feat(core): fit output limits to the context window #51021.

How did you verify your code works?

• Unit tests: cover Alibaba Chat and Messages at the 64k ceiling, a model whose output limit is below it, and Qwen on Alibaba Chat. The packages/core variant and models-dev suites pass.
• Live: with feat(core): fit output limits to the context window #51021 and fix(ai): fit thinking budgets to the output limit #51157, qwen3.7-plus max on Alibaba completed a title, tool steps, and a compaction summary. See fix(ai): fit thinking budgets to the output limit #51157 for the request fields.

Screenshots / recordings

N/A

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR