fix(core): preserve compatible reasoning replay

[opencode-agent]

Summary

• preserve cross-model reasoning as reasoning_content for OpenAI-compatible Chat targets instead of appending it to assistant text
• continue stripping model-bound continuation metadata on switches
• make protocol lowering safe: Anthropic, Bedrock, Gemini, and OpenAI Responses lower unsigned/foreign reasoning to ordinary assistant text, while signed native reasoning still round-trips
• preserve the same behavior through the native LLM adapter and V2 session projection

This keeps Kimi-compatible reasoning history structurally intact without reintroducing the unsigned thinking/signature failures fixed by #25303.

Fixes #32603

Tests

• packages/opencode: targeted message projection, native adapter, and provider transform suites (399 passed)
• packages/core: session runner message suite (6 passed)
• packages/llm: OpenAI Chat/compatible/Responses, Anthropic, Bedrock, and Gemini suites (165 passed)
• bun typecheck in packages/llm and packages/core
• targeted Oxlint: 0 errors
• Prettier check on all changed files

packages/opencode typecheck was also attempted but did not complete within the 10-minute local host limit; the targeted package tests and lint checks passed.