[Bug]: Assistant text emitted before a tool call disappears once later text in the same turn arrives

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/client (chat timeline)

Steps to reproduce

Provider: Claude Code harness, model claude-fable-5-1.

1. Ask a question the model answers in prose, then has it call a tool (in my case SendMessage to a resumed subagent), then write a short closing line. Turn shape is text → tool_use → text.
2. Watch the timeline: the first prose block streams in and is fully readable.
3. When the tool call completes and the closing text arrives, the first block is gone. Only the closing line (e.g. "Queued to the worker.") remains as the assistant message.

Reproduced 3 of 3 times with that turn shape. Turns with no tool call render fine. A turn with text → tool_use → text that ran later did not reproduce, so it may be timing-dependent.

Expected

All text blocks in the turn stay visible, interleaved with the tool-call cards.

Actual

The earlier text block is replaced by the last text block. The substantive answer is lost from the visible transcript.

Environment

• T3 Code 0.0.41-nightly.20260911.1533
• macOS 26.6 (Darwin 25.6.0)
• Claude Code provider

Notes

Possibly related to #7137 (stalled stream truncates the message to its first delta), but the symptom here is the reverse: earlier text is dropped in favour of the last block. Happy to add a screen recording on request.

Start in the apps/client chat timeline and reproduce the Claude Code turn shape text → tool_use → text described in the issue. Trace how streamed assistant text blocks are stored and rendered when the tool call completes; done means all text blocks remain visible, interleaved with the tool-call cards, and turns without tools still render correctly.