Emit <thinking> tag boundaries during streaming

[cpsievert]

Summary

• When ContentThinking chunks are emitted during streaming, the streaming loop now emits <thinking>\n before the first thinking chunk and \n</thinking>\n\n when transitioning to non-thinking content (or at end of stream)
• For content="text" consumers, tags are yielded as string chunks — concatenated output is well-formed
• For content="all" consumers, behavior is unchanged — typed ContentThinking objects are yielded, no tag strings
• ContentThinking now has a _complete PrivateAttr (default True). Streaming chunks are constructed via _as_chunk() with _complete=False, so __str__() returns bare text for fragments instead of wrapping each one in <thinking>...</thinking> tags
• Removes the synthetic "\n\n" separator from the OpenAI provider's reasoning_summary_text.done event (now redundant)

Companion to tidyverse/ellmer#975.

Motivation

Currently, streaming thinking content has two issues:

1. content="all" mode yields ContentThinking objects whose __str__() independently wraps each chunk in <thinking>...</thinking> — printing produces repeated tags around each fragment
2. content="text" mode yields thinking as bare strings indistinguishable from response text

After this change, concatenating a content="text" stream produces:

<thinking>
reasoning content here...
</thinking>

Response text here...

And calling str() on individual ContentThinking chunks in content="all" mode returns the raw thinking text without tag wrapping.

Test plan

• 10 unit tests covering sync/async, text/all modes, thinking-only streams, text-only streams, tag chunk boundaries, and str() on chunks
• Full existing test suite passes (190 passed, 3 skipped — only bedrock fails due to live API requirement)
• Pyright passes with 0 errors on changed files

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated no new comments.