fix(runtime): preserve OpenAI and Gemini reasoning replay

[dom-mp]

Summary

Fixes #257.
OpenAI reasoning models can reject multi-tool replay when a prior function_call is replayed without its paired reasoning item. NimbleBrain was applying the older-reasoning stripping optimization globally, even though that behavior was only intended for Anthropic thinking/signature bloat.
This PR makes reasoning replay policy provider-aware:

• Anthropic keeps the existing older-reasoning stripping behavior.
• OpenAI preserves reasoning history so Responses API tool-call replay remains valid.
• Gemini preserves reasoning/thought metadata so replay-sensitive thoughtSignature data stays intact.

Testing

• bun test test/unit/window.test.ts
• bun run check
• bun run format:check
• bun run lint