fix(core): preserve AI SDK provider metadata

[kitlangton]

Summary

• preserve content-level provider metadata when lowering canonical LLM messages into AI SDK prompts
• cover reasoning and tool-call replay metadata in the compatibility adapter

This keeps provider-specific continuation data such as thought signatures and provider item IDs available during history replay.

Testing

• bun run test test/aisdk.test.ts from packages/core
• bun typecheck from packages/core
• pre-push workspace typecheck: 31 successful, 31 total

[opencode-agent]

Audit note: #35817 overlaps this provider-metadata preservation path and is broader (namespace preservation across SessionRunner collection/replay plus AI SDK lowering). Please clarify whether #35817 builds on this PR or supersedes it before either merges.

[kitlangton]

Superseded by #35817, which is already merged into and preserves the provider metadata namespace while covering the same replay path. Rebasing this branch onto current leaves no diff. The focused AI SDK test and Core typecheck pass on the merged implementation.

[kitlangton]

Correction: superseded by #35817, which is already merged into v2 and preserves the provider metadata namespace while covering the same replay path. Rebasing this branch onto current v2 leaves no diff. The focused AI SDK test and Core typecheck pass on the merged implementation.