fix(core): preserve reasoning across model switches - #45401

Issue for this PR

Related to #45334. This addresses readable reasoning replay, not that issue's credential-switching bug.

Type of change

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

What does this PR do?

Keep readable reasoning as reasoning when replaying successful messages under a different model or provider, instead of making it ordinary assistant text. Destination adapters can then use their supported reasoning format.

Metadata stripping, encrypted-only omission, same-model replay, and failed-message handling are unchanged. Adapter policies also remain unchanged: Claude falls back to text for unsigned reasoning; Responses omits reasoning without replay state rather than receiving it as ordinary text.

Updates the existing model-switch test expectation. No new test files, credential changes, or metadata-schema changes.

How did you verify your code works?

• Core: 29 tests passed across session-runner-message and session-model-request.
• AI: 307 tests passed across the Chat Completions, Anthropic, Gemini, and OpenAI Responses adapters.
• bun typecheck from packages/core passed.
• Prettier and git diff --check passed. Oxlint reported no errors and two existing consistent-return warnings in unchanged functions.
• No live provider calls.

Screenshots / recordings

Not applicable.

Checklist

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