fix: harden kimi structured output and replay

[themauveavenger]

Issue for this PR

Closes #15197
Closes #15226

Type of change

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

What does this PR do?

This fixes two Kimi-specific protocol mismatches in opencode's current dev branch.

First, it preserves reasoning_content / reasoning_details when prior assistant turns are replayed through the OpenAI-compatible chat path. That addresses the replay failure reported in #15197 and also helps with the broader replay-stripping problem discussed in #19081.

Second, it stops forcing tool_choice: "required" for Kimi structured-output requests. Kimi thinking mode rejects that combination, so structured output now uses auto for Kimi-family models instead. That addresses #15226.

The fix is intentionally narrow:

• it adds the replay serialization on the OpenAI-compatible chat transport
• it moves the structured-output tool-choice decision into ProviderTransform
• it keeps the Kimi exception keyed off model.family
• it adds regression coverage for both behaviors

Related upstream work:

• fix(opencode): support interleaved reasoning_content for anthropic protocol non-Claude models #14637 and fix(opencode): support interleaved reasoning_content for anthropic protocol non-Claude models #14641 fix the same general replay problem for the Anthropic SDK path; this PR covers the OpenAI-compatible chat path instead, so it complements rather than duplicates those PRs
• feat: support OpenAI-compatible providers (Kimi, DeepSeek) with chat-completions API #18054 improves provider routing for OpenAI-compatible models like Kimi, but it does not address these two protocol mismatches directly
• Kimi k2.5 has issues with tool calling #20650 appears to be a broader provider/model malformed-tool-calling issue; this PR does not try to solve that class of failure, but it should reduce false Kimi failures caused by opencode's own request formatting

In practice, this should make Kimi sessions more stable in two places that currently fail immediately: multi-turn replay after tool use, and JSON-schema/structured-output requests.

How did you verify your code works?

I tested the touched paths locally with:

• bun test test/provider/copilot/convert-to-copilot-messages.test.ts
• bun test test/session/prompt.test.ts
• bun test test/provider/transform.test.ts
• bun test test/session/structured-output.test.ts

I also ran diagnostics on the changed files before committing.

Screenshots / recordings

N/A — no UI changes.

Checklist

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

[smashah]

why is this closed?

[themauveavenger]

why is this closed?

I modified the copilot sdk when it wasn't necessary.

The kimi issue of not sending "auto" might still be a problem. But I'm testing that more myself before I try again.