fix(opencode): replay Cerebras reasoning in assistant content

[ryanl-cerebras]

Issue for this PR

Closes #26762

Type of change

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

What does this PR do?

Cerebras models go through @ai-sdk/openai-compatible, which replays assistant reasoning parts as assistant.reasoning_content.

That breaks follow-up turns for Cerebras reasoning models like cerebras/zai-glm-4.7, which expect prior reasoning to be folded back into assistant.content.

This change rewrites prior assistant reasoning parts before serialization for @ai-sdk/cerebras:

• GLM reasoning is replayed as <think>...</think> text
• GPT-OSS reasoning is replayed as plain assistant text

It also adds regression tests for both cases.

How did you verify your code works?

• Reproduced the failure on opencode 1.14.41 with cerebras/zai-glm-4.7 using:
  • build an frontend for inferencex performance benchmark
• Verified the patched local build gets past the follow-up turn that previously failed
• Ran:
  • bun test test/provider/transform.test.ts
• Confirmed the new Cerebras regression tests pass

Screenshots / recordings

N/A

Checklist

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

[map9959]

This may be addressed through the SDK version in #30186

[ryanl-cerebras]

Closing as superseded by #34826, which fixes #26762 through the updated Cerebras SDK and adds request-level regression coverage. Thanks for getting this issue identified and helping validate the expected reasoning replay behavior.