fix(agents): keep streamed thinking ahead of text and tool calls - #11837

Fixed part of #11800

Summary

Both agent loops moved thinking that streamed before the text onto the tool-call message, so a turn the model produced as [thinking, text, tool_use] was persisted and replayed as [text, thinking, tool_use]. Anthropic binds each thinking signature to the blocks before it in the same message, so on Claude Fable 5.1 this alone drops the block on the next request of every tool round. Users lose the model's reasoning mid-task without any error.

Thinking now stays where the model produced it.

• Unrolled agent loop: the relocation is removed. Same-id chunks already coalesce in stream order. The only normalization left drops a copy of reasoning a provider repeats on the tool-call chunk.
• Legacy loop: the first split tool-call message carries only thinking that arrived on the tool-call chunk. Later split messages repeat the turn's thinking, which DeepSeek and Kimi require on every tool-call message; fix_conversation already removes the signed copies before an Anthropic request.

Net deletion. The wire payload for OpenAI-compatible, Google, and Responses providers is unchanged.

Testing

Existing DeepSeek, Kimi, and multi-tool reasoning tests pass unchanged. The Anthropic test now asserts signed thinking appears once and leads the tool-call message; the unrolled-loop pipeline test pins the block order. Wire-verified on Fable 5.1 with block binding: both loops replay [thinking, text, tool_use] with input_transformations: [].

Stack

1. fix(anthropic): preserved-thinking compliance for the provider layer #11836 provider layer: block binding, signature-aware history
2. fix(agents): keep streamed thinking ahead of text and tool calls #11837 agent loops: keep streamed thinking ahead of text and tool calls (this PR)
3. feat(agents): freeze the system prompt per conversation and deliver changes as system messages #11864 agent loops: freeze the system prompt, deliver changes as system messages