fix: preserve reasoning across provider steps

[Astro-Han]

Summary

• resolve each model runtime wire and reasoning replay contract once, then share that decision across model construction, execution, and connection probes
• generalize OpenAI Chat reasoning capture/replay from Kimi to compatible providers while preserving the observed reasoning or reasoning_content field and legacy persisted metadata
• keep Anthropic signed thinking, OpenAI Responses items/continuation, Gemini metadata, and structured reasoning_details on their existing protocol-specific paths

Closes #2241

Verification

• npm --workspace @maka/runtime test (3203 tests, 0 failures, 9 skipped)
• npm run lint
• npm run format:check
• focused durable DeepSeek Chat replay, shared runtime-contract, and DeepSeek Responses probe tests
• npm run build and npm run typecheck reach existing desktop protocol type errors also reproducible on the unchanged main checkout; all preceding affected workspaces build/typecheck successfully