fix(ai-sdk): preserve reasoning parts in message conversion

[hannesrudolph]

Problem

When using AI-SDK-backed providers (notably DeepSeek deepseek-reasoner) across tool-call turns, the provider may require the assistant's reasoning to be round-tripped (e.g. via reasoning_content). We were dropping that reasoning during our Anthropic → AI SDK message conversion, which can cause DeepSeek to reject follow-up requests after a tool call.

Context: DeepSeek “thinking mode” requires returning reasoning_content in subsequent requests within the same turn when tool calls are involved.

Closes #11199

What changed

1) Preserve reasoning through the AI SDK conversion layer

• convertToAiSdkMessages() now converts:
  • our stored { type: "reasoning", text: string } content blocks → AI SDK { type: "reasoning", text } parts
  • Anthropic extended thinking blocks { type: "thinking", thinking, signature } → AI SDK { type: "reasoning", text } parts
  • message-level reasoning_content (when present) → a canonical AI SDK reasoning part (avoids duplicating with content blocks)

This keeps the AI SDK ModelMessage stream complete, so the provider package (e.g. @ai-sdk/deepseek) can decide what to send to the native API (including mapping reasoning parts back into reasoning_content).

2) Store reasoning as structured content in task history (instead of <think> tags)

• Task.addToApiConversationHistory() now persists reasoning as a dedicated first assistant content block (type: "reasoning"), rather than embedding <think>...</think> into the text.

This enables consistent round-tripping and avoids mixing reasoning into user-visible text.

3) Cleanup

• Removed invalid "openai-compatible" from the API-provider reasoning allowlist (it is not a ProviderName; it’s used for code-index embedding).

Tests

• Added AI SDK conversion coverage in ai-sdk.spec.ts
• Updated task history persistence coverage in reasoning-preservation.test.ts

[roomote]

Oroocle   See task on Roo Cloud

Re-review after latest commit: follow-up resolved; no additional issues found.

• Ensure flattenAiSdkMessagesToStringContent still works when reasoning parts are present (avoid sending multipart content to string-only models like SambaNova DeepSeek).
• Revisit the aiSdkProviders reasoning preserve allowlist in Task.buildCleanConversationHistory (missing AI SDK providers like gemini and vertex, or switch to a capability-based check).

Previous reviews

• 5a3c173: Review #1
• 32d7325: Review #2

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}